Finance Adopted GenAI Everywhere and Can Observe Almost None of It

A new survey of financial services teams reports a number that is easy to read past and hard to defend. Ninety-four percent have adopted generative AI for observability work. Eighty-nine percent expect to observe their large language models. Six percent have actually implemented LLM observability. Read those three figures in sequence and you are looking at an industry that wired AI into the machinery that watches its systems, then forgot to watch the AI.

The data comes from Elastic’s State of Observability 2026, financial services cut, published in April. It is a vendor-funded survey, so treat the precise percentages as directional rather than census-grade. The shape of the finding holds regardless of decimal places. Adoption ran far ahead of the instrumentation that makes adoption safe, and it ran ahead in the one industry where unobserved automation is not a tooling oversight but a regulatory exposure.

Maturity Went Up. Visibility Did Not.

The same survey reports that 70% of financial services teams now rate their observability practices as mature or expert, up from 45% the prior year. That is a real jump in confidence. It sits oddly next to the 6% who can observe the LLMs they have deployed.

Both things can be true. Teams have genuinely matured at observing infrastructure, services, and traditional applications. The dashboards are better. The alerting is faster. The on-call rotations are calmer. What grew was competence with the previous generation of systems. What stayed flat was visibility into the new layer that generative AI introduced underneath those systems.

This is the trap of measuring maturity against last year’s terrain. A team can be expert at watching the building it already knows and blind to the new floor someone added while it was busy tuning alerts. The confidence is earned. It is also pointed at the wrong layer.

The Audit Question Is Already Live

The reason this matters more in finance than almost anywhere else shows up in two paired numbers. Sixty-one percent of these teams already use observability for real-time compliance and audit. Only 53% rate their tools as acceptable for audit readiness.

Sit with that pairing. A majority have made observability part of how they satisfy regulators in real time. A smaller majority actually trust the tools to hold up when an auditor asks them to prove it. The institutions did not wait for the instrumentation to be ready before they routed compliance through it. They routed compliance through it and are now hoping the instrumentation catches up before the examination does.

When an LLM sits inside a compliance-relevant workflow and nobody can reconstruct what it did, the failure is not a slow dashboard. It is an unanswerable question from a regulator. “Show me the decisions this system made and why” has a clean answer for a traditional rules engine. For an unobserved model, the honest answer is a shrug, and a shrug does not survive an audit.

Regulation Is the Headwind, Not the Excuse

Ninety-five percent of these teams report regulatory hurdles. Sixty-seven percent name GDPR specifically. It would be easy to read that as the reason observability lags: regulation is hard, so instrumentation is slow. The causality runs the other way.

Regulation is precisely why the 6% figure is alarming rather than merely awkward. In a lightly governed domain, deploying models you cannot observe is reckless. In a domain where 95% of teams already feel the weight of regulators and two thirds answer to GDPR, deploying models you cannot observe is a finding waiting to be written up. The regulatory pressure does not excuse the missing visibility. It raises the cost of it.

GDPR in particular makes the knowledge layer load-bearing. If a model processed personal data, you may have to explain what it did with that data, prove you can delete it, and show who had access. None of that is answerable without observability into the model itself. The 67% who name GDPR are naming the exact obligation their 6% implementation rate cannot yet meet.

And the Budget Is Moving the Wrong Way

Here is the part that turns a visibility problem into a governance one. Ninety-nine percent of these teams are actively cutting observability spend. Seventy-one percent report regular cost overruns on the observability they already run.

So the layer with the worst coverage is also the layer being defunded, inside teams that cannot keep their existing observability bills under control. The pressure to cut is real; nobody enjoys an observability invoice that overruns every quarter. But cutting the budget for the floor you have not built yet is how the 6% stays 6%. The instrumentation that would make LLM deployments auditable is competing for shrinking dollars against the instrumentation that is already overrunning.

This is where it stops being an engineering backlog item and becomes a leadership decision. Someone is choosing, implicitly, to keep models in compliance-relevant workflows without the observability to defend them, because the budget conversation is louder than the audit conversation. That choice is reversible. It is also, right now, being made by default rather than on purpose.

Do This Now

If you run AI in a regulated financial workflow, run one exercise this quarter. Pick a single LLM-backed process that touches a compliance or audit obligation. Then try to answer, end to end, the question an examiner would ask: what did this model do, on what inputs, producing what outputs, and can you prove it.

If you can answer cleanly, you are in the 6%, and you should document how you got there because the rest of your industry needs the pattern. If you cannot answer, you have found the exact place where your adoption ran ahead of your visibility. Do not start by buying more tooling. Start by deciding that this one workflow will be observable before it expands, and fund that decision specifically, ahead of the broader budget cut. One auditable workflow you can defend is worth more than ten you adopted and cannot explain.

We have argued before that observability without a governance loop that acts on what it sees is just expensive logging, that governance has to extend beyond engineering into the functions that own the risk, and that vertical industries each inherit a distinct version of the same governance problem. Finance just got its numbers. The numbers say the adoption is done and the instrumentation has barely started.

This analysis synthesizes State of Observability 2026: Financial Services (Elastic, April 2026).

Victorino Group helps regulated teams close the gap between adopting AI and being able to observe it. Let’s talk.