AI Cost Optimisation

Most teams pick a provider at the start and never revisit it, while the price of the same operation keeps falling. The work is matching each requirement to the cheapest option that meets it: a different vendor, a smaller model, a self-hosted open-source one, or an architecture change that cuts the number of calls instead of the price per call. Prices and options change month to month, so this is worth repeating on a schedule.

This if for you if

• Your provider bill grew faster than your usage and nobody has looked at why.

• The unit economics of an AI feature do not work, so you cannot offer a free tier or a cheap plan.

• You picked a model or a vendor when you started and have not revisited the choice since.

• You are about to build something with AI in it and want the economics settled before the code.

What we do together

Measure what each AI operation costs you today, per call and per month.

Test the alternatives against your real requirement: a different vendor, a smaller model, a self-hosted open-source one.

Look at architecture as well as price. Caching, batching and routing by difficulty often cut more than switching providers does.

Consolidate what is worth self-hosting onto shared infrastructure instead of paying per call.

Set a review cadence, because prices and options change month to month.

How it works

1. Discovery calls, free. Up to three, ending with a proposal and a price.

2. Cost audit. €1,500 for a single product or pipeline, €3,000 where several systems are in scope. Current spend broken down by operation, alternatives tested against your real requirement and priced side by side, with the trade-offs named.

3. Implementation. Priced whole against milestones once the audit has set the scope. Replace, re-architect and measure, so you can see the saving in your own bill.

4. Ongoing review. Optional retainer from €2,000 a month. Prices and options change month to month, and someone has to keep looking.

What you get

Your current AI spend broken down by operation, per call and per month.

Alternatives tested against your real requirement, priced side by side, with the trade-offs named.

The replacements implemented and measured against your previous spend.

A review cadence and the method behind it. Your team can run it, or I keep checking prices and options against your usage each cycle.

Proof

Modelled the unit economics of an AI product before it was built, then engineered to hit them: a paid attention-prediction API charged at $0.80 per image against a $1.50 selling price was replaced with an open-source model, dropping marginal cost to effectively zero. That made free trial analyses viable and the product cheap enough to sell to a mass market.

Replaced a per-minute video analysis service the same way.

Consolidated four models onto a single GPU server at a fixed $200 a month, shared across emotion recognition, gaze tracking, moderation and zone detection.