A worked example
| Line | Year 1 | Years 2–3 each |
|---|---|---|
| Build | £18,000 | — |
| Model usage | £3,000 | £3,300 |
| Hosting and infrastructure | £1,200 | £1,200 |
| Maintenance | £4,500 | £4,500 |
| Your team's time | £2,500 | £1,200 |
| Total | £29,200 | £10,200 |
Three-year total: about £49,600, of which the build is 36%. That ratio surprises people who budget only for the build.
The line nobody counts
Your own team's time — the evaluation set, the interface review, the ongoing quality ownership — is real cost. Leaving it out does not make it free; it makes the business case wrong.
Two to four days of a domain expert in year one, then a few hours a month thereafter.
What makes each line move
- Model usage: context size more than volume
- Hosting: whether you need high availability
- Maintenance: how variable your inputs are
- Team time: how clearly correct is defined at the start
Compare against the right alternative
The comparison is not against zero. It is against three more years of the current process, including its error costs and the capacity it consumes.
Frequently the current process costs more than the integration over the same period, and nobody has ever added it up.
Where costs are usually understated
- Maintenance, treated as optional
- The integration work, when systems have no API
- Content or data preparation before anything can be built
- The tuning period after launch, before quality settles