# What AI actually costs

Source: https://tpiros.dev/ai-cost

A taught half-day session for technology leadership teams on where AI cost comes from, which levers move it, and how to challenge a cost estimate before the decision is made.

Remote, half a day, one company's leadership team at a time (up to fifteen people). No pre-work: nothing to prepare, no data to send, no access to any system.

## Why it exists

By the time an AI invoice looks wrong, the architecture that produced it has been in place for two quarters. The decisions that set the number were made early, in a design review, by people who were not asked to think about cost. Reading the invoice more carefully does not help; recognising those decisions while they are still open does.

Two findings frame the session. The FinOps Foundation's State of FinOps 2026, covering 1,192 practitioners, found that 98% now manage AI spend, up from 63% in 2025 and 31% in 2024. KPMG's Global AI Pulse for Q2 2026, covering 2,145 senior leaders across 20 markets, found only 7% report an established return on AI, and that leaders with strong cost visibility are five times more likely to be among them (15% against 3%). Both figures are taken from the publishers' own reporting rather than secondary summaries.

## What it covers

- Where the number actually comes from: tokens in and out, why the reply costs more than the question, what context length does to a bill that looked fine in the pilot.
- The shape of a workload: taking common AI features apart step by step and finding where the money concentrates, which is rarely where people assume.
- The levers and their price: routing between models, caching, retrieval, batching, smaller models with checks. Each one trades cost against quality, latency or engineering time.
- The questions that expose a weak number: what happens at ten times the volume, what has been left out, what is measured, what breaks when the model is replaced.

## What a leadership team can do afterwards

- Read a cost estimate for an AI feature and say where it will be wrong.
- Ask a vendor the questions their pricing page is built to avoid.
- Tell the difference between a pilot that is cheap and a system that will be.
- Judge whether a proposed AI feature is worth building, before it is built.
- Explain to a board why the spend moved.

## Who it is not for

Organisations that have not started building with AI, engineering teams who want to implement the optimisations themselves (better served by a programme at https://tpiros.dev/programmes), and anyone looking for a vendor recommendation. The session is provider-agnostic.

## Before buying

A free forty-five minute online briefing runs monthly: thirty minutes taught, then fifteen minutes of questions. Registration is on the page; attendance is capped.

## Next step

- Register for the free briefing: https://tpiros.dev/ai-cost
- Email hello@tamaspiros.com with who is in the room and what they are being asked to approve.