How I scope an AI build (and when I tell you not to bother)

This is the method I use on paid work. The questions I ask on a first call, the criteria I use to kill projects, and the point where I tell someone to keep their money. I am publishing it because knowing the steps was never the valuable part. The moat is doing the work, not knowing the steps. Anyone can read what follows in ten minutes. Far fewer people will sit with a process owner for two hours, clean five years of inconsistent customer records, or chase down why the calendar sync fails on the last Friday of the month. So here are the keys. Use them yourself with my blessing, or read on and recognise the questions when I ask you them.

One thing before we start: roughly half the AI projects I get asked to quote should not exist. The method below is how I work that out, and it works just as well when you run it on yourself.

The five questions I ask before anything gets built

These run in order, and a bad answer at any step ends the conversation. That is the point of them.

1. What is the repetitive, expensive thing?

I want a named task you already pay for in staff hours. Quoting. Chasing invoices. Turning enquiries into confirmed bookings. Writing the same screening notes forty times a week.

If the answer starts with "we want to use AI to", I stop. The expensive AI projects are the ones that start with the word AI and go looking for a problem. The cheap ones start with a task someone is sick of doing, and the budget conversation almost answers itself from there.

2. How often does it happen?

A rough weekly count, honestly estimated. Daily and weekly tasks justify systems. A task that happens four times a year almost never does, however painful those four occasions are, because the build cost is the same and the payback is a quarter of the payback on a weekly task. When the frequency is low I will usually suggest a checklist and move on. Checklists are free.

3. What does an error cost?

This is the question most people have never been asked, and it drives more of the price than anything else. An automation that drafts internal notes can fail safely: someone reads it, winces, fixes it. One that sends payment links to bereaved customers, or commits you to a contract, cannot. The second kind needs review steps, fallbacks and alerting, and those guardrails are real engineering work, not paranoia.

If the error cost is high and the appetite for review steps is zero, I walk. Not because it cannot be built, but because it should not be.

4. Where does the data live?

Count the systems the build has to read from or write to. CRM, calendar, payment provider, accounting software, phone system. Each one adds integration work, and ageing or niche software adds more than modern tools with good APIs. The AI itself is the cheap bit; model access costs pennies per run at small business volumes. The money goes on the plumbing.

Then the harder part: what state is the data in? If your customer records are inconsistent, duplicated or spread across spreadsheets, cleaning them up becomes part of the project whether anyone names it in the quote or not. Better it gets named.

5. Who owns the process?

I need one person who can describe the process end to end, including the ugly exceptions, and who will answer questions while I build. Not a committee, not "ask whoever is on shift". If nobody owns the process, the project does not have a missing stakeholder, it has a missing process, and no amount of software fixes that. No owner, no build. I have never regretted enforcing this and I have regretted every time I let it slide.

When I tell you not to bother

Three honest exits, and I take each of them regularly.

A spreadsheet wins when the task is low volume, one person owns it, the work is genuinely exploratory, or the process is still changing shape every month. Building software around a process that has not settled means rebuilding it twice. Spreadsheets are for thinking; systems are for operating. If you are still thinking, keep the spreadsheet.

Off-the-shelf wins when your painful task is one that thousands of businesses share. Common problems get products, and products are cheaper than builds every single time. I run Bidwell, an AI tender-writing product, precisely because drafting bid responses is the same structured job for every small firm that does it, so it belongs in a subscription, not a per-client build. The test I hold myself to in the cost guide applies here word for word: if you cannot articulate why a subscription or a platform assembly fails for your case, you are not ready to commission a build. That sentence has saved my clients more money than anything I have ever built them.

Nothing wins when the problem is a process problem or a people problem wearing an AI costume. Automating a broken process gives you the same broken process running faster, which is worse, because now it breaks at machine speed and nobody notices for a week. If two departments disagree about who handles an enquiry, software will not referee that. Fix the disagreement, then call me.

What it costs, honestly

I keep the full numbers, with sources, in the cost guide, so here is the short version. Assistant subscriptions run roughly £14 to £25 per person per month: Microsoft 365 Copilot Business is £19.32 per user per month on monthly billing, excluding VAT. The platform tier, wiring AI into the tools you already own, is a small monthly subscription plus days of setup work per workflow. Bespoke builds start in the low thousands and rise with the number of systems they touch.

The most useful benchmark I know comes from UK government research: businesses adopting AI spent a median of £2,000 on it in 2024, against a mean of £19,000. Most adopters spend a little; a small minority make a deliberate, larger bet. If your first quote is ten times the median, the burden of proof is on the quote, not on you.

Where this goes wrong

Even with the five questions answered well, the same failures keep turning up. Watch for these.

Scoping the demo, not the exceptions. The happy path is always quick. The expense lives in the customer who replies in an unexpected format, the third-party API that goes down, the two bookings that collide. If a quote does not mention edge cases, the edge cases are coming out of your patience later.

Pretending the system is unattended. Government research found 84% of businesses using AI maintain at least some human checking. That review time is a genuine cost and a worthwhile one. Price it in from day one rather than discovering it in month two.

No maintenance plan. Models change, APIs change, your business changes. A build with no maintenance plan is a build with a hidden second invoice. Ask who fixes it when it breaks at 7am on a Saturday, and get the answer in writing.

Letting the vendor scope it. Vendors price what they sell. A software company will tell you a subscription covers it; an agency will tell you it needs a build. Neither is lying, and neither is neutral. Run the five questions yourself before you talk to anyone who profits from the answer, including me.

Expecting revenue instead of time. Among UK businesses that adopted AI, 75% reported workforce productivity improvements, while most saw no direct revenue change. This technology earns its keep by removing cost and time before it adds sales. Scope and measure accordingly, or you will judge a working system a failure.

What it takes if you do it yourself

Now you have the method, here is what using it costs. Mapping one process honestly takes hours of unglamorous conversation with the person who runs it, and they will remember exceptions in week three that they swore did not exist in week one. Building the workflow is days of focused work, then more days testing it against real data rather than the tidy examples everyone reaches for. You need enough technical confidence to wire systems together, enough patience to handle the edge cases as they surface, and enough stamina to keep maintaining the thing after the novelty wears off. None of it is beyond a capable operations person. All of it is work, which is why most people who read this will nod and do nothing, and why I am relaxed about giving it away.

If you would rather skip straight to the part where it runs, book a consultation and I will run this exact method on your business, including telling you if the honest answer is a spreadsheet.