Replace Your First 5 Hires With AI

What a Part-Time Marketer Actually Costs You

Before you hire, get honest about the number. A fractional content marketer or social media contractor typically runs $2,000–$5,000/month for 10–15 hours a week. A full-time marketing coordinator starts around $48,000–$66,000/year base — and once you layer in payroll taxes (7.65% FICA match), health insurance (roughly $700/month employer share), recruiting fees, and a 10–12 week ramp where you get maybe 60% productivity, that base becomes $65,000–$92,000 loaded. A boutique content agency retainer covers roughly the same surface area for $3,500–$7,000/month — no sick days, but also no institutional memory of your brand.

The honest comparison: an AI marketing agent covering content drafts, scheduling, email, and basic SEO costs you the Axiom plan you are already on. You are not getting a strategist. You are getting the 60–70% of a marketer's week that is repetitive production work — blog drafts, social variants, metadata, email copy — and that is exactly where the math makes sense. Strategy, channel decisions, and anything requiring judgment about your specific business stay with you.

What to Delegate to the Agent

• Blog first drafts from a brief. You write 5 bullets: target keyword, audience, angle, 2–3 points to make, any product tie-in. The agent writes 800–1,200 words. You edit for accuracy and voice. Net time: 15 minutes instead of 3 hours.
• SEO metadata at scale. Title tags (under 60 characters), meta descriptions (under 155), image alt text for every product photo or blog graphic. This is where human marketers let things slip — the agent does not skip files.
• Social post variants from existing content. One blog post becomes 4–6 platform-specific variants: a tight 280-character X post, a LinkedIn hook-plus-paragraph, an Instagram caption with a CTA, a Google Business Profile update. Feed it the URL, specify the CTA, done.
• Email newsletter drafts. Give it your "what happened this week" notes — a new product, a customer win, an upcoming promotion — and it returns a draft with subject line, preview text, body, and a single CTA. You rewrite the voice where it sounds off.
• Monthly performance narrative. Paste your GA4 or Search Console numbers. The agent writes the "here is what changed and why it probably happened" paragraph for your own records or a client report.
• A/B copy variants. Three subject line options for every email. Two hero headline variants for a landing page test. Four CTA button texts. Humans are bad at generating genuine variants without anchoring to the first option they wrote; models are better at this specific task.
• Repurposing queues. Backlog of blog posts that never got turned into social content? The agent can work through a list — URL in, 5 social posts out — as a batch job.
• Google Business Profile post drafts. A 150-word weekly GBP post with relevant keywords, a photo prompt suggestion, and a local CTA. Most small businesses skip this entirely; in 2026 it feeds directly into Google's local AI-search answers.
• Competitor content gap analysis. Give the agent your top 5 competitors' blog URLs and your own sitemap. It returns a list of topics they cover that you do not, rough search intent included. You decide which gaps to fill.
• AI-search optimisation tasks. Rewriting FAQ sections into direct question-and-answer format so Google AI Overviews, Perplexity, and ChatGPT Search can surface your content as a cited answer. Drafting JSON-LD schema for FAQ, HowTo, and Product pages.

The Guardrail: The Approval Calendar

Nothing the agent drafts goes live without passing through a single chokepoint: a content calendar you review once a week, or once before any scheduled publish. In Axiom, the marketing agent operates on the propose-then-approve loop — it stages drafts and queues them for your sign-off; it does not publish autonomously. That is the correct default. Do not override it until you have reviewed at least 30 pieces of output and genuinely trust the voice. Your weekly review slot should take 20–30 minutes: skim each draft for factual accuracy, brand voice, and CTA accuracy, then release the queue.

Content Brief Template (Copy and Use This)

CONTENT BRIEF
─────────────────────────────────────────────
Format:         [Blog post / Email / Social batch / GBP post]
Target keyword: [Primary keyword you want to rank for]
Audience:       [Who is reading this — be specific, e.g. "solo e-commerce founders, 1–5 years in"]
Angle / hook:   [The one sentence that makes this piece worth reading over the 40 others on the topic]
Key points:
  1.
  2.
  3.
Product tie-in: [Optional — where/how your product appears, if at all. Be specific: feature name, use case.]
CTA:            [What you want the reader to do at the end — one action only]
Tone notes:     [Anything that deviates from your baseline voice — e.g. "more direct", "drop the jargon"]
Do NOT include: [Common mistakes, competitor names to avoid, topics out of scope]
Length target:  [Word count or character limit]
─────────────────────────────────────────────
Output requested:
  [ ] Draft body copy
  [ ] Title tag (max 60 chars)
  [ ] Meta description (max 155 chars)
  [ ] 3 social variants (X / LinkedIn / Instagram)
  [ ] Subject line + preview text (if email)
  [ ] FAQ schema markup (if the piece answers specific questions)

Weekly Marketing Output Checklist

1. Monday — brief submission. Drop your content brief for the week's primary asset (one blog post or one email). Five minutes if you use the template above.
2. Tuesday — draft review. The agent returns the draft. Read for factual accuracy first, voice second. Mark any claim about your product, pricing, or results that needs verification before the piece goes anywhere.
3. Wednesday — social queue build. The agent takes the approved draft and generates 4–6 social variants. Review takes 10 minutes: check CTA accuracy and that nothing reads off-brand.
4. Thursday — email draft if applicable. If you send a weekly newsletter, the agent drafts from your bullet notes by Thursday. You add any personal note or story, then approve.
5. Friday — schedule and release. Approve the calendar in Axiom. Posts go to the schedule you set. GBP post gets queued. You are done with marketing production for the week.
6. End of month — performance pull. Paste Search Console impressions, clicks, and top queries into the agent. It drafts a one-paragraph narrative. You add your interpretation and save it to your records.

The 2026 Reality: Local SEO and AI-Search Are Not Optional

Two surfaces now intercept your potential customers before they reach your site: Google AI Overviews and AI chat tools (Perplexity, ChatGPT Search, Copilot). A meaningful share of commercial queries — especially "best X for Y" and "how do I do Z" — now return a synthesised answer at the top of the page. Users read it and move on. The only way to appear in those answers is to have content structured the way these systems can parse: direct answers to specific questions, FAQ and HowTo schema on your pages, and citations from sources these models already index (industry publications, review platforms, your own high-authority content). Your AI marketing agent can draft FAQ JSON-LD, rewrite FAQ sections into clean Q&A format, and pull together a list of questions your competitors rank for that you do not. It cannot tell you which of those questions your customers actually care about — that is your call.

A support rep is often the first hire founders make — and the most emotionally-driven one. A customer complains twice in a week, you're drowning in email, and you post a job listing before doing the math. Let's do the math first.

What This Hire Actually Costs You

A full-time support rep in the US runs $38,000–$52,000 in base salary. Add employer FICA (7.65%), state unemployment insurance, and a modest health contribution and you're at $48,000–$73,000 loaded per year — call it $4,000–$6,000 per month, every month, whether your ticket volume is 3 or 300. That doesn't include recruiting ($4,000–$5,000 is a typical cost-per-hire for an entry-level role), a 4–8 week ramp before they're useful, or the three to four months of institutional knowledge required before they can handle anything off-script without pulling you in.

The structural problem outlasts the cost: a human rep works roughly 40 hours across five days. Your customers contact you on Saturday afternoons, Sunday evenings, and at 11 PM on a Tuesday. First-response time — the metric most directly tied to support satisfaction — suffers every time a ticket lands outside business hours. You hired someone to fix the gap, and the gap is still there for roughly 40% of the week.

What to Delegate vs. What to Keep Human

The Guardrail That Makes This Work: Propose → Approve

The failure mode with AI support isn't that it sends a wrong answer — it's that it sends a confident wrong answer, at scale, before anyone notices. The fix is a hard rule: anything non-routine goes through a propose-then-approve loop. The agent drafts the reply, stages it, and waits. You see it in a queue. You approve, edit, or reject. Only then does it send.

This isn't a theoretical safety switch — it's the operating model. Over time, as you verify that the agent handles specific categories reliably, you can open the autonomy tap category by category. But you open it deliberately, based on observed performance, not by default.

If you're using Axiom by Digitalix Hub, this is how the support agent works out of the box: it proposes, you approve. Routine tasks (tagging, follow-ups, CSAT) run autonomously. High-stakes replies — anything involving money, commitments, or an unhappy customer — sit in your approval queue until you act. The agent cannot send those on its own.

The Support Answer Library: Build This Before You Launch the Agent

An AI support agent is only as good as the answer library you give it. Skip this step and it will improvise — and improvisation is where errors live. Pull your last 90 days of support tickets. Find every question that appeared more than twice. Write a canonical answer to each. That is your answer library. Here is the minimum structure for each entry:

# Support Answer Library Entry

question_id: faq-001
trigger_phrases:
  - "what's your refund policy"
  - "can I get a refund"
  - "money back"
  - "cancel and refund"
category: billing
urgency: normal
tier: 1  # 1 = autonomous send | 2 = propose/approve | 3 = human only

canonical_answer: |
  Hi [first_name],

  Our policy is a 14-day money-back guarantee on your first plan. No
  questions asked, as long as you haven't yet completed your onboarding
  setup. If you're within that window, reply here and we'll process it
  the same day.

  If you're past 14 days or have completed onboarding, reach out and
  we'll look at your specific situation. We're not going to fight you
  over it.

  [Agent name], [Your Company] Support

edge_cases:
  - Customer mentions being a long-term subscriber: escalate to Tier 2, do not auto-send
  - Customer mentions a charge they don't recognize: escalate to Tier 2
  - Customer uses the word "lawyer" or "dispute": escalate to Tier 3 immediately

last_reviewed: 2026-06-01
reviewed_by: [owner name]

Build 15–20 of these entries before you go live. Audit them every 90 days. Your refund policy changes. Your product changes. Your tone changes. An answer library that hasn't been reviewed in six months will produce confident, outdated replies — and outdated replies at scale are worse than slow replies.

Go-Live Checklist

• Answer library complete: Minimum 15 canonical entries covering your most common ticket categories. Each entry has explicit edge-case escalation rules — not vague instructions like "escalate if complex," but named triggers (specific words, customer attributes, dollar thresholds).
• Tier classification decided: Every answer library entry is labeled Tier 1, 2, or 3. Nothing is ambiguous about which replies go to a queue versus send autonomously.
• Escalation path wired: Angry and legal tickets trigger an immediate notification to you — email, Slack, or Discord ping — not just a queue entry. You need to see these within minutes, not at your next inbox check.
• Brand voice document written: 3–5 sentences describing exactly how your brand sounds in support. Formal or casual? First names? Apologize freely or stay factual? Sign off as a named agent persona or as the company? This goes into the agent's system instructions verbatim.
• Prohibited actions list explicit: Write out what the agent is never allowed to do in a reply — offer discounts above a specific amount, make promises about unbuilt features, comment on competitors, share any customer data. Explicit beats implicit every time.
• Human review queue tested: Before launch, send yourself 5 test tickets and confirm the propose-then-approve flow surfaces them correctly — in the right interface, with the right notification, to the right person.
• First-response SLA defined: Set internal targets: e.g., Tier 1 answers in under 3 minutes, Tier 2 replies in the approval queue within 1 hour, Tier 3 escalation pings you immediately. Measure against these weekly.
• CSAT survey templated and scheduled: 24 hours after resolution, dispatch a one-question rating survey. Decide in advance what score triggers a personal follow-up from you.
• Product feedback loop has a destination: Where do tagged feature-request tickets actually land? A Notion database, a weekly digest email, a GitHub issue? Define the destination before tickets start flowing — otherwise the tags are meaningless.
• Audit schedule blocked: 30 minutes per month, read a random sample of 10 resolved tickets — not just the flagged ones. This is how you catch tone drift and outdated answers before they become customer complaints.

Reply Template: The On-Brand Holding Response

One of the highest-leverage things AI support can do is send an immediate, human-sounding acknowledgment the moment a ticket arrives — before anyone has read it. This single message does two things: it resets the customer's anxiety clock (they know someone received it), and it buys time for the Tier 2 approval queue without the customer feeling ignored. Below is a working template. Adapt the tone to match your brand voice document.

Subject: Got it — [ticket subject or first 6 words of their message]

Hi [first_name],

Thanks for reaching out. I've got your message and I'm on it.

If this is time-sensitive, reply to this email with the word URGENT
in the subject line and I'll move you to the front of the queue.

You'll hear back from me with a real answer [within X hours / by
end of business today / within 24 hours — pick one and mean it].

[Agent name]
[Your Company] Support

—
Message received: [timestamp]
Ticket reference: [ticket_id]

A few things this template does deliberately: it doesn't say "your ticket has been received" in corporate-passive voice. It doesn't promise a specific answer yet — just a real one. The URGENT escape hatch gives the customer agency without creating a separate process you have to build. And the timestamp plus ticket ID make the interaction traceable without requiring the customer to dig through their email.

Reply Template: The Tier 3 Handoff (Escalation to Human)

When the agent detects an escalation trigger — anger, legal language, a chargeback threat — it should send one message and stop. Not attempt to resolve. Not apologize profusely. Not explain policy. One message:

Hi [first_name],

I've flagged your message for [owner name / 'our team lead'] and
you'll hear directly from them within [2 hours / by end of day].

If it's urgent, you can also reach [owner name] directly at
[direct email or phone].

[Agent name]
[Your Company] Support

Short, human, and it gets out of the way. The agent's job at Tier 3 is detection and handoff — not resolution. Anything more risks making a tense situation worse.

The Metrics That Actually Tell You If It's Working

Don't measure ticket volume — that's input, not outcome. Measure these four:

Check these weekly for the first month. After that, monthly is enough unless something spikes. The goal isn't to hit arbitrary benchmarks — it's to catch problems before they become patterns.

What This Hire Actually Costs

A US-based SDR runs $45,000–$65,000 base salary plus commission, with OTE typically landing at $60,000–$85,000. Load that with payroll taxes, benefits, a LinkedIn Sales Navigator seat ($99/month), a CRM seat, and a sequencing tool like Apollo or Outreach ($100–$200/month), plus the recruiter fee to find them in the first place — you are looking at $90,000–$115,000 per year all-in. For that investment, a realistic SDR will spend roughly 60% of their day on work that does not require human judgment: finding prospects, researching accounts, writing first-touch emails, logging activity, and nudging stale threads. The remaining 40% — running a discovery call, reading hesitation in a prospect's tone, building the kind of relationship that actually closes — still needs a person. The question is whether your revenue justifies paying full-time rates for the 60% before you have the pipeline to prove it.

What You Can Delegate to an AI Agent

• Lead research and enrichment: Given a target company name or domain, pull firmographic data (industry, headcount, funding stage, tech stack via BuiltWith), identify the right contact by title, and surface recent trigger events — a new funding round, a job posting for a role that signals pain, a press mention. Tools like Apollo, Clay, and Clearbit automate the raw data pull; an AI agent can orchestrate the enrichment pipeline and produce a one-paragraph research brief per account so you are not starting from a blank page.
• Drafting personalized first-touch outreach: Not mail-merge personalization (Hi {{FirstName}}, I noticed you work at {{Company}}). Actual contextual reasoning: 'They just posted three SDR roles, which usually means they are scaling a sales motion and probably feeling the pipeline-quality problem.' The agent reads the trigger, maps it to the pain your ICP commonly has, and drafts a short email with a specific hook. You review before anything leaves your domain.
• Multi-step follow-up sequences: Most replies come after the third or fourth touch. An agent can stage a five-to-seven step sequence — email on day 1, LinkedIn connection request on day 3, a value-add email on day 7 (a relevant case study or piece of content, not a pitch), a break-up email on day 14 — and pause the sequence automatically when a prospect replies or books a meeting. You never have to remember to follow up; the agent manages the queue.
• CRM hygiene: Logging calls, updating deal stages, flagging contacts who have not been touched in 30-plus days, noting when a prospect changes jobs (a high-intent signal worth a separate outreach), and generating a pipeline freshness report every Monday. This is the work most SDRs hate and most CRMs automate badly.
• List building and segmentation: Pulling accounts that match your ICP from Apollo or LinkedIn Sales Navigator, deduplicating against existing CRM contacts, and organizing into priority tiers — tier 1 (perfect ICP fit plus a live trigger event), tier 2 (ICP fit, no trigger), tier 3 (partial fit). The agent does the filtering; you decide whether the criteria are pointed at the right market.
• Reply triage and routing: When a prospect responds, the agent reads the reply, classifies it (interested, not now, wrong person, unsubscribe), drafts a suggested next message, and flags it for your review. You approve and send. Nothing goes out without you reading it.

What Must Stay Human

• The discovery call: You cannot automate the conversation where you figure out whether this prospect has a real problem, a real budget, and a real timeline. An agent can prepare you — research brief, suggested questions, summary of prior touches — but the call is yours.
• Relationships with your top 20 accounts: The prospects most likely to close have talked to you personally more than once. A prospect who eventually realizes they have been running on AI autopilot will not buy. Use the agent to keep the thread alive, but show up yourself at the moments that matter.
• Pricing and deal structure: Anything involving negotiation, custom terms, or discounting requires judgment about that specific deal and your own revenue priorities. No agent should be authorized to make pricing commitments.
• Live objection handling: When someone says on a call that they looked at you six months ago and went with a competitor, responding well requires context, empathy, and real-time strategic thinking. No sequence can replicate that.
• Deciding which accounts to pursue at all: Your ICP will shift as you learn what actually closes. The agent executes against the criteria you set; you are responsible for whether those criteria point at the right market.

The Honest Limits of AI Outreach

Cold email reply rates are low across the board. For most B2B outbound, a 3–8% reply rate on a well-targeted sequence to a warm list is realistic, and AI-generated copy is not going to change that math dramatically. What changes is the volume you can sustain without burning your own time, and the consistency of follow-up — most founders give up after two touches, and most SDRs after three; a configured sequence runs to seven. AI does not solve deliverability. If your domain is under 90 days old, if you are sending from an unwarmed inbox, or if your subject lines contain trigger words that spam filters flag, the best-written email lands in junk. Set up SPF, DKIM, and DMARC before you run sequences at any scale, and use a warm-up tool (Mailreach, Lemwarm) if you are starting a new sending domain. Personalization at the 'I read your last three blog posts' level sounds compelling in theory and often reads as uncanny valley in practice. The best AI outreach names one concrete trigger and stops — the first email should be under 100 words.

SDR Agent Setup: Step-by-Step

1. Define your ICP in writing before touching any tool. Industry, company size (headcount range), geography, tech stack signals, buyer title, and the two or three pain indicators that predict a fit. If you cannot write this in a paragraph, the agent will enrich the wrong leads and you will not know until you have wasted three weeks of sequences.
2. Pick your data source. Apollo.io covers most B2B use cases at $49–$99/month and includes built-in sequence functionality. Clay (from $149/month) is more powerful for enrichment when you are pulling from multiple sources and want to run conditional logic on each row. LinkedIn Sales Navigator ($99/month) gives the most precise title and company filtering but requires a separate outreach tool.
3. Build your trigger list. What events signal a prospect might be ready right now? Funding announcements (Crunchbase, Tracxn), job postings that reveal pain (LinkedIn, Greenhouse), product launches (ProductHunt, press releases), tech stack changes (BuiltWith alerts). Pick two or three triggers you can monitor systematically — not twelve.
4. Write your sequence templates — five steps maximum to start. See the template table below. Keep step 1 under 100 words. Each subsequent step either adds value (a resource, a specific example relevant to their industry) or makes a direct ask. Never send two pitches in a row.
5. Set the approval rule explicitly. In whatever tool you use, configure sequences to stage but not send until you review. In Axiom, this is the default: your sales agent queues outreach in a propose-and-approve loop — nothing moves until you confirm. If your sequencing tool does not have a staged-approval mode, treat every sequence as a draft until you have manually reviewed the first 20 sends.
6. Define your weekly review ritual. Twenty to thirty minutes on Monday morning: review staged outreach for the week, approve or edit, check replies from the prior week, update deal stages. That is your SDR function for the week.
7. Track three numbers weekly. New leads enriched and staged. Reply rate on sequences that have run at least 14 days. Meetings booked. Pipeline freshness — deals with no activity in 21-plus days — is your lagging indicator that something has stalled and you have not noticed yet.

5-Step Outbound Sequence: Starter Templates

ICP Definition Worksheet

Trigger Event Monitoring: Setup Checklist

• Funding announcements: Set a Crunchbase Pro alert for your target industries and headcount range. New rounds (especially Series A/B) mean the company has capital and is under pressure to show growth. Reach out within 5 business days of the announcement — the window closes fast.
• Job postings that signal pain: Set a LinkedIn Jobs alert for specific titles at companies in your target segment. A company hiring for a role your product supports is advertising their problem in public.
• Tech stack changes: BuiltWith lets you monitor when a company adds or drops a specific technology. If your product integrates with HubSpot, a company that just adopted HubSpot is a warm lead.
• Press mentions and product launches: Set a Google Alert for target company names plus terms like 'launches,' 'announces,' 'partners with.' ProductHunt surfaces companies at the exact moment they are building something that creates demand for what you sell.
• Job changes at warm contacts: A contact who moves to a new company carrying the same role is one of the highest-intent leads you will find. LinkedIn Sales Navigator's job-change alerts exist for this. Apollo also flags it in the contact activity feed.

Domain Authentication: Pre-Sequence Checklist

Run this before you send a single sequence. Deliverability problems are invisible until your reply rate is inexplicably zero — by then you may have burned the domain.

Weekly SDR Review: 30-Minute Ritual

Using Axiom for Sales

Axiom's sales agent is not a sequencing tool — it is one specialist inside your company's shared AI roster, sitting alongside your support, finance, and growth agents in a single shared memory. That distinction matters in practice. When a prospect replies asking about pricing, the agent already knows your current plan structure, your refund policy, and any prior interactions that were logged. It does not need to be briefed each time.

The sales agent operates on the same propose-and-approve loop as every other Axiom agent. It drafts outreach, stages follow-ups, prepares research briefs, and surfaces pipeline freshness alerts — then waits for your approval before anything externally-facing moves. Routine internal tasks (logging, segmenting, flagging stale deals) run without intervention. Anything that touches a prospect waits for you.

Axiom starts at EUR 19/month with no free usage tier and no API keys to configure — AI usage is included in every plan. Your first plan carries a 14-day money-back guarantee, which runs until you submit your first onboarding answer. After that, the guarantee does not apply. You fund the account, pick a plan, and the agents go to work.

A freelance bookkeeper working 10–15 hours a month costs $300–$800/month. Bring one in-house part-time and you're looking at $45,000–$60,000/year once you load in benefits, payroll taxes, and the weeks you spent finding them. Spend a day watching what they actually do and you'll notice most of those hours are transactional: copying numbers between systems, sending the same late-payment email for the third time this month, sorting a bank export into categories you defined two years ago. That work is not skilled — it's disciplined and consistent. An AI agent is better at disciplined and consistent than any human. The final numbers, the tax return, anything that carries a professional license or touches legal liability: that stays with a human. This section draws a hard line between those two worlds, then gives you the exact templates and schemas to hand the transactional work off immediately.

What to delegate — and what never leaves your desk

The payment chaser sequence — exact templates

These four messages form a complete sequence triggered off the invoice due date. Load them into your agent as a scheduled workflow. The tone is professional and direct — it gets progressively firmer without becoming aggressive. Replace the sign-off name with whoever owns the client relationship. At T+14 the sequence pauses and routes to you; the agent does not send a fourth automated chase without your instruction.

--- T-3 DAYS (3 days before due date) ---
Subject: Quick reminder — Invoice #[INV-NUMBER] due [DATE]

Hi [FIRST NAME],

Just a heads-up that Invoice #[INV-NUMBER] for [AMOUNT] is due on [DATE].
Payment details are on the original invoice — if you need a copy resent,
just reply and I'll send it straight over.

Thanks,
[YOUR NAME]


--- DAY 0 (due date) ---
Subject: Invoice #[INV-NUMBER] — due today

Hi [FIRST NAME],

Invoice #[INV-NUMBER] for [AMOUNT] falls due today.
If you've already sent payment, please ignore this.
If there's a hold-up on your end, let me know and we can work something out.

Payment link: [LINK]

Thanks,
[YOUR NAME]


--- T+7 DAYS (one week overdue) ---
Subject: Invoice #[INV-NUMBER] — 7 days overdue

Hi [FIRST NAME],

I'm following up on Invoice #[INV-NUMBER] for [AMOUNT], due on [DATE] and
now 7 days past due. If there's a problem with the invoice or you need
different payment arrangements, I'm happy to talk it through.

Otherwise, I'd appreciate payment at your earliest opportunity.

Payment link: [LINK]

[YOUR NAME]


--- T+14 DAYS — ESCALATION TO OWNER (sequence pauses) ---
[Internal flag to owner — not sent to client]

Invoice #[INV-NUMBER] | Client: [CLIENT NAME] | Amount: [AMOUNT]
Status: 14 days overdue. Automated sequence paused.

Recommended next step: a direct call or personal email from you.

Options:
  A) Resume automated chasers (not recommended past this point)
  B) Draft a formal final-notice letter
  C) Hold — I'll wait for your instruction

Which would you like?

The weekly money snapshot — schema and delivery

The most useful financial habit for a small-business owner is reading one consistent report every Monday morning before opening email. Not a full P&L — just enough to know whether you need to make calls today. An AI agent generates this automatically from your accounting tool's API or a weekly CSV export. The schema below defines exactly what it should contain and what the agent should flag automatically.

WEEKLY MONEY SNAPSHOT — Week ending [DATE]

1. CASH POSITION
   Current bank balance:            $[X]
   vs. last week:                   +/- $[X]
   vs. 4-week rolling average:      +/- $[X]

2. RECEIVABLES (what customers owe you)
   Total outstanding:               $[X]

   Breakdown by age:
     Current (not yet due):         $[X]  ([N] invoices)
     1–30 days overdue:             $[X]  ([N] invoices)
     31–60 days overdue:            $[X]  ([N] invoices)  ← flag for owner review
     60+ days overdue:              $[X]  ([N] invoices)  ← escalate: human contact required

   Auto-flag: if 1–30 day bucket > 20% of total receivables, note it here.

3. DAYS SALES OUTSTANDING (DSO)
   Current DSO:                     [X] days
   Prior week DSO:                  [X] days
   Your payment terms:              Net [X] days
   Trend:                           Improving / Stable / Worsening

   Auto-flag: if DSO exceeds payment terms by more than 5 days, note it here.

4. TOP 5 EXPENSES THIS WEEK
   [Category]                       $[X]
   [Category]                       $[X]
   [Category]                       $[X]
   [Category]                       $[X]
   [Category]                       $[X]

   Auto-flag: any category more than 20% above its 4-week average.

5. INVOICES SENT THIS WEEK
   Count: [N]     Total value: $[X]

6. PAYMENTS RECEIVED THIS WEEK
   Count: [N]     Total: $[X]

7. ONE-LINE SUMMARY
   [Agent drafts one sentence. Example:
   "Cash is up $4,200 on last week; two invoices totalling $8,500
   are now 31+ days overdue and need your attention before Friday."]

The two numbers that tell you where you're headed

Most small-business owners check the bank balance and stop there. The balance tells you where you are right now. These two numbers tell you what's coming.

• Days Sales Outstanding (DSO): the average number of days between sending an invoice and receiving payment. Formula: (total receivables ÷ revenue in the period) × number of days in the period. If you bill $30,000/month and have $25,000 outstanding, your DSO is roughly 25 days. If your terms are net-30 and your DSO climbs above 35, clients are stretching you — that's the week to tighten the chase sequence, not a month later when it becomes a cash problem.
• Overdue invoice count by age bucket: raw number of invoices in the 1–30, 31–60, and 60+ day bands. Watch the 60+ bucket specifically. That's where write-offs come from. Anything that lands there should trigger a phone call from you, not another automated email.
• Why not profit? Profit is a trailing indicator your accountant calculates after the month closes. DSO and overdue count are leading indicators — you can act on them this week. A business can show healthy profit and still hit a cash crisis because clients pay slowly. These two numbers give you the early warning.

Invoice draft checklist — what the agent must include every time

When you instruct an AI agent to draft an invoice, it should produce a document that could go straight to the client after a 60-second review. That means every field below is present, accurate, and populated from company memory — not left blank for you to fill in manually.

• Invoice number — sequential, never reused. The agent maintains the counter in company memory and increments it automatically.
• Issue date and due date as actual calendar dates (e.g. 15 July 2026, not 'net 30'). An explicit date removes the ambiguity that clients exploit.
• Your legal business name and registered address — exactly as they appear in your business registration, not a trading name variant.
• Client's legal name and billing address — ask once, store in company memory, never retype.
• Line items: description specific enough that someone unfamiliar with the project could understand what was delivered. 'Consulting — June' fails this test. 'Brand strategy workshop, 3 hours, 4 June 2026' passes it.
• Quantity, unit rate, and line total for each item, listed separately.
• Subtotal, tax rate and tax amount shown as separate line items (required for VAT compliance in most jurisdictions; check your local rules).
• Total amount due, in the agreed currency.
• Payment instructions in full: bank name, account number, sort code or routing number — or a payment link. Both if you accept both. Do not make the client hunt for how to pay you.
• Late payment policy stated plainly on the invoice face, e.g. 'Invoices unpaid after 30 days may incur a late fee of 1.5% per month.' Check enforceability in your jurisdiction before including a rate.
• A unique payment reference the client should quote — typically the invoice number. This is what links their bank transfer to your open invoice when you reconcile.

How Axiom handles this

Axiom by Digitalix Hub runs a Finance agent alongside your other specialist agents — support, sales, content — all sharing the same company memory. That means the Finance agent already knows your standard payment terms, your chart-of-accounts categories, your clients' billing addresses, and your bank details the moment you set them up. It doesn't need a separate integration wizard for each piece. The propose-approve model applies to everything financial. Routine work — drafting invoices, sending scheduled payment chasers, generating the weekly snapshot — runs autonomously. High-stakes actions — credit notes, write-offs, anything that moves money out — come to you as a flagged approval before anything happens. You confirm or adjust; the agent executes. Axiom plans start at EUR 19/month. There's no free usage tier, but the first plan carries a 14-day money-back guarantee — provided you haven't yet submitted your first onboarding answer, which is when the system begins building your company memory. AI model usage is included in every plan; there are no API keys to source or manage.

The hard line — a one-page policy to post inside your business

Confusion about what the AI agent can and can't do on its own causes more problems than the agent itself. Write this out and share it with anyone who touches your finances.

FINANCE AGENT — WHAT IT DOES AUTONOMOUSLY / WHAT REQUIRES HUMAN SIGN-OFF

AUTONOMOUS (agent runs without asking):
  - Draft new invoices from job records or your instructions
  - Send the T-3, day-0, and T+7 payment chaser emails
  - Categorize bank and card transactions against defined rules
  - Flag uncategorizable or ambiguous transactions for review
  - Generate and deliver the weekly money snapshot every Monday
  - Match bank export lines to open invoices
  - Alert you when DSO exceeds your payment terms by more than 5 days
  - Alert you when any invoice hits the 60+ day overdue bucket

APPROVAL REQUIRED (agent drafts, owner confirms before action):
  - Any outbound payment or bank transfer
  - Credit notes, refunds, or invoice adjustments
  - Writing off a bad debt
  - Sending the T+14 escalation or any contact beyond the standard sequence
  - Any communication that quotes a legal right or threatens further action

HUMAN ONLY (agent does not touch):
  - Approving month-end or year-end financials
  - Filing tax returns (VAT, GST, income tax, payroll tax)
  - Giving advice on deductibility or tax treatment
  - Making judgment calls on disputed invoices past T+14
  - Anything requiring a professional license to perform

The researcher is the first hire most solo founders mentally skip — and the first function that quietly collapses. You started strong: you did a competitive sweep before launch, you read the reports, you knew your market. Then the first customer showed up, the first invoice needed paying, and the next time you looked up, six months had passed. You're making pricing decisions based on what you knew at launch. Your biggest competitor shipped a new tier in February and you found out when a prospect mentioned it on a call in May. This is not a discipline failure. It's a capacity one. Research is the easiest thing to deprioritize because its cost is invisible until a decision goes wrong — and by then you're playing catch-up instead of calling the shot.

What You're Actually Delegating

The research function has two layers. The first is surveillance: watching a defined set of signals on a fixed schedule so nothing material changes without you knowing. The second is synthesis: when you need to understand something quickly — a new competitor, a market shift, a pricing question — pulling together what's knowable in an hour rather than burning a day on it. Both layers are repetitive, schedulable, and require almost no judgment in execution. The judgment lives entirely in what you do with the output. That distinction is the whole design principle here: you're not outsourcing thinking, you're outsourcing retrieval and first-pass summarization so your thinking starts from a better position.

What to Delegate, What to Keep, and Where the Line Is

The Guardrail That Makes This Work

Your Weekly Research Brief (Template)

Fill in the bracketed sections once. After that, the AI does the work. Save this as a recurring task — in a Claude Project with web search enabled, in Perplexity, or as an Axiom research agent brief. The output format is designed to be skimmable in under five minutes: you want to see what changed, what needs your attention, and confirmation that the quiet areas actually ran.

WEEKLY RESEARCH BRIEF — [YOUR COMPANY NAME]
Run every: Monday 08:00
Output format: Slack-ready bullet digest + flagged items

== COMPETITOR WATCH ==
Check the following URLs for any changes since [LAST_RUN_DATE]:
- [Competitor 1 pricing page URL]
- [Competitor 1 changelog or blog URL]
- [Competitor 2 pricing page URL]
- [Competitor 2 changelog or blog URL]
[Add up to 5 competitors]

For each: note what changed, what did not, and whether the change affects our positioning.
If nothing changed: say so in one line. Do not pad.

== BRAND MENTIONS ==
Search for mentions of [YOUR BRAND NAME] on:
- Reddit (past 7 days)
- G2 / Capterra (new reviews only)
- X / Twitter (past 7 days)

Also search for mentions of [COMPETITOR 1] and [COMPETITOR 2] alongside the words
"switched", "left", "moved from", "cancelled", "alternative to".

Summarize: overall sentiment trend, any recurring complaints, any direct comparison to us.

== MARKET SIGNALS ==
Scan the following topics for anything published in the last 7 days that a founder
in [YOUR INDUSTRY] should know about:
- [Topic 1 — e.g. "AI agent pricing models B2B SaaS"]
- [Topic 2 — e.g. "small business software consolidation 2026"]
- [Topic 3]

Limit to 3 most material items. For each: one-sentence summary + one sentence on why it
matters to a company at our stage.

== OUTPUT FORMAT ==
**Changed this week:** [bullet list — real changes only, with source URLs]
**Mentions worth reading:** [bullet list with direct links]
**Market signals:** [3 bullets max]
**Flagged for your decision:** [anything that warrants a call from you — competitor move,
  pricing gap, product threat, unexpected praise]
**Quiet this week:** [what you checked and found unchanged — so you know the sweep ran]

== CONFIDENCE LABELLING ==
Tag every numerical claim or market-size figure with one of:
[AI-synthesized] — treat as directional; verify before quoting
[Source-confirmed] — verified against a named primary source (include URL)

The Scheduled Research Stack (What to Run and When)

• Weekly (Monday morning): Run the brief above. Competitor pricing and feature changes, brand mentions, three market signals. The digest should take under five minutes to scan. Anything flagged gets a calendar block the same day.
• Monthly (first Monday): Full competitor feature matrix refresh. Pull each competitor's feature page and update your comparison table below. Note any new capability they shipped that you don't have — this is your earliest roadmap signal.
• Monthly (first Monday): Review digest — pull the last 30 days of G2 and Capterra reviews for your top two competitors. Categorize by theme: what are customers asking for that the competitor isn't delivering? Those gaps are worth more than any analyst report.
• Quarterly: Deep-dive on one strategic question you've been deferring — market sizing, pricing benchmarking, ICP expansion, a new vertical. Give the AI a full brief and block 30 minutes to read and annotate the output before doing anything with it.
• Triggered (ad-hoc): Any time a prospect mentions a competitor you haven't tracked, run a one-shot company brief before your next call. Twenty minutes of prep replaces an analyst hour. The template is below.

The Competitive Intelligence Schema (Copy Into Notion, Obsidian, or Your Notes App)

One profile per competitor. Update the 'Last updated' line and the Recent Changes section every time your weekly sweep flags something. Everything else updates on your monthly refresh. The Confidence Tier section at the bottom is non-optional — it tells future-you which numbers you actually checked.

## Competitor Profile: [Competitor Name]
Last updated: [DATE] | Source: AI research sweep / manual review

### Basics
- Website:
- Founded / HQ:
- Pricing (current):
- Free tier: Yes / No / Freemium
- Primary ICP (who they're actually selling to, based on their copy and case studies):

### Positioning
- Tagline / hero headline (exact text, with date captured):
- Core differentiator they claim:
- Tone: enterprise / SMB / developer / prosumer
- What problem they lead with:

### Features vs. Us
| Feature area       | Them | Us  | Notes                           |
|--------------------|------|-----|---------------------------------|
| [Feature 1]        |  ✓   |  ✓  |                                 |
| [Feature 2]        |  ✓   |  ✗  | On our roadmap — Q3             |
| [Feature 3]        |  ✗   |  ✓  | Active differentiator for us    |
| [Feature 4]        |  ?   |  ✓  | Unconfirmed — check their docs  |

### Pricing Model
- Structure: per seat / flat / usage-based / hybrid
- Entry tier:  $[X]/month for [what's included]
- Mid tier:    $[X]/month for [what's included]
- Top tier:    $[X]/month for [what's included]
- Annual discount: Yes ([X]%) / No / Unknown
- Free trial: [X] days / money-back guarantee / none
- Last price change: [DATE or Unknown]

### Recent Changes (last 90 days)
- [DATE]: [What changed] — [source URL]
- [DATE]: [What changed] — [source URL]

### Customer Complaints (G2 / Capterra / Reddit — with dates)
- Most common complaint:
- Second most common:
- Recurring feature requests they're not meeting:
- Praise they consistently receive (so you know where they're genuinely strong):

### Our Positioning Response
- Where we win against them in a direct comparison:
- Where they win against us — be honest:
- How to handle them when a prospect brings them up:
- Red flags that suggest a prospect is actually a better fit for them than for us:

### Confidence Tier
- Pricing data:  [AI-synthesized / Source-confirmed — URL]
- Feature data:  [AI-synthesized / Source-confirmed — URL]
- ICP data:      [AI-synthesized / Source-confirmed — URL]
- Complaints:    [AI-synthesized from reviews / Manually read]

One-Shot Company Brief (Ad-Hoc Template)

Use this when a competitor surfaces unexpectedly — on a prospect call, in a Reddit thread, from a churned customer. Run it before your next conversation touches that name.

ONE-SHOT COMPETITOR BRIEF — [COMPETITOR NAME]
I'm a founder at [YOUR COMPANY]. I just heard about this company and need a fast brief
before a sales call. Give me:

1. What they do in plain language (2 sentences max)
2. Who they're built for — their stated ICP and what their copy actually suggests
3. Pricing — current plans with prices; flag if you can't find a public price
4. Top 3 features or capabilities they emphasize
5. What customers complain about (G2, Capterra, Reddit — recent only)
6. How they position against companies like mine ([YOUR CATEGORY/NICHE])
7. Anything that changed in the last 60 days

For every specific claim: cite a URL or label it [AI-synthesized].
Do not speculate. If you don't know something, say so.

A Note on AI-Search Visibility (2026 Reality)

Your competitors' research presence now extends beyond Google rankings. When a prospect asks ChatGPT, Perplexity, or Google's AI Overview 'what's the best [your category] tool for small businesses,' the answer is generated from a synthesis of what AI systems have indexed and weighted — your G2 reviews, your blog posts, what people say about you on Reddit, how consistently your product description appears across the web. This matters to your research stack in two ways. First, run your weekly brand-mention sweep across AI-search surfaces, not just traditional review sites — search your brand name in Perplexity and note whether it appears, and what it says. Second, the competitive intelligence you gather on your competitors' content tells you what topics they're reinforcing. If they're consistently publishing on a topic and you're not, they're building citation surface in AI systems that you're not. This is not a reason to panic. It is a reason to include 'AI-search presence' as a column in your competitive feature matrix.

How Axiom Runs This Role

In Axiom, the research function runs inside the Growth agent — a specialist that shares company memory with your Support, Sales, and Content agents. That shared memory matters: when the Growth agent surfaces a competitor insight, it's already in context for your next campaign brief or sales sequence, rather than sitting in a doc nobody opens. You set the competitor list, topic watchlist, and schedule once during onboarding. From then on the agent runs the weekly sweep on its own and surfaces a digest for your review.

The approve loop is worth understanding here. Axiom agents propose; you approve before anything consequential goes out. For research, that means the digest lands in your queue — you review it, flag what matters, and the relevant insight is available to the other agents immediately. The agent doesn't make strategic calls. It makes sure you're never the last person in the room to know something changed. Axiom accounts start free with no card required; plans start at EUR 19/month and include AI usage — no API keys to manage. The first plan comes with a 14-day money-back guarantee, valid until you submit your first onboarding answer.

The biggest mistake founders make with AI agents is starting with the role they're most excited about rather than the one that will build trust fastest. Trust matters because you're not just adopting a tool — you're changing how decisions get made in your business. Start where volume is highest, judgment requirement is lowest, and a mistake costs you an awkward email rather than a lost customer or a compliance problem. Then expand, week by week, once you've seen what the agent actually does with your real work.

Pick your first role: the sequencing logic

Before you open any software, answer this honestly: which role is eating two or more hours of your day right now, produces mostly repetitive output, and has a clear right-or-wrong answer most of the time? For most solo founders and small teams, that's one of two things — support triage (answering the same dozen questions on repeat) or content production (writing posts, emails, and copy that follow a consistent format). Either works as a starting point. Support triage has slightly lower initial stakes because a staged draft never reaches a customer until you say so. Content has a natural review gate before anything publishes. Neither is forgiving of vague instructions, which is why you write those down first.

Your week-one plan: five steps, one agent, no shortcuts

1. Day 1 — Pick exactly one role and write down what it actually does. Open a notes doc and list every task that role would handle in a typical week. Be specific: not 'handle customer questions' but 'reply to refund status requests, answer the password-reset question, triage bug reports to the right Slack channel, and follow up with anyone who hasn't heard back in 48 hours.' This list becomes the agent's scope. Everything not on it stays with you. Spend 20 minutes here — it will prevent hours of confusion later.
2. Day 2 — Complete the onboarding Q&A. In Axiom, this is the structured intake where you describe your business context, tone, approval preferences, and what the agent must never do without asking. Answer as specifically as you can. 'Friendly but professional' is not useful. 'Write as a knowledgeable colleague, not a customer-service script — use the customer's name, reference their specific issue, never promise a refund without the phrase "let me check on that for you"' is useful. The quality of your answers here directly determines the quality of the agent's first outputs.
3. Day 3 — Define your approval gates explicitly. Decide which outputs are routine (agent handles autonomously; you review asynchronously) and which are high-stakes (agent drafts; you approve before anything happens). A sensible default for week one: everything goes through approval. You're building trust, not automating yet. In Axiom, the propose → approve loop is on by default — the agent stages its output and waits. You loosen this over time, for patterns you've confirmed hold up. Write a short list of your hard lines — things the agent must never say or commit to without you — and put them in the agent's instructions explicitly.
4. Day 4 — Give the agent one real piece of work. Not a simulation. One real support email to draft a reply to, one real blog brief to turn into a draft, one real month of transactions to categorize. Watch the output. Don't just check whether it's 'good enough' — check whether it did what you meant, used the right tone, flagged the right things, and stayed within scope. Edit what needs editing. Note what surprised you, both ways. This is your calibration session.
5. Day 5 — Review the outputs and set next-week scope. Block 30 minutes. Ask: which outputs needed zero edits? Which needed heavy edits, and why? What did the agent miss that it should have caught? What did it flag for approval that was clearly routine? Use your answers to tighten the instructions before week two. The goal on day five is not perfection — it's a clear picture of where the agent is reliable enough to run with less oversight. Expand scope only in areas where it earned that.

Week-one setup checklist

• Role scoping doc complete: every in-scope task written out in plain English, nothing vague
• Onboarding Q&A answered: tone, context, audience, hard-line rules — specific, not generic
• Approval gates defined: written list of what requires your sign-off vs. what can run autonomously (default for week one: everything requires sign-off)
• Hard-line rules written into agent instructions: things it must never say, never commit to, never send without you
• One real task queued: not a test prompt — a live piece of work from your actual backlog
• Escalation path clear: if the agent flags something it cannot handle, where does it go? Your inbox, a specific Slack channel, a tag in your project tool — decide now
• 30-minute review blocked for day 5: on the calendar, not optional
• Expansion criteria written down: what would have to be true for you to give this agent more autonomy in week two? Write it now so you do not move the goalposts later

Instruction template for your first agent

Copy this into your agent's instruction field and fill in every bracketed section. Specificity here is the difference between an agent that saves you two hours a day and one that creates editing work. Leave nothing generic.

ROLE: [Support Triage Agent / Content Agent — pick one]

BUSINESS CONTEXT:
We are [company name], a [one-sentence description of what you do].
Our customers are [who they are — e.g. 'independent restaurant owners in the UK'].
Our product/service is [brief, specific description].

TONE:
Write as a knowledgeable, direct colleague — not a corporate script.
Use the customer's name. Reference their specific situation.
Do not use: 'I hope this finds you well', 'Please don't hesitate', 'As per my previous email',
'Happy to help', 'Great question'.
Match the customer's formality level. If they write casually, reply casually.
Maximum reply length: [e.g. 150 words]. If it takes more than that, something is wrong.

IN SCOPE (handle or draft a reply):
- [Specific task type — e.g. 'Refund status questions']
- [e.g. 'Password reset instructions']
- [e.g. 'Billing cycle questions']
- [e.g. 'Shipping timeline questions']

OUT OF SCOPE (flag for human review — do not draft a reply):
- Any complaint mentioning legal action or a regulatory body
- Any refund request above [your threshold — e.g. £50]
- Any message where the customer appears distressed or angry
- Any request for account deletion or data export
- [Any other hard lines specific to your business]

APPROVAL REQUIRED (stage the draft and wait — do not send):
- Everything, during week one
- [After week one, list the specific task types cleared to run autonomously]

NEVER:
- Promise a specific resolution timeline without checking
- Confirm a refund is approved without explicit human sign-off
- Refer to internal team names, tools, or processes
- Use the phrase [any phrase you have banned]

ON UNCERTAINTY:
If the situation does not fit a clear category above, flag it for human review with a
one-line summary of why you are unsure. Do not guess.

What week two looks like

By the end of week one you will have a clear view of which task types the agent handles reliably and which still need your hand on them. Week two is about two things: tightening the instructions based on what you saw, and selectively loosening the approval gate on the patterns that held up consistently. A reasonable week-two move for a support agent: switch routine password-reset replies and shipping-timeline questions to autonomous, keep refund requests and anything with a complaint on full approval. For a content agent: let the agent schedule social drafts autonomously; keep email newsletters on approval until the voice is fully calibrated.

The agents that become genuinely useful at month two are the ones where the owner treated week one as a proper calibration, not a quick setup to tick off. The time you put in on the instruction template and the day-five review compounds — every refinement improves every output that follows.

The most common objection to running AI agents in your business is not cost. It is trust. You have spent years building customer relationships, a reputation, and a way of doing things — and handing any of that to software feels like a gamble. That objection is worth taking seriously. So before explaining how Axiom works, here is the honest version of what running AI agents actually means in practice — and what it does not mean.

Short answer: the agents do not have a free hand. Every action they take either runs inside a narrow lane you defined during onboarding — routine, low-stakes, reversible — or it stops and waits for you to approve it. You decide where each boundary sits. You can start with the most restrictive settings, where almost everything comes to you for sign-off, and move the dial as you build confidence. Nothing goes to a customer, gets published, or touches your accounts unless you either pre-cleared that category or clicked approve.

The propose → approve loop

This is the core mechanism. When an Axiom agent picks up a task, it does not execute and then tell you what happened. It produces a draft — a reply, a post, a schedule update, an outreach message — and surfaces it for your review before anything leaves the system. You see exactly what it intends to send or do. You approve it, edit it, or reject it. Only after approval does it act.

The loop has three stages. First, the agent picks up the task from your company's shared memory — it already knows your voice, your policies, your pricing, and your customer history because every agent on your roster draws from one shared context. Second, it produces a draft: the email copy, the content outline, the support reply, whatever the task calls for. Third, it either sends the draft to your approval inbox (for anything high-stakes) or, for task types you have pre-cleared as routine, it executes and logs what it did so you can audit later. Nothing is hidden.

Your approval inbox

Think of the approval inbox as a manager's review queue. Every proposed action that crosses a threshold — contacting a customer, publishing content, updating a deal, sending a proposal — lands here as a card. The card shows you what the agent drafted, what triggered it, what context it used, and exactly what will happen if you approve. You approve with one click, edit inline and approve, or reject with a note. That rejection note feeds back into the agent's understanding of your preferences, so the same mistake does not recur.

On a typical day as a solo operator, the inbox is not overwhelming. You might see three to eight cards: a support reply drafted for a tricky customer question, a social post queued for the afternoon, a follow-up email staged for a prospect who went quiet. Each card takes ten to thirty seconds to review. The daily oversight commitment on a basic plan is closer to fifteen minutes than two hours. That is the trade: you replace the hours you would spend doing these tasks yourself with a shorter block spent reviewing and approving what the agents drafted.

What runs autonomously vs. what requires your approval

This is not binary — it is a spectrum you configure. The table below shows the defaults Axiom starts you on and how they map to common agent tasks. "Autonomous" means the agent acts and logs. "Gated" means it drafts and waits for you. You can move any configurable row.

Dialing autonomy up over time

Think of autonomy as calibration, not a one-time switch. You start conservative. You watch what the agents propose. When a category of draft — support replies for billing questions, for example — looks right fifteen times in a row, you release that category. You keep everything else gated. Over the first thirty to sixty days, most operators settle on two or three categories they are comfortable releasing, and the rest stays in the inbox. That is normal and healthy. The agents handle the volume; you stay close to anything that touches a customer relationship or your public reputation.

Axiom gives you a per-task-type toggle in the settings panel — no engineering required. You flip "support reply sending" from gated to autonomous, and from that point forward that category runs without interrupting you. You can revert any setting at any time. If you notice a pattern of errors in a released category, gate it again, review a week's worth of drafts, and release it again once you understand the failure mode.

What the approval card looks like

Every item in the approval inbox follows the same structure so you can scan it quickly. This is the format Axiom uses:

APPROVAL REQUEST — [Agent Name]

TASK TYPE:       Support reply / Outreach email / Content post / Other
TRIGGERED BY:    [Customer name] sent a message at [time] / Scheduled post at [time]
CONTEXT USED:    [What the agent knew: customer tier, history, last interaction]

DRAFTED ACTION:
—————————————————————————————————————————
[The exact text, content, or action the agent proposes — in full, nothing summarised]
—————————————————————————————————————————

RISK LEVEL:      Low / Medium / High
IF APPROVED:     Will send / publish / update at [exact time and channel]
IF REJECTED:     Task closes; no action taken; you can add a note for training

[ APPROVE ]   [ EDIT & APPROVE ]   [ REJECT + NOTE ]

You always see the full proposed output — not a summary. If an agent drafts a 400-word email, you read the 400-word email before you approve it. There is no "the agent will send something appropriate" — you see exactly what that something is.

The owner's control checklist (run this at onboarding)

Before you let any agent work on live customer data, run through this checklist once. It takes about thirty minutes and it is the difference between confident delegation and nervous second-guessing every morning.

1. Write your company voice in 150 words. Formal or casual? Contractions or not? First name or last name with customers? Do you ever use humour? The agents read this on every task — vague input here produces vague output forever. Be specific: "We write like a knowledgeable friend, not a corporate memo. Short sentences. No exclamation marks. We always acknowledge the customer's frustration before offering a solution."
2. List your hard constraints. These are the things an agent must never do regardless of context. Examples: never offer a discount without approval; never commit to a delivery date; never respond to a refund request with a rejection — always escalate to the owner. These go into company policy and the agents treat them as non-negotiable.
3. Set your customer data access scope. Decide which agents can see what. Your support agent needs ticket history and order data. It does not need your revenue dashboard or investor communications. Axiom lets you scope data access per agent role at setup.
4. Review the default autonomy table and move at least two rows. Do not leave defaults unchanged without thinking about them. Ask yourself: "If this agent sends this without my review and gets it wrong, what is the worst realistic outcome?" If the answer is "annoyed customer" you can probably release it. If the answer is "lost deal" or "public embarrassment", keep it gated.
5. Send yourself five test approvals before going live. Trigger five tasks manually — one support reply, one social post, one CRM note, one report, one outreach draft — and review the approval cards. You are checking two things: that the output quality is acceptable, and that you understand what context the agent used to produce it. If you cannot follow the agent's reasoning from the card, your company memory is incomplete.
6. Set a weekly audit reminder for the first month. Every Friday, spend ten minutes reading the autonomous action log — not the gated items, but everything that ran without your review. You are looking for drift: cases where the output was technically correct but slightly off-voice, slightly wrong in tone, or missing a nuance you care about. Catch drift at week two, not month four.
7. Define your escalation path for edge cases. Some customer situations should never be handled by an agent, even a well-briefed one: legal disputes, threatened chargebacks, media enquiries, anything involving personal data concerns. Write a short list of these and add them to company policy. The agents will route anything matching those patterns to your inbox as high-risk, regardless of how you have set the autonomy dials.

What Axiom does not do

Being clear about the limits is as important as describing the capabilities. Axiom agents cannot make purchases, move money, access systems outside the Axiom platform, or take actions on external websites without you explicitly connecting and authorising those integrations. They do not self-modify their instructions. They do not communicate with each other outside the shared company memory — there is no agent-to-agent chatter you cannot see. Every action is logged. You can pull an audit trail for any agent, any task, any time period.

The agents also do not get better by training on your customer data in the background or sharing anything about your business with other Axiom users. Your company memory is yours. The shared context that all your agents draw from is scoped to your account and goes nowhere else.