What SMBs actually pay for AI

Real numbers from thirty engagements, including the ones where we told the client not to proceed.

Pricing in this market is opaque enough that most buyers cannot tell whether a quote is reasonable. Here is what our own engagements have actually cost, across thirty projects for companies between 20 and 250 staff.

The two-week assessment is £4,500. Roughly one in four ends with a recommendation not to build anything yet, usually because the process is about to change for other reasons or the data history is too short to be worth modelling. That fee is credited against a build if one follows.

Builds have ranged from £18,000 to £61,000, with a median around £34,000. The largest single driver of cost is not model complexity but integration surface: how many systems have to be read from and written to, and whether any of them are on-premise with a bespoke API. A well-scoped document pipeline into one modern accounting system is cheap. The same pipeline into three systems, one of which is a 2009 ERP, is not.

Ongoing support runs £1,200 to £3,800 a month depending on volume and how often the model needs retraining. About 80% of clients take it after handover; the rest run the system themselves, which is what the documentation is for.

Payback, where clients have measured it, has landed between four and fourteen months. The variance is mostly about volume: the same automation across 400 documents a day pays back in a quarter, and across 40 a day it may never justify itself. That is the arithmetic the assessment exists to do before anyone signs anything.