It took twenty-eight days to get a Business Associate Agreement signed for a clinical-documentation AI at a regional health system. Almost all of that time went to arguing over indemnification language. The schedule-critical path ran through a legal document.
Aaron Levie put it sharply in March 2025: software is no longer aiding the worker; software is becoming the worker. A system that does the job has to touch everything the worker touched, so the integration surface is the production system.
That surface breaks into four layers: data, identity, governance, and audit. Each has a team inside the customer that can stop the deployment. The engineer who understands all four gets short meetings. The one who doesn't gets a redesign on day fifty.
Data and identity: start on day one
The most common failure in a forward deployed engagement is the gap between the data the customer described and the data it actually has. The sample is clean. Production isn't. Fields the documentation calls required are empty thirty percent of the time. Dates live as strings in three formats. The customer-identifier field has nine variants because of a 2014 acquisition nobody finished integrating.
So look at production data on day one, not day forty. You'll have the data-access fight either way, and having it early costs you a delayed validation rather than a failed delivery. Expect data plumbing (schema mapping, validation, error handling) to take roughly as much time as prompt engineering. Engineers who underestimate that ratio ship systems that work in the demo and fail in the rollout.
Identity is the layer that has moved fastest. At Ignite 2025, Microsoft made Entra Agent ID generally available: a directory for non-human agent identities, with sponsors, conditional access, and access reviews. Every agent you deploy needs:
- an identity of its own
- a human sponsor inside the customer who is accountable for what it does
- scoped permissions for the data it reads, the systems it writes to, and the tools it calls
- an expiration, renewed on the customer's normal access-review cadence
- an audit trail tying every action back to that identity
In a regulated environment, the security team will ask how the agent's actions trace back to a person who can be paged at 2 a.m. Walk in without an answer and you lose the deployment. Walk in with one and the meeting takes an hour. Identity work has long lead times inside large companies, so pick the scheme in the first week of validation and get it on the request queue.
Governance and audit: design for the explanation
Governance in 2026 usually means three frameworks. HIPAA applies if the system touches health information, which puts the privacy officer in the design conversation and the data inside a BAA. SOC2 applies if the system runs inside the customer's audited perimeter, so its controls for change management, access, and incident response now cover your AI system as well. Then there is the EU AI Act (Regulation 2024/1689, in force since August 1, 2024), which can reach a US vendor serving a US customer if the outputs affect EU residents.
The customer's compliance team remains the compliance authority. The engineer's job is to know enough to ask good questions and to build the system so compliance is easy.
Audit is the layer most engineers ignore until something goes wrong, at which point it is the only thing anyone wants to see. The principles come straight from transactional systems: actions that replay safely, actions scoped to one identity and window, every model call and tool invocation logged, and a clean way to roll back. The working heuristic is simple. Assume every decision the system makes will one day have to be explained in writing to a regulator or a court, and design the trace so the explanation is producible.
Four layers, one engagement
That health system's deployment is the clearest illustration I've seen of the four layers working together. The agent drafted clinical notes from the audio of patient visits, with a promised forty-minute cut in after-hours charting per clinician per day. The team mapped them in the first two weeks.
Data. The audio, the transcript, the draft, and the patient's prior notes were all protected health information, which is why the BAA sat on the critical path. Get it on the queue before kickoff.
Identity. Each clinician got a personal agent identity through Entra Agent ID, scoped to their own patient panel and reviewed on the customer's ninety-day cadence. When a clinician left two months in, offboarding revoked her agent identity within four hours, a window the security team had insisted on.
Governance. The privacy officer joined in week two with three non-negotiables. Audio could not leave the customer's cloud tenant. Notes stayed marked as drafts until a clinician signed. And the system had to reproduce the full chain of inputs and prompts behind any note for at least seven years.
Audit. That last rule collapsed into one invariant. Every draft note in the EHR maps to a single trace ID, which resolves to the audio reference, transcript, prompt, model version, output, clinician edits, and signature event. In week eight, the privacy officer picked a random encounter from a random clinician and asked for the chain. Reconstruction took eleven minutes. Compliance signed off the following week.
The deployment cleared a six-month internal review in eleven weeks. Nobody pushed harder. Every gatekeeper's question had been answered before they asked it.
Three sponsors, and when to push back
Regulated verticals rarely have one sponsor. Shiv Rao of Abridge describes three on the hospital side. The CMIO cares about clinical accuracy and workflow. The CIO owns identity, integration, and security. The CFO cares about revenue cycle, because documentation feeds billing. "Being able to resonate with all three," Rao told Harry Stebbings, "is really how you win the day." Treat the deal as single-sponsor and one of the other two will surprise you, late.
IT and security will push back on your requests, and most of the time they're right. Sometimes they're wrong in predictable ways: a request refused because it doesn't fit the standard shape, or a control built for a different threat model. That is when you go to the sponsor and ask for a one-time exception with an owner, a revert path, and a deadline by which it either becomes standard or closes. The framing makes the request smaller and lets IT say yes without setting a precedent.
Sometimes the pushback improves the design. In one composite case, a global asset manager's research-summarization agent ran on a single shared service account. The CISO killed it in thirty minutes. A shared account couldn't tie an analyst's read of a research note to that analyst, and the rules on material non-public information require exactly that link. The redesign gave each analyst an agent identity scoped like their own login. It cost two weeks and still shipped on schedule, because the CISO sat in the design review in week one instead of the validation review in week eight. Ten months into production, the compliance team was asked to evidence that audit trail for an SEC inquiry. The two weeks paid for themselves.
Rao has a line for all of this: the industry moves at the speed of trust. Scoped identities and producible traces are what that trust looks like when someone finally asks to see it.
Thinking about moving into this work?
I wrote a free guide for people considering forward deployed engineering: what the 4.6x growth figure means, how to tell a real FDE role from a renamed sales job, and a 30-day plan that ends with your first piece of proof.