Every AI cost line, including the ones nobody counts as AI.
Ten classes of AI spend inventoried, sorted by cost profile, and attributed to a team, a product and a cost per unit of business work.
from £12,000 · 3 weeks · fixed fee, scope set by the size of your estate
Who this is for
Organisations that have moved AI beyond a single pilot and cannot yet produce a breakdown of what it costs. Any sector, anywhere. Most score low here even where cloud cost discipline is strong, because the assumptions do not carry across.
How it runs
Nothing to install, nobody to second, no steering group.
Find
Discovery across all ten classes of AI spend, including generative media on departmental cards, document processing billed per page, and AI seats inside software subscriptions. Endpoint and provisioned capacity inventory pulled directly from consoles.
Attribute
Spend mapped to consumers, products and use cases. Where everything runs through one API key, allocation is built by sampling and normalised against your actual invoice total.
Quantify
Cost to serve for your top use cases, waste register with values, driver-based forecast including an autonomy expansion scenario, and the readout.
What we look for
Not an exhaustive list. These are the findings that recur.
Common findings
- Dedicated endpoints and provisioned throughput running well below meaningful utilisation
- Endpoints left running from pilots that concluded months ago
- A frontier model doing work a small model would do: classification, extraction, routing
- No prompt caching where system prompts repeat on every call
- Full conversation history replayed each turn instead of summarised
- Retrieval returning far more context than the answer uses
- Development and test traffic hitting production keys and production tiers
- AI seats inside software subscriptions bought and never signed into
- Per-page and per-minute rates never renegotiated at current volume
- No spend caps or anomaly alerts on any provider account
What you get
Written so your own team can execute it. There is no phase two you have to buy.
What it costs you in time
Your organisation's total involvement. Four conversations and one data request.
What we need
- Read access to billing and usage exports
- One hour with your executive sponsor
- One hour with finance
- One hour with your platform or cloud lead
- One hour with an engineer close to the workloads
- Two weeks of application or usage logs
What we do not need
- Production system access
- Customer, member or personal data
- Software to buy, install or integrate
- An agent or collector in your estate
- A procurement exercise
- An internal project team or steering group
Questions
"Our AI spend is small. Is it worth reviewing?"+
It is usually larger than the finance view suggests, because it arrives in seven or eight places and only one of them looks like an AI invoice. If you are starting with cloud, most of your AI spend sits inside that bill anyway. And the cheapest moment to set up attribution is before it matters.
"We have one API key for everything. Can you still attribute it?"+
Yes, and this is the common case. Where there is no native attribution, spend is allocated by inventorying every consumer, sampling activity, measuring representative calls and normalising against your invoice total. Every figure is labelled high, medium or low confidence with the method shown.
"Who should own AI cost, technology or finance?"+
Both, and that is precisely the problem. The budget usually sits with finance and the data sits with engineering, so neither can answer alone. The readout is deliberately run with both in the room.
More at the questions page.
Independent means independent. Nivaan sells no cloud, no tooling, no licences and no managed services, and takes no commission from any vendor. What that means. A recommendation to keep doing exactly what you are doing is an outcome we are paid the same for.
Start with a question, not a contract.
Twenty minutes, no deck. Bring your last cloud invoice if you have it to hand.
Book a 20-minute call