The insurability gap: why UK boards will cap AI agents long before the models do
TL;DR: NVIDIA has signed memorandums of understanding with six of the largest names in institutional capital to build financing platforms aimed at mobilising over $500 billion of third-party capital for AI infrastructure, assessed on the same variables any infrastructure lender uses: the customer, demand, utilisation, cash flow and residual value. Writing in Insurance Journal, James Felton Keith argues that no comparable measurement exists for the layer above the hardware, the layer where an autonomous system is granted authority to act. His proposed unit, Agent Insurable Value, is his attempt at giving underwriters something they can actually price. The British consequence is sharper than the American one. UK financial firms already run AI at scale, the London market has just been warned it may be covering AI risk it never priced, and boards facing that combination will do the rational thing: keep delegating advice rather than authority. The binding constraint on UK enterprise AI is not model capability. It is that nobody can yet write a policy against what an agent is allowed to do.
Two markets, one chain, and only half of it has a price
The financing announcement is the more interesting half of the story precisely because it is unremarkable. NVIDIA’s own framing is that “In AI, compute is revenue”, and that “An AI factory turns energy and data into valuable intelligence”. Strip the language back and what six of the largest infrastructure investors in the world have agreed to underwrite is a power station with better margins. Utilisation, contracted demand, cash flow, residual value: these are the questions asked of a toll road or a fibre network. Compute became financeable the moment it became measurable in those terms.
Keith’s point is that the chain does not stop at the AI factory. Capital funds compute, compute trains models, models are wrapped into agents, agents are inserted into business processes, and business processes produce economic output. The financiers are pricing the top of that chain with real precision. The bottom of it, where an agent is authorised to approve an invoice or reallocate cloud spend, has no comparable unit of account at all.
Strategic Reality: An organisation can now raise institutional capital against the machine that produces intelligence more easily than it can buy cover against the decisions that intelligence is permitted to make. That asymmetry is not a technology problem. It is a measurement problem, and measurement problems get solved on the timescale of market convention, not the timescale of model releases.
Keith’s proposal, Agent Insurable Value, tries to close it. The idea is deliberately not about the software. A cheap agent with a corporate card and a procurement mandate carries far more exposure than an expensive model that only drafts copy, and the framework is built to separate the two. What it asks about instead is how much of the business leans on the system, what it has been authorised to decide and to transact, how critical the process is if it stops, which resources and data it can reach, whose intellectual property it touches, how concentrated the underlying dependency is, and how much governance and human supervision actually sits around it. The insight underneath is one line long. Value the authority, not the asset.
The real story is the correlation term
British firms are already further into this than the debate assumes. The Bank of England and FCA joint survey of artificial intelligence in UK financial services found 75% of responding firms already using AI with a further 10% planning to, against 58% and 14% two years earlier. Insurance was the heaviest user of any sector surveyed at 95%. The numbers that matter for underwriting, though, are the uncomfortable ones further down the report.
| Critical numbers | Value |
|---|---|
| Third-party capital NVIDIA’s financing platforms aim to mobilise | Over $500bn |
| Financial institutions signed to those platforms | 6 |
| UK financial services firms already using AI (BoE/FCA survey, 2024) | 75% |
| UK insurance sector firms already using AI | 95% |
| Firms reporting only “partial understanding” of the AI they run | 46% |
| Firms reporting “complete understanding” | 34% |
| AI use cases rated high materiality | 16% |
| Share of all named third-party model providers held by the top three | 44% |
| Share of AI use cases that are third-party implementations | One third |
Read those together and the underwriting problem writes itself. Insurers price frequency, severity and correlation. Frequency and severity for an agentic failure are hard but tractable, and are broadly what Keith’s framework attempts to structure. Correlation is the one that should worry the London market, because 44% of named model providers sitting with three firms means a single model defect does not produce a claim. It produces a book-wide event, arriving on the same Tuesday, across insureds who have no commercial relationship with each other and no idea they share a dependency.
Critical Context: The 46% of firms with only partial understanding of their own AI are not badly run. The survey attributes the gap largely to third-party models, which is to say the understanding is missing because the system was bought rather than built. Underwriters asking whether a client uses AI on a proposal form are asking a question a large minority of respondents genuinely cannot answer accurately.
This is the structural cause behind a symptom the London market has already flagged. The International Underwriting Association has warned that standard computer-system definitions in existing cyber wordings will very likely capture generative AI as businesses currently use it, meaning insurers may already be covering AI risk they never priced. Silent cover is what happens when a market has exposure but no measurement object. It is not a drafting failure. It is the vacuum that forms where an underwriting language should be.
What happens in the boardroom when risk cannot be transferred
Here is the part that connects a Lime Street underwriting problem to a British company’s AI roadmap.
Enterprises do not deploy capability. They deploy capability minus whatever residual risk they cannot transfer, insure, or contractually push onto someone else. That is the whole logic of how large organisations adopted cloud, outsourcing and offshore manufacturing. Each of those became boardroom-acceptable at the point a professional indemnity, cyber or contingent business interruption product made the tail bearable.
AI agents have no such product at scale. A handful of affirmative offerings exist, and the IUA’s underwriting director counts three markets preparing or already writing AI-specific cover, but a board approving a system that can move money is currently approving an uninsured tail. Faced with that, directors do not refuse AI. They do something more damaging to the business case: they approve it with the authority stripped out.
Hidden Cost: The rational response to an uninsurable tail is to cap delegated authority, and capped authority is precisely what destroys the return. An agent that must route every action through a human approver reproduces the cost structure it was bought to remove. The programme does not fail. It underdelivers, gets labelled a disappointing pilot, and the conclusion drawn is that the models were not ready. The models were fine. The risk transfer was missing.
That mechanism explains an otherwise puzzling British pattern. UK insurers have embedded AI in core functions whilst the gap between ambition and operational scale widens, and the wider market has committed serious capital without publishing evidence that it works. Both look like execution failures. Read through the insurability lens, a good share of it is authority rationing.
Why a governance certificate does not answer the underwriting question
The obvious objection is that this is what governance frameworks are for, and UK boards have spent two years building exactly that apparatus. ISO/IEC 42001 certification, NIST AI Risk Management Framework alignment, named accountable owners, model registries.
Keith’s argument against sufficiency is the strongest passage in his piece, and it survives translation to the UK context intact. Two companies can hold identical governance credentials whilst running agents with wildly different economic authority. One agent recommends a purchase order. The other executes it. Both tick the same boxes. Their severity distributions are not remotely alike. Governance evidence tells an underwriter that a process exists. It does not tell them what is at stake when the process fails.
This is the same fault line we identified in the AI accountability gap facing UK procurers, approached from the opposite side of the contract. There, the failure was buyers accepting opacity from vendors. Here, it is buyers being unable to describe their own exposure to a market that needs to price it. Both resolve the same way: by making the specific, contractual, quantified answer a requirement rather than an aspiration.
| Stakeholder | What the insurability gap changes | The question they now have to answer |
|---|---|---|
| Board and audit committee | Approving agentic authority is approving retained risk, not transferred risk | What is the maximum economic value any single agent can commit without a human? |
| CFO | Business cases assume a benefit that only lands at full autonomy | What does the return look like at capped authority, honestly modelled? |
| Chief risk officer | Existing cyber cover may respond, or may not, with nobody sure which | Which wordings currently touch our AI, and has anyone tested that in writing? |
| Broker | Proposal-form questions do not capture agentic exposure | Are we asking clients what their AI can do, or only whether they use it? |
| CTO and platform owners | Spend controls become a risk control, not a finance control | What are the hard ceilings on spend, transactions and system changes per agent? |
Five things to do before the renewal conversation finds you
None of this requires waiting for a market to form. The organisations that will be insurable first are the ones that can already describe themselves in underwriting terms.
1. Build an agent register with authority attached. Not a model inventory. A list of every autonomous or semi-autonomous system, and against each one the money it can commit, the systems it can alter, the data it consumes and the human checkpoint it can bypass. Most firms have the first column. Almost none have the rest.
2. Put a number on the exposure inside each agent’s remit. The annual value of the process it touches, and the largest single action it can take unsupervised. This is the core of what Agent Insurable Value is reaching for, and it is calculable today with a spreadsheet and an honest afternoon.
3. Impose hard economic ceilings in the platform, not the policy document. Spend caps, transaction limits, rate limits, blast-radius constraints enforced at the API and account level. An agent that provisions runaway infrastructure or loops expensively is not being attacked, it is doing exactly what it was permitted to do, badly. Control that fails closed is the difference between an incident and a loss.
4. Map your concentration honestly. If your agents, your suppliers’ agents and your outsourcers’ agents all sit on the same foundation model, you hold correlated risk regardless of how well governed each deployment is. The BoE and FCA data on provider concentration is the market-level version of the same picture.
5. Ask your broker the specific question, in writing. Does any policy we currently hold respond to loss caused by an AI system acting within its granted authority? Record the answer. Given the IUA’s warning about silent cover, that answer is genuinely uncertain today, and uncertainty is worth documenting before a claim rather than during one.
Take Action: Sequence these by maturity. Firms still running advisory-only AI need steps 1 and 5. Firms with agents touching production systems need 3 immediately. Firms with agents holding transaction authority need all five, and need step 2 in a form a third party could audit.
Four problems that will not appear in the first board paper
The measurement itself creates disclosure risk. Producing a credible agent register means producing a document that quantifies your exposure. That document is discoverable, and it will be read in hindsight against whatever went wrong. The answer is not to avoid writing it. It is to write it with the same care as a risk register, and to act on what it says.
Priced cover may arrive at a price nobody wants. The comfortable assumption is that an insurable market makes agentic AI cheaper to run. The first honest pricing of correlated model risk could easily be expensive enough to change the build-versus-buy calculation, or to make certain classes of delegated authority commercially unattractive on their own terms.
The capital and risk markets are drifting apart. Infrastructure financing is racing ahead while risk transfer stands still, and the strain is already visible in credit markets, where banks have been distributing AI data-centre debt as single-borrower exposure approached internal limits and the ECB has warned about the fragility of AI-driven valuations. A correction in compute financing would not fix the agency measurement problem. It would arrive on top of it.
Attackers are already operating at the layer insurers cannot price. Cyber underwriters have warned that autonomous attack chains bring smaller British firms into range because the cost of running an attack collapses. The exposure Keith describes is the benign twin of that: same autonomy, same speed, no attacker required. A market that cannot yet price the hostile version has no chance of pricing the self-inflicted one.
Authority is the thing that needs a number
The strategic reading of NVIDIA’s financing announcement is not that AI infrastructure is now investable. It is that a measurement language turned an expense into an asset class in a matter of quarters, and that no such language exists one layer up.
Three things determine whether a British organisation gets past advisory-grade AI in the next eighteen months. First, whether it can state, in pounds, how much economic activity sits inside each agent’s remit. Second, whether its controls constrain economic authority rather than just documenting intent. Third, whether it can hold a coherent conversation with an underwriter who is about to start asking much harder questions than whether the firm uses AI.
Strategic Insight: The firms that treat agentic authority as a quantified, capped, auditable thing are not just better governed. They are the ones that will still be able to buy cover when the market prices this, and the ones that can safely delegate more in the meantime. Measurement is not the compliance cost of autonomy. It is the permission to have any.
Capital markets have learnt to price the machine. Until somebody can price the mandate, British boards will keep buying capability and switching most of it off, and will keep blaming the technology for a decision they made in the risk committee.
Source and attribution
This analysis responds to “Viewpoint: If AI Compute Is an Investable Asset, AI Agency Must Become Insurable” by James Felton Keith, published by Insurance Journal on 21 August 2026. The Agent Insurable Value framework and its associated risk categories are Keith’s, presented as part of the InclusionScore framework; the UK application, the argument about authority rationing, and the recommendations are ours.
Supporting figures come from primary sources: NVIDIA’s announcement of AI compute infrastructure financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, and the Bank of England and FCA joint report on artificial intelligence in UK financial services.
Resultsense provides independent analysis of AI developments for UK businesses. Read more of our strategic insights, follow the stories behind this analysis in our AI news coverage, or build your team’s vocabulary with the AI glossary.