The number on the vendor's quote is the cheapest part of any AI system you will ever run, and everything else lives below the waterline. Once it is in production, integrated, governed and maintained, the licence or token cost is typically a small share of what it costs you over three years, in our founder's experience 8 to 15 per cent.
This is why so many AI business cases fall apart eighteen months in. Independent surveys through 2026 keep finding that most leaders have yet to see material revenue gains or cost savings from AI, and Dun and Bradstreet's 2026 survey found only 6 per cent of businesses say their data is even fully ready to support AI at scale. The technology is not the reason, and the cost base was scoped like a software purchase when it should have been scoped like an operating model change.
For a COO this is the central question, because you own the processes AI is supposed to improve, the people whose work it changes and the operational risk when it goes wrong. If anyone in the business should hold the true total cost of ownership picture it is you, and most of the time nobody does.
Why AI breaks the normal TCO model
When you buy a conventional SaaS product the total cost of ownership is reasonably predictable, covering licence, implementation, some integration, training and support. You can scope it, fix it and hold a vendor to it.
AI does not behave that way, for three reasons.
The cost is consumption-based and it moves. A simple linear workflow might cost a few pence per interaction, and a more complex 2026-era system that calls tools, reasons over several steps and loops until it gets an answer can cost a pound or more per interaction. That is an order of magnitude higher, and as you make the system more capable you make it more expensive to run. The relationship is not linear, and your unit economics shift under you as usage grows and as the system gets smarter.
Most of the work is not the model. The often-quoted figure for data preparation as a share of AI effort has no primary source we could trace, so we will not repeat it. What is measured is starker: Dun and Bradstreet's July 2026 survey of 10,000 businesses found only 6 per cent say their data is fully ready to support AI at scale. The model is the easy bit, and getting your data clean, accessible, correctly permissioned and continuously maintained is where the real spend sits, and it is spend that recurs.
It is never finished. A traditional system is built, deployed and then maintained at low cost. An agentic AI system needs continuous evaluation, prompt tuning, monitoring and adaptation as the underlying models change and your data drifts. Most enterprise deployments carry a real monthly running cost the original business case never mentioned: tokens, vector storage, observability, security and tuning. That makes it an operating function, not a project.
Treat AI like a one-off purchase and you will underestimate the true cost badly, in our founder's experience by 40 to 60 per cent. That is not a rounding error, and it is the difference between a business case that holds and one that quietly comes apart.
The seven cost layers below the waterline
When we build a TCO picture we account for seven layers, and the vendor quote covers, at most, the first.
1. Licence and consumption. This is the visible cost, charged per-seat, per-token or per-call. Model this at realistic production volume, not pilot volume, and stress it for the case where adoption succeeds and usage triples.
2. Data readiness. This covers cleaning, structuring, pipelines, permissioning and ongoing data quality, and it is the single largest line in most AI initiatives and the one most often left out entirely. If your data is not ready, and for most mid-market operations it is not, this dominates everything else.
3. Integration and engineering. This layer covers connecting the system to your existing stack, building the orchestration around the model and handling the cases where it fails gracefully. The model itself is a small fraction of the build cost, and this integration layer is most of the rest.
4. Governance and assurance. This layer runs to model validation, monitoring, audit trails, human-in-the-loop review and the controls a regulator or your own risk function expects. In financial services, insurance and legal this is non-negotiable, and it is a standing cost, not a one-off.
5. People and change. This layer holds training, role redesign, the productivity dip while people learn the new way of working and the management time to lead it. The technology is often the cheapest part of the change, and the human side is where value is won or lost. We have written before about why AI transformation is mostly people.
6. Run and maintenance. This is the recurring monthly cost of keeping it working: tuning, monitoring, security and re-evaluation as models are deprecated and replaced. It belongs in the budget as a permanent operating line.
7. AI debt. This is the structural cost of moving fast without architecture: duplicated tools, ungoverned pilots, models nobody owns and integrations that quietly rot. It does not show up on any invoice, and that is exactly why it is dangerous, as we cover in AI debt: the cost of moving fast without architecture.
A mid-complexity operations agent that a naïve estimate prices at around £150k over three years routinely lands, on our estimates, closer to £350k once these layers are counted, and the difference is the six layers below the licence. The agent is the same, and what changed is how thoroughly it was costed.
What this means for the COO specifically
The CFO will ask what it costs, the CTO whether it works, and the COO sits at the only vantage point where the whole picture is visible. The hidden costs are operational costs, and they land in your function, on your people and in the processes you run.
You see the consumption curve before finance does. Usage-based pricing means the cost follows operational adoption, so if a tool genuinely lands and your teams lean on it, the bill climbs. That is success creating cost, and you will see it in the workflow long before it shows up in a quarterly variance report.
You own the data readiness problem. The effort that goes into data is effort on your operational data, the records, the processes and the exceptions your teams handle every day. No vendor can fix that for you, and it is operational hygiene that determines whether any of the rest works.
You absorb the change cost. The productivity dip, the retraining, the role redesign and the people who need a new path are all operational reality in your function. Scoped properly it is manageable, and ignored it is the reason adoption stalls and the business case never pays back.
You carry the run cost forever. Once it is live, keeping it working is an operational function, and someone owns the tuning, the monitoring and the response when a model is deprecated. If that someone is not named and funded, the system degrades quietly until it fails loudly.
The COOs getting real return from AI in 2026 are not the ones with the cleverest models, and they are the ones who scoped the true cost up front. They sequenced the spend deliberately and refused to let a pilot become a production dependency without an operating model behind it.
How to build a TCO picture that survives contact with reality
You do not need a procurement spreadsheet with forty tabs, and what you do need is disciplined answers to a small number of hard questions before you sign anything.
Model production volume, not pilot volume. Ask what this costs at full adoption and at three times that, and if the economics only work at pilot scale, you have a demo, not a business case.
Cost the data work properly. Before you commit to the model, get a real assessment of whether your data can feed it, and if it cannot, that work is your first and longest cost. It belongs in the business case from day one, and that is exactly what a structured readiness assessment is for.
Name the run owner and fund the run line. Decide now who owns the system in production and what the standing monthly cost is, because an AI system with no named owner and no run budget is AI debt with a launch date.
Count the people cost as a line, not an afterthought. Training, change, role redesign and the productivity dip are real and they are yours, so put a number on them.
Sequence the spend. Start where the cost is low and the value is obvious, build the governance you will need anyway and prove the economics on something small before you commit to something large. Pilots stall between an impressive demo and anything that runs in the business far more often than they reach production, and the reason is usually the same: nobody costed the path from pilot to production.
Counting every layer is the cheaper path
It feels more expensive to count all seven layers, and it is not, because the expensive path is the one where you scope AI like a software licence. You get six months in and discover the real cost as a series of unbudgeted surprises, and by that point you are committed. The pilot is load-bearing and walking away is harder than overspending.
A true total cost of ownership picture, built before you sign, does three things. It stops the initiatives that were never going to pay back before you spend on them, and it properly resources the ones that will. It gives you, the COO, the one thing AI business cases almost never come with: a number you can actually stand behind.
That is the difference between AI as a line of operational cost you control and AI as a liability you discover. The work to tell them apart is not glamorous, and it is just a clear-eyed cost picture and the discipline to act on it.
If you want help building a real total cost of ownership picture for an AI initiative or pressure-testing one you already have, get in touch. Our Breathe assessment is built to surface the costs below the waterline before you commit, and our fractional Chief AI Officer support keeps the run economics in check once you are live.