What to delegate to AI in a 1-5 person business (and what not to)

For a 1-5 person business in 2026, AI consistently pays back on six task classes and consistently fails on six others. The line is not capability. The line is whether the customer is buying your output or buying you. The 90-second test below answers it before you sign up for any new agent or workflow.

Most “AI for solo founders” content covers what AI can do. The list grows every quarter, and the answer is increasingly “more than you think.” That framing has stopped being useful. The harder question, and the one that actually determines whether delegation pays back at owner-operator scale, is which task classes survive contact with your customer trust contract.

This piece is the delegation framework the rest of the operators register links back to. Pillar reading. If you are running a 1-5 person business and you have ever asked yourself “should I have AI do this,” start here.

Why “what to delegate” is the wrong question to ask first

The reflex when a new agent or workflow lands in your feed is to ask “can AI do X.” Almost always the answer is yes, technically. Drafting? Yes. Summarising a thread? Yes. Writing a refund email? Yes. Negotiating with a vendor? Technically yes. Drafting your own marketing post? Yes.

The reflex is wrong because it ranks tasks by AI capability instead of by customer trust contract. In a 1-5 person business, the trust contract IS the differentiator. The customer is not buying your output the way they buy from a 5,000-person SaaS company; they are buying the fact that you, specifically, are running this. Substitute AI at the surface where the customer expects you, and you have not gained productivity. You have spent a brand asset.

The reframe: AI delegation is a brand decision before it is a productivity decision. The right first question is not “can AI do this” but “does delegating this to AI fit my trust contract with the people who pay me.” The capability question is downstream. Start with the brand question and the capability question answers itself most of the time.

The corollary, useful for founders who are over-indexed on either side: AI is not categorically dangerous, and AI is not categorically a productivity bonanza. Six task classes pay back reliably at this scale. Six others reliably do not. The rest of this piece is the two lists and the test that distinguishes them on a new task you have not seen yet.

The six task classes AI does well in 1-5 person businesses

These are the surfaces where the trust contract is intact, the failure mode is recoverable, and the productivity gain is real and compounding. Treat the list as a default-on register: unless one of the six categories trips a specific brand consideration in your business, delegate.

1. Drafting. Emails, listing copy, proposals, contract first drafts, follow-up sequences, replies to inbound. The pattern is identical across all of them: AI produces the first draft, you edit, you send. The customer reads what you sent under your name, with your judgement embedded in the edits. The productivity gain is the time-to-first-draft, not the published output. ChatGPT Plus and Claude Pro cover this surface well at consumer-tier subscription pricing.

2. Summarising. Long documents you need to triage, transcripts of calls you ran (or missed), customer threads spanning weeks, vendor proposals that arrive as 40-page PDFs. AI summarisation has been reliable since 2023; in 2026 the only meaningful failure mode is hallucinated specifics, which the editing pass catches. Treat summaries as input to your judgement, not as the judgement.

3. Scheduling and calendar coordination. Calendly handles the booking surface; Reclaim AI (or equivalents like Motion) handles the prioritisation and rescheduling. The combination removes the coordination overhead that historically eats two to four hours per week at owner-operator scale. The trust contract is unaffected because nobody expects you, personally, to be running calendar Tetris.

4. Research synthesis. Competitive scans, sourcing comparisons across vendors, market sizing for a new offer, regulatory landscape reads. AI does the corpus assembly and the first synthesis pass; you do the verification on the load-bearing claims and you do the strategic read. The productivity multiple here is the highest of the six categories: tasks that used to take a half-day Saturday compress to thirty minutes plus a verification pass.

5. Code generation for solo developers. Cursor, Claude Code, and GitHub Copilot have moved from “experimental” to default tooling for solo developers building production code. The trust contract here is internal, not customer-facing: the code ships under your responsibility regardless of who or what wrote the first version. Stack-aware completions, multi-file refactors, and test generation are the load-bearing wins; the failure mode is silent introduction of subtle bugs, which the test pass catches.

6. Image and asset production. Midjourney, Adobe Firefly, and DALL-E (via ChatGPT Plus) cover the surface where you need backgrounds, hero images, social-post visuals, and stock-equivalent imagery without a stock subscription. The trust contract is intact because nobody assumes a 3-person business has an in-house illustrator. The exception is the brand-distinctive creative output that defines your visual identity, which belongs in the second list below.

The six task classes AI fails on in 1-5 person businesses

These are the surfaces where the trust contract is the artefact, the failure mode is irreversible at owner-operator scale, or both. Treat the list as default-off: unless you have done the explicit work to engineer disclosure or human-in-the-loop, do not delegate.

1. High-stakes customer-facing decisions without disclosure. Firing a client. Refunds above your judgement threshold (the number varies by business; €500 is a useful default for service businesses, higher for product). Sensitive complaints where a customer is angry, hurt, or both. The failure mode is not output quality. It is the relationship cost on discovery. Run the five-category piece on when not to substitute AI for the cited evidence base.

2. Regulatory or legal advice. Contract terms, tax structuring, employment law, GDPR posture, EU AI Act compliance. AI is acceptable for first-pass research and for drafting a question you will then take to a licensed professional. AI is not acceptable as the answer. The cited consequence on any of these surfaces, when the recipient acts on AI output as if it were licensed advice, is disproportionate to the productivity saved.

3. Complex multi-party negotiations. Vendor pricing across three or more counterparties, partnership terms with revenue-sharing implications, anything where the dynamics involve reading the room across multiple humans. AI can prepare you (research the counterparty, draft openers, score scenarios). AI cannot run the negotiation. The information bandwidth, the relational cues, and the strategic patience that complex negotiations require are surfaces where 2026 models still degrade quickly.

4. Brand-distinctive creative work. The 10% of your output that defines you. The signature framework. The voice fingerprint. The visual asset that anchors the brand. AI can produce competent generic work in any of these registers; competent generic is the opposite of brand-distinctive. The risk is not that AI will produce bad output. The risk is that AI will produce indistinguishable output, and indistinguishable output is the wrong product for a 1-5 person business that competes on differentiation.

5. Anything requiring physical presence. Shipping and fulfilment QC, manufacturing inspection, in-person sales calls, site visits, hardware diagnostics. Obvious, but worth listing because solo founders sometimes try to substitute AI assistants at the back-office layer of these surfaces in ways that introduce drift. The frame: AI handles the planning around physical work; AI does not handle the physical work.

6. Trust-of-authenticity work. Testimonials. Customer reviews you write or solicit. Founder posts on social. Anything where the recipient is consuming the artefact specifically because a human wrote it. EU AI Act Article 50 makes the transparency obligation explicit for some of these surfaces; the broader principle is that authenticity is the artefact, and synthetic output, even competent synthetic output, is the absence of the artefact. Disclosure helps where it is feasible. On founder posts and testimonials, disclosure usually defeats the purpose.

The 90-second test before you delegate any new task

When a task arrives that is not on either list above, run three questions in ninety seconds.

(a) If AI gets it wrong, what is the worst-case cost? Recoverable in under a day at low cost: delegation candidate. Irreversible (lost client, regulatory breach, public-facing reputational damage): not a delegation candidate without human review at the output.

(b) Does the customer expect a human authored this? No or neutral: delegation candidate. Yes, and disclosure breaks the implicit contract: not a delegation candidate.

(c) Is upfront disclosure feasible without breaking the trust contract? Yes (internal drafting, research synthesis, background imagery): delegation candidate. No (founder posts, testimonials, complaint resolution): treat the task as not delegation-ready until you can engineer disclosure that works.

Score: two answers favouring delegation, delegate. One, delegate with material human review at the output. Zero, hold the task in human-only mode until the cost surface, the authorship expectation, or the disclosure feasibility changes.

The test takes ninety seconds because longer tests get skipped. Three questions on a Post-it covers the same ground a 30-row delegation matrix would at this scale.

The disclosure rule

The rule that ties the two lists together: when AI does customer-facing work without disclosure, the trust cost compounds the longer the substitution runs before discovery. When disclosure is upfront, the customer self-selects in or out, and the relationship survives whichever way they go.

Disclosure does not mean “AI was used at some point in this workflow.” That framing dilutes to meaninglessness. Disclosure means specifying which surface AI touched and which surface a human owned. “First draft was AI-generated; I edited and sent” is a disclosure. “We use AI to summarise call transcripts; I read the summary, you talk to me” is a disclosure. “AI-assisted” is not a disclosure.

The legal floor in 2026 is EU AI Act Article 50 transparency obligations for some interaction surfaces. The editorial floor is higher: anywhere the customer would feel deceived if they later discovered the substitution, disclosure beats the alternative. The customer who self-selects out of an AI-disclosed surface is a customer who would have churned painfully on discovery anyway. Disclosure is the cheapest version of that conversation.

Related operators reading

For the inverse list with cited cases of substitution-cost-more-than-it-saves, see when not to use AI for your small business. For the four-question filter once a use case clears the delegation framework, see picking your first AI agent. For the policy artefact that codifies the framework across your team, see the 1-page AI policy. For solo-founder customer service stack specifics, see the customer service AI stack. For the foundation tooling decision underneath all of this, see Claude Pro vs ChatGPT Plus for the solo founder. The Holding-up record for this claim is at /holding/?claim=OPS-061.

What changes this list

Cadence is 45 days because two of the six “fails on” categories sit on regulatory surfaces (legal advice, transparency obligations) that move on quarterly enforcement cycles. Three conditions would move this claim to Partial.

• A model capability shift makes one of the six fail-on categories pass the 90-second test reliably. The likeliest candidate is multi-party negotiation, where 2026 frontier models are visibly improving on long-horizon strategic patience.
• An SMB-scale liability insurance market emerges that prices the trust-failure risk on customer-facing AI substitution. Today the risk sits on the operator’s balance sheet; insurance would convert several decisions from absolute to price comparison.
• A regulatory shift codifies disclosure formats that make AI substitution acceptable on surfaces where it currently is not. EU AI Act Article 50 enforcement guidance is the load-bearing surface to watch.

We will re-test this framework against actual SMB outcomes on or before 21 Jun 2026. If any of the three conditions has triggered, this claim moves to Partial.