A well-run AI build starts with someone studying how your business works before quoting, then gives you a written scope and price, a build inside your own accounts, testing you can see and sign off, a staged launch and a clear plan for after launch. You should own the code, the accounts and the data. If a builder skips any of those steps, treat it as a warning.
Do you need a builder at all?
Start here, because the honest answer is sometimes no. If an off-the-shelf chatbot or your CRM's built-in AI features cover what you need, use them. They're cheaper, and you can switch them off.
A custom build earns its cost when the AI has to work across your own systems: reading your booking calendar, updating your CRM, following your qualification rules, and handing over to your team in a specific way. We set out where that line sits in custom AI chatbot vs off-the-shelf.
What each stage should look like
Every builder uses different names, but a sound project moves through the same stages. Here's what you should receive at each one, and what should worry you.
| Stage | What you should get | Red flag |
|---|---|---|
| Audit | A map of how enquiries arrive and where work gets stuck, with the cost of each problem | A quote before anyone has asked how your business works |
| Design | A written scope: what is built, what isn't, the systems it connects to, who can do what, and how success is measured | "An AI solution" with no list of what it does |
| Build | Regular short updates, in your own accounts | Silence for weeks, then a demo on their account |
| Test | Real and awkward cases run before any customer sees it, with results you sign off | "We tested it" with nothing to show |
| Launch | A staged rollout, team training and written documentation | Everything switched on at once, with no handover |
| Improve | A review of real conversations and numbers, with agreed next changes | No plan for anything after go-live |
This is the process we follow. Testing is the easiest stage to rush under deadline pressure, which is why it gets its own section below.
What it should cost, and what you pay separately
Expect two kinds of cost, and make sure the proposal separates them.
The build. A fixed price for the agreed scope. As a reference, our paid audit costs £750 to £1,500 and is credited against an agreed sprint, and a bounded 14-day sprint costs £5,000 to £20,000. Larger custom projects are scoped separately. Be cautious about any builder who promises every custom project in 14 days.
Running costs. AI usage, hosting, phone minutes and software subscriptions. These should be paid directly to the providers from your own accounts, so you can see them and nobody marks them up. Ask for a monthly estimate at your expected volume before you sign.
Ongoing support is a third, optional cost. Ours is a monthly retainer from £1,000, with a three-month minimum and then 30 days' notice. Whoever you hire, get the terms in writing.
What you should own
Before you sign, get these in writing. If the answer to any of them is "we keep that", ask why.
- The code sits in a repository in your name, or you receive it in full
- The AI, hosting and phone accounts are in your name, with your payment details
- Your data, including every conversation, is stored in your accounts and can be exported
- You hold admin access to every system, even if the builder also has access
- The approved information the AI answers from is written down and editable by you
- Documentation explains every connection, setting and account
Ownership matters most on the day you want to change supplier. If everything already sits in your accounts, that's an email. If it sits in theirs, it's a rebuild.
Questions to ask before you sign
Good builders answer these clearly and without hesitation.
| Question | A good answer sounds like |
|---|---|
| What happens when the AI doesn't know the answer? | It says so and hands the conversation to a named person, with the context attached |
| How will you test it? | A written list of cases, including awkward ones, that you review and sign off |
| What can it never do? | A specific list: no advice in regulated areas, no prices outside the approved list, no complaint handling |
| What does it cost to run each month? | An estimate at your volume, paid to the providers from your accounts |
| Who owns the code and the data? | You do, in your accounts |
| Can I see something similar you've built? | A live example or a case study with specifics |
The first question matters most. The builder is building something that speaks in your name, and you remain responsible for what it says. In Moffatt v Air Canada (2024), a Canadian tribunal held the airline liable for incorrect fare information its website chatbot gave a customer, and rejected the argument that the chatbot was responsible for its own actions.
What you'll need to provide during the build
AI projects slow down when the builder is waiting on access or decisions. Have these ready before the build starts:
- Access to the systems it connects to, such as your website, booking system, CRM and inbox.
- The approved information it answers from: services, published prices, policies, opening times and common questions with your preferred answers.
- One person who can answer questions and approve decisions within a day.
- Real examples of enquiries, including the difficult ones, to test against.
Testing you should see and sign off
Before any customer uses the system, you should see it handle real cases: an angry customer, a question it mustn't answer, a request to stop contacting them, an enquiry at 3am, someone asking whether they're talking to a person. Ask for the test list, the results and what was changed as a result. We describe a practical test plan in how to test an AI agent before it talks to a customer.
After launch
AI systems aren't set and forget. Your prices change, your services change, the underlying models change, and real conversations always show gaps the tests missed. Decide before launch who reads the conversations each week, who updates the approved information and who fixes problems, whether that's your team or the builder on a retainer.
Do this week
Write one page describing the problem you want solved: where enquiries come from, what happens to them now, where they get stuck and what it costs you. Take that page to any builder you're considering and see whether their questions go deeper than yours. The ones that do are worth a second conversation.
If you'd like that conversation with us, book a call. We'll start with the audit questions, not a pitch.
