Realistic Numbers for an AI Build

Build cost by scope

Ranges are wide because integration complexity dominates, and that is knowable only after looking at your systems.

Where the build money goes

• Integration and data plumbing — frequently 40%
• The AI pipeline itself — 20–25%
• Review and correction interface — 15–25%
• Evaluation, monitoring and failure handling — 15%

The AI is rarely the largest line. Proposals where it is are usually missing the interface or the operational work.

Running cost

Model usage for a typical business application runs £80–£600 a month; hosting adds £40–£200. Higher if you are processing tens of thousands of documents, lower than most people fear otherwise.

We model it against your actual volumes before you commit, and build in a hard cap so a bug cannot produce a surprise.

Maintenance is the line people forget

Twenty to thirty per cent of build cost annually, higher than conventional software because of quality drift. That covers evaluation reruns, threshold tuning, prompt adjustments after provider changes and test set refresh.

Budget it explicitly. Unbudgeted maintenance means an unmaintained system, which means a system nobody trusts within a year.

What makes a quote go up

1. Systems without APIs, needing import routes or interface automation
2. Documentation that must be organised before it can be used
3. Regulatory requirements needing audit trails and retention controls
4. Multiple document types or languages at launch rather than phased

Frequently asked questions