What should your first AI budget buy?

Picture a regional distributor whose service team spends part of each day putting together order-status replies. For this example, the first AI project drafts a response from approved order data; a colleague still sends it. That is a clear job to buy. Refunds, warehouse changes and every other customer channel can stay outside it. Write down both the job and its limits before asking agencies for prices.

Put the request, the records it may use and the expected reply on a simple scope sheet. Name the person who checks the answer and the cases that go back to a colleague. Attach examples you have permission to share. If the order record cannot answer a question, the draft should say so. Giving every agency the same examples makes their answers easier to compare and exposes different assumptions before they become extra work.

Brainpool describes fixed-price work to check project feasibility. Fifty One Degrees publishes a typical proof-of-concept offer below £15,000 and describes production deployment separately.^1,2,3

Those are different points in a purchase. You might need a study because nobody knows whether the available data can support the task. A prototype lets you inspect an approach that works on examples. A service used by staff also needs agreed access, recovery, supervision and support. A low entry price or a short quoted timetable does not tell you that all of that is included. Ask what will be usable when the paid stage ends.

For the distributor, a study could assess a sample of order records, explain where they fail and recommend whether to proceed. A running service would also need an approved system connection and a queue staff can review. Ask each bidder which result it is selling, what you keep if you stop and what you must buy next. A quote that does not answer those questions is not yet clear enough to approve.

Nexus tells us it charges per role automated, without day rates or consultant-time charges. Its public commercial material describes a proof of concept before an annual commitment.^4,5

Whatever the pricing model, ask for the same cost headings: the build, data preparation, connected services, running costs, support, changes and handing the work back. Add your staff’s time separately. With a role-based offer, define the tasks within that role and which changes make the price rise. With a project fee, ask what triggers an extra charge. With a value-linked fee, agree what evidence makes payment due and the most you could owe.

Put a quoted amount or “still to be agreed” against each heading. A blank price is not zero. Use your own records to compare an ordinary workload with a plausible busy period. The distributor should include repeated requests and the time colleagues spend checking drafts, as well as replies that work first time. Keep forecast savings separate from the cash needed to pay the agency. Freeing staff time does not automatically reduce the wage bill.

The approval note can be short, but it needs the business problem, what you are buying, the maximum commitment and the work your team must do. Add the evidence needed to accept the result and the next spending decision. For the distributor, the question is whether checked drafts reduce total handling effort without unacceptable errors. Start with a measure collected by the business and targets its owner agrees. There is no universal success percentage in this guide.

Approve the next purchase when the agreed evidence is available. If access is blocked, establish whether someone can fix that prerequisite or whether the approach itself has failed. If the result cannot be used, record what you learnt and whether further work deserves funding. An existing prototype is not, on its own, a reason to keep spending. Someone who missed the sales meeting should be able to read the note and understand what arrives, what it costs and when to look again.