When I ran an IT asset management business, I often used a simple model to explain licence compliance risk to customers. There were broadly three ways an organisation could try to manage it: commercially, operationally or technically. The commercial route, typically buying an enterprise agreement that licensed everything for everyone, was always the most expensive, particularly when utilisation across many of the products was poor. The operational route, through policies and role-based access, was cheaper, but it needed a comprehensive set of processes and still lacked the granularity that real usage data provides. The technical route, understanding utilisation, reharvesting licences that weren't being used and putting controls in place so people only had what they needed, was consistently the most cost-effective.
I'm now seeing organisations take the same three routes with AI, and I think the same lesson applies. I'd describe them as paying, policing and proving.
My view is that many organisations are doing this in the wrong order: buying AI for everyone, writing contracts and policies to cover the risk, and only then, if at all, looking at how AI is being used. It should be the other way round.
Pay: buying or contracting your way out
The commercial approach to AI tends to take one of two forms.
The first is buying your way out: licensing an enterprise AI tool such as Microsoft 365 Copilot for everyone, on the basis that a sanctioned option will bring use under control. There's some merit in this, as a good approved tool does reduce the pull towards personal accounts. But it is the most expensive way to buy control, and the evidence suggests it delivers less than expected. A large US survey of workers with paid AI subscriptions, by Recon Analytics, found that where they had access to both Copilot and ChatGPT, only 18% made Copilot their main tool and 76% chose ChatGPT. Buying a tool for everyone doesn't stop people using the one they prefer, which is very much the "E5 for everyone" problem in a new form.
I've seen the scale of this first hand. In my earlier IT asset management work, one global company held a very large enterprise agreement covering a whole portfolio of products. Around 75% of those products had adoption and utilisation below 10%. It had paid for the right to use almost everything, and was using very little of it.
The second is contracting your way out: pushing responsibility for AI onto others through legal constructs. Customer terms disclaim responsibility for AI-assisted outputs, employee contracts and acceptable use policies make staff responsible for how they use AI, and supplier contracts ask suppliers to warrant or indemnify their AI use. These clauses have their place. They allocate liability after something has gone wrong, which does nothing to prevent it, and they have real limits. Lawyers point out that nobody can realistically warrant the behaviour of a third-party AI model they don't control, so a warranty is only as good as your ability to verify it. The contracting also works in both directions, with AI vendors frequently seeking broad indemnities from their own customers for how inputs and outputs are used.
Most importantly, accountability doesn't transfer. Data protection regulators are clear on this; the UK's ICO, for example, states that the organisation remains responsible for complying with data protection law, and for demonstrating that compliance, in any AI system processing personal data. It is equally clear that senior management are accountable and can't simply delegate the issue to technical teams. Contracts move the cost of a failure; they don't reduce the chance of one.
Police: governing through policy
The operational approach is to govern AI through policy: acceptable use policies, approved tool lists, guidelines and training. It is considerably cheaper than the commercial route and an essential part of good governance, but on its own it runs into a familiar problem.
Almost every organisation now has an AI policy. Delinea's 2026 Identity Security Report, a survey of more than 2,000 IT and security leaders and a similar number of employees at organisations with 500 or more staff, found that 99.7% have a formal policy governing what data AI tools can access, yet only around half check that access against policy in real time. Three-quarters of the employees surveyed said they had bypassed formal approval at some point, and C-level executives were the most likely to do so. A separate survey for PagerDuty of 1,250 office professionals at large companies found that two-thirds had used AI tools at work that they believed weren't permitted.
The gap is in fit and enforcement. A policy applied uniformly across a business can't reflect how different roles need to use AI, and without usage data there is no way of knowing whether anyone is following it.
Prove: managing AI on actual use
The technical approach manages AI based on what is happening in practice: which tools are installed and used, by which teams and through which accounts, what they connect to and what they cost. Access can be shaped around role and requirement. Unused seats can be reclaimed and reassigned, consumption budgeted and owned, and controls applied where the risk sits instead of uniformly across the business. Tools are already appearing in the Microsoft ecosystem that use usage analytics to find Copilot licences that aren't being used so they can be reassigned, which is licence reharvesting under another name.
It is the most cost-effective of the three approaches, for the same reasons it was in software licensing. You stop paying for what isn't used, you target spend and policy where they add value, and you reduce risk at the point it occurs. It isn't free: it needs the right tooling, trustworthy data and some expertise to interpret it, and controls set too tightly create friction of their own. But the return is considerably higher.
To make that concrete, take an illustrative organisation of 1,000 people and an AI add-on priced, for the sake of argument, at £25 per user per month. Licensing everyone costs £300,000 a year. If usage data shows that around a third of staff need and use it, licensing those roles costs around £100,000. Even after allowing for the cost of discovery, tooling and the time needed to interpret the data, the difference is substantial, and it repeats every year of the agreement. The figures are illustrative rather than a quote, but the shape of the calculation is one I saw repeatedly in software licensing.
This is also where IT asset management has evolved. Its modern counterpart, FinOps, began as cloud cost management and has since extended its framework to cover AI, SaaS and software licensing. The FinOps Foundation's State of FinOps 2026 report found that 98% of FinOps teams now manage AI spend, up from 31% in 2024. What few organisations have done is connect that discipline to how they govern AI.
There is also a reason beyond cost. The people you answer to increasingly want evidence. Customers are adding AI questions to their due diligence, and more of them are asking for proof of AI safeguards in their contracts, and insurers are starting to move in the same direction. In the US, standard policy wordings now allow generative AI exclusions, and some advisers are warning that AI governance which isn't documented will be treated as absent. The cyber insurance market is still working out its position. Earlier this year some insurers proposed caps on certain AI-related losses, while others have since affirmed AI cover and launched AI-specific endorsements. Either way, how AI is governed is becoming part of the underwriting conversation. The direction of travel is clear, and answering those questions will take usage data and evidence of control, which a policy document on its own can't provide.
The case for licensing everyone
The strongest argument for licensing AI for everyone is that it builds an adoption culture targeted licensing doesn't, and that low usage often reflects a lack of training instead of a lack of need. There is truth in that. The answer is to pilot widely, train properly, and then let the evidence decide how far to scale.
The other argument I hear is that enterprise-wide agreements come with a substantial discount, which makes licensing everyone cheaper per user. Having negotiated plenty of these agreements, I'd be cautious. It is one of the oldest approaches in software licensing. The customer accepts the discount on the basis that they will optimise before the next renewal, the discount then disappears at renewal, and the agreement doesn't allow them to reduce their numbers. A discount on licences nobody uses is still waste, and it tends to get more expensive over time.
The approach also needs care in one other respect. Proving how AI is used should mean visibility of tools, accounts and data flows across the estate. It shouldn't extend to surveillance of individual employees. In countries with strong employee privacy protections, and in organisations with works councils, that distinction matters a great deal, and it is something I'll come back to in a separate piece.
What I expect to happen next
Three things seem likely to me over the next couple of years. The first is that insurers will do more than regulators to shape how organisations govern AI, simply because they will price the absence of evidence. The second is that AI licensing will follow the path software licensing did, with compliance reviews built around premium features reaching unlicensed users and consumption nobody owns. The third is that the organisations that bought AI for everyone will spend the next renewal cycle trying to work out who needed it.
There is precedent for the second of these. Organisations holding enterprise agreements for database and middleware platforms often assumed the agreement entitled them to everything. In practice, database options and management packs were licensed separately, and deploying them without the right entitlement produced compliance settlements running into millions. I'd expect premium AI features to follow a similar path.
There is a sharper version of this to come. When AI lets an organisation spend less with an incumbent vendor, that vendor's revenue is at stake, and audits have long been one of the ways vendors protect it. The organisations best placed in that conversation will be the ones that can already show what they use and what they are entitled to.
How the three fit together
You will still need contracts with your AI suppliers and clear policies for staff, so the question is the order in which you approach the three. In software licensing, the organisations that achieved the best outcomes started with trustworthy usage data and let it shape both the commercial agreement and the policies around it. They bought only what they needed and applied policy where it made a difference.
| Approach | How it works | Cost of control | How precisely it matches real use |
|---|---|---|---|
| Pay | Buy AI for everyone, or contract the risk to others | Highest | Lowest |
| Police | Policies, approved tool lists and training, applied evenly | Lower | Partial, and hard to check |
| Prove | Manage access, spend and controls from evidence of actual use | Lowest overall | Highest |
AI is no different. Start by proving what is in use, and let that evidence determine what you pay for and where you need to police.
Three questions for the board
- How much are we spending on AI subscriptions and consumption, and how much of it is actually being used?
- If our insurer, a major customer or a regulator asked how we control AI, what evidence could we show them?
- Are we relying on contracts and policies to manage AI risk, or do we know how AI is really being used?
Sources
- Delinea, 2026 Identity Security Report: The AI Enforcement Gap (September 2026)
- PagerDuty, 2026 Shadow AI Survey, conducted by Wakefield Research
- Recon Analytics, AI Choice 2026: Why Licenses Don't Equal Adoption (US survey of paid AI users, July 2025 to January 2026, published February 2026)
- FinOps Foundation, State of FinOps 2026 (1,192 respondents)
- ICO, accountability and governance implications of AI
- Lexology, pitfalls of AI warranties
- The Insurer, Beazley launches cyber cover for AI shutdowns and regulatory risks (September 2026)
- Insurance Business, ISO's generative AI exclusion is already on thousands of CGL policies