You've been using ChatGPT or Claude for a while. You can see the obvious time savings. But you've also noticed something: the copy-paste-a-prompt routine only gets you so far. The real value comes when AI does a specific job for your business, reliably, in the background, using your data, in your voice, without you babysitting it.

That's a custom AI tool. And if you've never commissioned one before, the whole process probably feels like a black box — how do you scope it, how do you cost it, who do you hire, what can go wrong?

This guide walks you through every stage. Written for non-technical founders and operators of businesses under a hundred people. No jargon. Real steps you can act on this week.

Step 1 — Find the process worth building around

The single biggest mistake in custom AI builds is building the wrong thing. Not a technical failure — a scoping failure. You build something clever that nobody uses.

The fix is to start from a process, not from a technology.

A "process" is a thing that happens in your business regularly — weekly or more often — that takes meaningful time and has a clear output. Writing a proposal. Qualifying a lead. Reconciling an invoice. Briefing a salesperson before a call. Generating a Monday-morning report.

The test for a build-worthy process is four yes-or-no questions:

Is it repeating? If it happens once a month or once a quarter, the ROI rarely justifies a build. Weekly or more.
Is it documented, even roughly? If you can't describe the steps someone else would follow, no AI can do it either. If it only lives in your head, write it down first.
Does it produce a clear output? A document, a record, a report, a decision. AI thrives on producing things. It struggles with ambiguous "let's just think about this."
Is the cost of a mistake survivable? Early builds work best when an occasional error is caught and fixed. If the process touches financial transfers, regulatory filings, or medical decisions, you need a different approach.

Four yeses, it's a candidate.

Step 2 — Scope it as a "job," not a "system"

The second biggest mistake is over-scoping. Non-technical buyers assume they need to specify everything, so they write 40-page requirement docs that collapse under their own weight.

The right scoping exercise fits on one page. Three sections.

The job. A single sentence describing what the tool does. "Drafts personalised sales follow-ups from call transcripts." "Flags customer accounts likely to churn in the next 60 days." "Generates a Monday-morning briefing from last week's pipeline and support data."

The inputs. Where the tool gets its raw material. Your Gmail. A specific Google Drive folder. Your CRM. A Slack channel. Your meeting transcripts. Be specific.

The output. Where the tool puts the result. A draft email in your inbox. A Slack post at 8am. A new row in a Google Sheet. An email to you and two colleagues. Be specific.

If you can fill in those three sections, you can get a quote. If you can't, talk to a build shop that does a scoping session before quoting — that's a normal and often free first conversation.

Step 3 — Decide build vs. buy vs. DIY

Three paths, very different economics.

Buy an off-the-shelf tool. If a well-reviewed SaaS exists that does exactly the job, buy it. Right answer about 30% of the time. The limitation is that SaaS is built for the median user — your business is specific, and the tool won't be shaped to your actual workflow.

DIY with no-code tools. Using tools like n8n, Make, or Zapier combined with Claude or ChatGPT, a non-developer can build surprisingly capable workflows in a weekend. Right answer about 25% of the time. Works when the workflow is simple, when you've got some technical curiosity, and when the cost of it breaking is low (because you'll have to maintain it).

Commission a custom build. A specialist builds it for you, hands it over working, documents it, and maintains it on a light retainer. The right answer the remaining 45% of the time. Higher upfront cost, but the tool is shaped to your business, reliable, and supported.

The breakeven calculation: if the tool will save a person in your business more than 5 hours a week, a custom build almost always pays back within 6–9 months. Less than 2 hours a week, DIY or buy first.

Step 4 — Get realistic about cost

Custom AI tools for small businesses sit in a different pricing band than enterprise "AI transformation" projects. Honest lay of the land in 2026:

Simple internal tool, one workflow, one data source. £3,000–£6,000. Shipped in 1–2 weeks. Example: a proposal drafter fed by past proposals and call notes.
Smart website tool that customers interact with. £5,000–£9,000. Shipped in 2–3 weeks. Example: a conversational lead qualifier that replaces your contact form.
Internal dashboard pulling from multiple data sources. £4,000–£7,000. Shipped in 2 weeks. Example: a weekly briefing dashboard pulling from CRM, support tool, and analytics.
Staff training portal. £6,000–£9,000. Shipped in 3 weeks. Example: an onboarding assistant that knows your SOPs and can answer any new hire's question.
Automated background workflow with agents. £5,000–£9,000. Shipped in 2–3 weeks. Example: a scheduled agent that processes inbound documents, extracts key data, and files them in the right places.

Anything quoted significantly above these numbers for a small business use case — and you'll see quotes from larger agencies at 2–5x — is usually pricing their overhead into your project, not pricing the actual work.

Step 5 — Choose a builder, not a consultant

The commercial AI market has two very different kinds of supplier. Make sure you're talking to the right one.

Consultants sell strategy. They run workshops, produce roadmap decks, identify opportunities, and hand you a plan. They don't usually ship working software. For a business under a hundred people, this is almost never what you need — the strategy step is worth at most a two-hour conversation, not a six-figure engagement.

Builders ship. They scope, build, and deliver working tools. They'll spend thirty minutes understanding your process, give you a fixed quote, and start building. You'll have something running within three weeks.

The four questions to ask any prospective builder:

Can you show me something you shipped in the last 60 days?
What does the handover look like — do I own the tool or rent access?
If I want to take this in-house later, what's the process?
Can you quote a fixed scope and fixed price?

Hesitation on any of those four is a warning sign. A legitimate builder can answer all four inside five minutes.

Step 6 — Scope the first build small

Even when you've found the right builder, resist the urge to make your first build ambitious. First builds should:

Solve one clear job.
Use one or two data sources, not eight.
Ship in under three weeks.
Be something you'd use personally, or that one specific person in your business would use daily.

The reason is learning. The first build teaches you how AI tools fit into your business, what maintenance looks like, what breaks, what surprises you about adoption. That learning is worth more than any specification document. Your second build will be ten times better scoped because of it.

Step 7 — Plan for maintenance

AI tools are not buy-and-forget software. The underlying models change (usually for the better, occasionally with awkward behaviour shifts). The integrations you depend on get updated. Your business changes, and the tool needs to follow.

A realistic maintenance budget is 10–20% of the original build cost per year. For a £6,000 build, £600–£1,200/year keeps it working. Most builders will offer this as a small monthly retainer — £50–£100/month — and it's almost always worth paying.

The alternative — buying the build and trying to maintain it yourself with no technical support — usually results in the tool slowly drifting into brokenness over 6–18 months, at which point you either pay to fix it or let it die. Cheaper to maintain from day one.

Step 8 — Set up the feedback loop

The last step, and the most skipped one. After the tool ships, for the first month, set a 10-minute weekly slot on your calendar to check: did it do the job this week? What did it get wrong? What did you manually correct?

After one month, revisit the build with your builder. Most problems surface in week two or three — a tone issue in drafted emails, an edge case in the data, a workflow step that needs refining. A 20-minute tune-up after a month usually takes the tool from "useful" to "essential."

After three months, if you're using it daily and it's saving the hours it was scoped to save, you've shipped a real thing. Most business software doesn't cross that bar. Custom AI tools, scoped this way, do.

What could go wrong — and how to avoid it

Three common failure modes:

"It hallucinates." The tool makes things up. Happens most in tools that try to do too much without enough grounding in your data. Fix: narrower scope, more specific data sources, a QA check in the workflow.

"Nobody uses it." The tool works, but the team ignores it. Usually because the workflow it replaces didn't hurt enough, or the tool adds a step rather than removing one. Fix: build for a process that genuinely consumes time, and integrate the tool into where people already work (Slack, email, inbox) rather than a new interface.

"The model changed and it broke." The AI provider updates the underlying model and the tool behaves differently. Fix: a maintenance retainer with a builder who'll catch this and adjust prompts or logic within a week of it happening.

None of these are reasons not to commission a build. They're reasons to commission carefully, with the right supplier, and to budget maintenance from day one.

Want to keep reading?

The 10 AI tools worth paying attention to in 2026 — what your team should already be on before you commission a build.
The AI Chief of Staff — when one build turns into six and you've quietly built yourself a senior operator.
Want to see what fixed-scope custom builds look like in production? The Buyer Intelligence, Anti-Ad Builder, and Uncomfortable Truths tools on this site are all examples of the same pattern — narrow scope, clear output, shipped in weeks.