AI and ML Consulting for US Healthcare and Fintech: Compliance First
How regulated US industries deploy AI without failing audit: HIPAA and SOC 2 considerations, model governance, vendor due diligence and what auditors actually ask for.

Most AI projects in US healthcare and financial services do not fail technically. They fail at review, because nobody can produce written evidence of how the model behaves, who reviews its output, and how a change to a prompt is controlled.
The engineering is usually the easy part. This is about the rest.
Scope the compliance workstream first
The single most consequential decision is treating compliance as a workstream with its own weeks in the plan, rather than a checkpoint at the end.
Across our engagements, regulated AI work adds 30-50% to the timeline of an equivalent unregulated feature. That time goes to data handling and vendor agreements, evaluation evidence, human-review workflow, documentation and change control. Those five items are not overhead — they are what makes the thing deployable.
A project plan that shows model development and then "compliance review" as a one-week box at the end has already gone wrong. The reviewer will ask for artefacts that should have been produced continuously, and producing them retrospectively is both slower and less credible.
Data first: what can touch what
In healthcare, any vendor in the path of protected health information needs a business associate agreement. That includes the model provider, the hosting provider, any vector database, any logging or observability service, and any error-tracking tool that might capture a payload. The last three are where the gaps usually are — error tracking that captures request bodies is a well-trodden route for PHI to end up somewhere with no BAA.
Several major model providers offer BAAs on enterprise tiers. Consumer tiers generally do not. Check the specific product and tier rather than the company name.
De-identification reduces scope substantially, and it must be done properly. HIPAA provides two routes: Safe Harbor, removing eighteen specified identifier categories, or Expert Determination. Structured fields are straightforward. Free-text clinical notes are genuinely hard, because identifiers appear inline in prose and removal is imperfect. Treat de-identification as an engineering component with its own test suite and measured recall, not as a regular expression someone wrote once.
In financial services, the analogous questions are what data leaves your environment, whether the model provider retains it, and whether your existing vendor management process covers them. GLBA safeguards obligations flow to service providers, and card data brings PCI scope — the useful decision there is usually to keep card data out of the AI path entirely.
The evaluation harness is your evidence
In an unregulated setting the evaluation harness makes development measurable. In a regulated setting it is the primary artefact you will be asked for.
Build it before the model: a held-out set of real examples with known correct answers, scored automatically, with the methodology written down. Then keep the results. Every version, every date, every change.
When an auditor or an enterprise customer asks how the model performs, the answer needs to be a document with numbers and a described methodology. "It performed well in testing" is not evidence, and in a regulated review it will not be accepted as any.
Include known weaknesses. A written statement of where the model does badly is more credible than a claim of uniform accuracy, and it is also what protects you when the failure mode you documented occurs.
Human in the loop, deliberately placed
For a first regulated feature, the model should propose, rank, extract or summarise — and a qualified person should confirm before anything is acted upon.
This is not timidity. It is the design decision that removes most of the regulatory risk, because the consequential decision remains with an accountable human. It also produces better evaluation data as a by-product, since every review is a labelled example.
Where the human sits matters. Reviewing before action is meaningful. Reviewing a sample afterwards is monitoring, not control, and it should be described as such rather than presented as human oversight.
Fair lending changes the model choice
In fintech, ECOA and Regulation B require adverse action notices stating the specific principal reasons for a denial. A model whose reasoning cannot be articulated in those terms is a compliance problem irrespective of its accuracy.
That pushes credit decisioning toward interpretable approaches — models whose contribution per feature can be stated — or toward using an opaque model only as one input to a decision process that is itself explainable. Disparate impact analysis across protected classes is also expected, and it needs to be run and documented, not asserted.
This is one of the clearest cases where the regulatory constraint should drive the technical choice from the start rather than being discovered at model selection.
Change control: where audits actually find gaps
The most common SOC 2 finding on AI features is change management, and the cause is almost always the same: prompts edited directly in a production console, models switched by changing a configuration value, with no record of who changed what or whether anything was re-evaluated.
Treat prompts and model versions as code. In version control. Changes reviewed. Evaluation run automatically before promotion. Results recorded against the version. A rollback path that has been exercised.
This is not burdensome once set up, and it converts the hardest audit question — how do you control changes to model behaviour — into a matter of showing a pull request history.
Model cards, written once
A written description of what the model does, the data it was evaluated on, how it performs including known weaknesses, what it should not be used for, and who is accountable for it.
Not formally mandated in most US contexts, but auditors, enterprise procurement, internal risk committees and your own future engineers all ask for the same information. Writing it once and keeping it current saves answering the question repeatedly and inconsistently.
What reviewers actually ask for
In our experience, six things:
How the training and evaluation data was obtained, and whether the consent under which it was collected covers this use. Performance on a held-out set, with methodology. What happens when the model is wrong — the specific failure path, not a general assurance. Who reviews outputs, and at what point. How changes to model behaviour are controlled. And vendor agreements for every service the data passes through.
All written. All dated. All produced during the project rather than assembled after it.
Self-hosting: a real option, not a default
Self-hosting an open-weight model removes the vendor from the data path, which simplifies BAAs and residency arguments considerably. It costs you capability relative to frontier commercial models, and it adds operational burden your team has to carry.
We benchmark self-hosted candidates against your evaluation set alongside commercial APIs, so the decision rests on measured quality for your specific task rather than a preference. For narrow, well-defined tasks — classification, extraction, routing — self-hosted models frequently perform adequately and the compliance simplification is worth real money.
For open-ended reasoning or generation, the gap is usually still material. Measure it rather than assuming either way.
What regulated AI work adds to a project
Blurbee delivery data across regulated engagements, 2019-2026. Weeks added beyond the equivalent unregulated build.
Why regulated US teams work with Blurbee
We treat the audit trail as a deliverable, not a log file. Evaluation evidence, model cards, human-review workflow and change control are scoped as their own workstream from week one — because in a regulated setting they are the difference between a feature that launches and a pilot that never leaves review.
- Fixed-scope first projectTwo to four weeks, priced as a number rather than an estimate, so you have evidence before a longer commitment.
- One delivery lead, namedYou always know who to ask. No account manager relaying questions to a team you never meet.
- Code and IP are yoursAssigned in writing, handed over with repositories, infrastructure definitions and documentation whether or not you continue with us.
- Performance budgets in CICore Web Vitals and bundle size gate the build. Slow does not ship.
- Accessibility to WCAG 2.2 AAReviewed with automated and manual passes, not claimed on a slide.
- Written overlap disciplineStanding calls inside your working window, everything else asynchronous with a decision log.