The AI cost calculator your vendor doesn't want you to see

Three real client engagements. Actual costs, actual returns — including the one that lost money. We're showing you all three because the vendor who only shares the wins is the vendor you should not trust with your capital.

Client A: The clear win

A professional services firm with a six-person outbound sales team. They wanted AI-assisted lead qualification. We built an agent that scored inbound leads, drafted personalised outreach, and flagged high-priority prospects for human follow-up. Build cost: £28,000. Monthly operating cost: £1,400. Monthly senior sales time recovered: 84 hours. At their billing rate, that's £12,600 per month of senior time freed for higher-value work. Payback period: three months. This is the case study vendors show you. It's real — and it is not typical.

Client B: The break-even

A retailer wanted AI-powered demand forecasting to reduce overstock and stockouts. Build cost: £45,000. Monthly operating cost: £2,800. Inventory cost reduction in year one: £38,000. Gross saving against total first-year cost of £78,600: not profitable in year one. In year two — without the build cost — clearly positive. We modelled this upfront and they proceeded, correctly: the forecast model compounds value as it learns more data. But year one on paper looks bad. Expect stakeholder questions, and answer them with the multi-year model you agreed before you started.

Client C: The loss — and the lesson

A financial services firm wanted AI-generated regulatory reports to replace a contracted compliance function. Build cost: £62,000. It worked. Then their regulator changed the reporting format. Rebuild: £22,000. Then a key data source changed schema without notice. Another £8,000. Total spend: £92,000. Expected annual saving: £30,000. Break-even: somewhere around year four. The client cancelled at month 18. The lesson: AI systems in regulated environments with frequently changing requirements carry a hidden maintenance cost that most vendors fail to model — and that we under-modelled in this engagement.

How to model it properly

We use a four-part framework before committing to any AI build. First: identify the measurable output — not 'productivity improvement' but specific hours, error rates, or revenue figures. Second: model the maintenance cost — plan for 20 to 30% of build cost per year for ongoing changes and schema updates. Third: run a sensitivity analysis — what if data quality is 20% worse than expected? What if adoption is 60% of target? Fourth: agree the success metrics before you start, in writing, with sign-off from the stakeholder who will judge the outcome.