In a regulated business the binding constraint on AI automation is almost never model capability. It is whether you can reconstruct, eighteen months later, how a specific decision was made — which inputs, which version, which human signed it off. An accurate system you cannot explain is unusable in a lending decision, a claim denial, or a clinical pathway.
That single requirement changes the architecture, not just the paperwork. Here is what it actually changes.
Which rules bind an automated workflow?
More than teams expect, and they stack rather than replace each other. A single automated workflow in a European bank can sit under GDPR, the EU AI Act, FCA conduct rules and the firm's own model risk framework simultaneously — each with a different definition of what "adequate" looks like.
The EU AI Act is the one with teeth and a calendar. Penalties reach €35 million or 7% of global turnover, and it names credit scoring and employment decisions as high-risk by category — meaning the obligations attach whether or not you think of your tool as consequential.
GDPR adds a separate hook that catches more workflows than people assume: a lawful basis is needed for inference as well as training, retention rules reach into embeddings and prompt logs, and Article 22 governs decisions made without meaningful human involvement.
Sector rules then layer on top — HIPAA for protected health information, FCA guidance on consumer outcomes, SOC 2 commitments you have already made to customers. None of these replace the others.
What makes a workflow automatable here?
The same four properties that hold anywhere — high volume, rule-heavy, low-judgement, with a human checkpoint that already exists — plus a fifth that only regulated work imposes: the decision has to be reconstructable.
Reconstructable means that months later you can replay exactly what happened: the inputs, the retrieved evidence, the model and prompt version, the output, and who reviewed it. If any of those is missing, you cannot answer an ombudsman, a regulator, or your own second line.
This is why the durable pattern in regulated work is narrow rather than autonomous. Use models for extraction, retrieval and drafting — tasks where output is checkable against a source — and keep the decision logic in deterministic rules a person can read.
The practical effect is that the model never decides. It assembles the evidence and proposes; the rule decides; the human approves. That architecture survives scrutiny. An end-to-end agent making the call rarely does.
What do regulated buyers demand that others don't?
A specific, fairly consistent set of controls that have nothing to do with model quality. Non-negotiables for tools in regulated environments include SOC 2 compliance, role-based access control, audit logging, and private or self-hosted deployment options.
Add the ones that come up in every enterprise security review: where inference runs, whether the vendor may train on your inputs, how long prompts and traces are retained, and the sub-processor list.
This is why governance work is not a tax bolted on at the end. Roughly 88% of AI agents never reach production, and the attributed causes are infrastructure gaps (41%), governance and security barriers (38%) and ROI measurement failures (33%). Governance is the second-largest reason things die, and it dies in procurement rather than in engineering.
The cheapest time to satisfy those controls is while designing the system. The most expensive is during a security review on a deal you have already forecast — which is the work our AI governance and assurance practice exists to prevent.
Where does this genuinely differ by sector?
The constraint changes; the use case barely does. Every sector below automates document-heavy, rule-bound work. What differs is the rule that decides how it must be built.
| Sector | The rule that binds | What it forces |
|---|---|---|
| Healthcare | PHI must not reach an unbound model | De-identify before inference; BAAs; decision support only |
| Financial services | Every decision reconstructable | Deterministic decision logic; full replay logging |
| Legal | Privilege and information barriers | Access enforced at retrieval, not by prompt |
| Insurance | Adverse decisions must be explainable | Rationale emitted as a first-class artefact |
In healthcare that means anything touching diagnosis or triage stays decision support with a named reviewer and logged overrides. In financial services it means the model extracts and retrieves while the lending rule itself stays readable. In legal and professional services it means retrieval honours matter permissions, because a prompt instruction is not an access control.
Note what does not change: the workflow worth automating is document-heavy and high-volume in all four. Sector expertise earns its money at design time, not at idea time.
What ships first?
Document extraction, almost without exception. It is high volume, schema-defined, checkable field by field against a source page, and it already has a human reviewing the output — which means you can deploy at 90% accuracy and route the remainder rather than needing near-perfection before launch.
The scale is real. JPMorgan Chase is reported to have automated document review to the tune of roughly 360,000 hours a year — one workflow, in one firm.
After extraction, the reliable next steps are complaint and dispute handling, file assembly, and alert triage. All share the same shape: the work is retrieval-heavy, the judgement stays human, and the volume justifies the build.
What consistently fails is starting with the autonomous decision. It is the hardest thing to evidence, the slowest thing to approve, and the one most likely to be blocked after the money is spent — a pattern we traced in which AI automation applications actually reach production.
The short version
Regulation does not stop you automating. It decides the architecture: models for extraction and retrieval, deterministic rules for decisions, humans at the checkpoint that already existed, and logging complete enough to replay any single case on demand.
Design for the replay from day one. Retrofitting an audit trail onto a working system costs several times what building it in does, and it is the step that stalls deals — which is why our AI automation work scopes the evidence layer in the same sprint as the workflow, not after it.