The Visibility Dividend

You cannot govern, price, or defend what you cannot see

This is the companion to The Accountability Gap, which argued that AI adoption has outrun AI ownership. This paper takes up the prerequisite. Accountability is impossible without instrumentation: you cannot hold anyone answerable for a system whose behaviour, data flows and cost nobody can observe.

Two 2026 surveys measured visibility from opposite ends. The small-business survey found that a majority cannot establish whether they were breached in the last year. The enterprise survey found that only about a third can see what their AI actually costs to run. Both are the same deficiency: no telemetry on the systems that matter most. And in both datasets, the organisations that closed it outperformed — by multiples.

Two surveys, one missing instrument

The security half

Half of small businesses — 50.5% — reported a confirmed or suspected security incident in the past twelve months. A further 5.6% answered that they did not know. That 5.6% is the more interesting number. A confirmed breach is a bad outcome; an unknown breach status is a structural failure. It means no logging anyone reviews, no alerting anyone trusts, and no way to answer the first question a customer, insurer or regulator will ask.

Every sector above the line handles regulated or commercially sensitive data, which is precisely where an undetected intrusion is most expensive.

Not knowing is not a milder version of being breached. It is being breached with the detection removed.

The economic half

KPMG asked enterprises a structurally identical question about their AI systems: how visible are the operating costs today? Thirty-five percent have full, actively monitored visibility. Forty-two percent describe it as somewhat visible. Thirteen percent see costs only after billing arrives. Eight percent describe them as largely invisible.

Sixty-five percent of organisations are running AI without a real-time view of what it costs. Thirteen percent learn when the invoice arrives — under usage-based pricing, the equivalent of running a data centre with no power meter.

What visibility is worth

KPMG compared organisations with full, actively monitored cost visibility against those without, on the hardest measure in the survey: established ROI.

Visibility does not itself create value — a meter does not reduce a bill. What it does is make three otherwise impossible things routine.

It lets you kill what isn't working

That is not retreat — confidence rose on every measure in the same survey. It is portfolio discipline, and it is available only to organisations that can see which deployments are losing money. The other half are not making better bets; they are making the same bets blind.

It converts "meaningful value" into a number

It turns cheaper models into an actual saving

Access to lower-cost, high-fidelity models was the fastest-rising strategic influence in the survey, from 15% to 22% in three months. But routing a workload to a cheaper model is only rational if you know what the workload costs, how it performs, and which requests genuinely need the frontier model. Without per-use-case telemetry, "switch to a cheaper model" is a guess with a quality risk attached.

Three properties that make AI harder to see

Cost is a function of usage, not provisioning

Conventional software cost is fixed at purchase. AI cost is generated per request and varies by prompt length, retrieved context, model choice and retry behaviour — changeable by a single developer's deployment. One-third of senior leaders cite AI cost and economic literacy as a barrier to deploying agents; 29% report difficulty controlling operating costs as systems scale.

The exposure happens at the prompt

Traditional data-loss prevention watches files, email and endpoints. AI exfiltration happens when an employee pastes a customer record into a chat window — a keystroke inside an approved browser session that most security stacks never see as a data movement. Data security, privacy and risk remained the largest strategic concern for enterprises at 33%, ahead of every economic pressure.

Behaviour drifts without a deployment

The same input can produce different outputs. Providers update models underneath you. Retrieved context changes what the system says without your code changing. Testing at release tells you what the system did once, in a lab. Only continuous observation tells you what it is doing now.

Four layers, in order

Layer 1 is the one most organisations skip. Expect the real inventory to be substantially longer than the procurement record — AI features embedded in software you already bought almost never appear on an approved-tools list, because nobody purchased them separately.

Layer 2 is the highest-leverage thing to build, because one collection point serves security, finance, quality and compliance at once. It has to sit where every request passes through; bolted onto each application individually, coverage becomes optional and therefore incomplete.

Instrument the path, not the applications

Most organisations have poor AI visibility not because they rejected observability, but because they tried to add it application by application — and application-by-application coverage always converges on partial. Every new tool is a new integration. Every team that ships without it creates a gap.

The alternative is to make the instrumented route the route. If every model call passes through a single gateway, then logging, cost attribution, data-classification checks and policy enforcement are properties of the path rather than features each team must remember to adopt. Coverage stops depending on discipline.

Any control that makes the sanctioned path slower than the unsanctioned one will be routed around — and the resulting shadow usage is invisible by definition.

The visibility ladder

Four levels, each with a single test. Answer it from memory, without commissioning an exercise, or you are on the level below.

Most organisations, on this data, are at Level 0 or 1 and believe they are at Level 2. That gap between assumed and actual position is the same gap the confidence paradox describes — 86.3% confident, 56.1% unable to confirm a clean record — and it closes only by running the tests rather than estimating the answers.

The meter pays for itself

Both surveys were designed to measure something else. One was measuring small-business security readiness; the other, enterprise AI value realisation. They converged on the same finding from opposite directions, which is usually a sign the finding is structural rather than sampled.

Spending on AI is not the constraint: planned investment held at a weighted average of US$188m per enterprise, and 79% would sustain it through a recession. The constraint is that most of that spend is being made without a meter attached. The dividend for attaching one shows up in the data as a factor of five.