How a care-coordination startup rolled out AI workflows inside HIPAA constraints
Security had frozen every AI tool request for a year at this Columbus health-tech company, and shadow usage was spreading anyway. Reviewing one platform once, instead of eight tools separately, got the team from freeze to approved rollout in six weeks, with shadow-AI accounts down from eleven to zero.
The situation
The security lead was not being difficult; he was being asked to do the impossible. Eight separate AI tool requests sat in the review queue, each needing its own BAA, its own data-flow review, and its own prompt-injection assessment. So everything stayed frozen, and the freeze created exactly the risk it was meant to prevent: an internal check found eleven personal AI accounts being used for work, some with pasted chart-note fragments.
The VP of Product needed the productivity and the security lead needed the perimeter. Neither was getting either.
The turn
The pitch to security was structural: review one platform once, then let every workflow inherit that review. Second Axis ran its session isolation, encryption, access boundaries, and no-training posture through the full assessment in six weeks, BAA included. Sentinel's screening then became the per-workflow gate: a new workflow now clears in about three days instead of joining an eight-week queue.
Shadow usage ended without a crackdown. Once the sanctioned workspace was faster than a personal account, the eleven shadow accounts went quiet on their own; the follow-up check two months later found zero.
“Security said yes in six weeks because they reviewed one platform once, instead of reviewing every tool my PMs were smuggling in. That is the whole story.”
How Second Axis fits
One security review, inherited by every workflow
Session isolation, encryption, and access boundaries assessed once. Every subsequent workflow inherits the approved posture.
Sentinel screening as the per-workflow gate
Prompt injection and data-access checks run before a workflow reaches the team, turning an eight-week queue into a three-day check.
Guardrails matched to PHI boundaries
Access rules define what agents can touch, so PHI-adjacent workflows run inside limits security set once.
A sanctioned path faster than shadow AI
The approved workspace outperformed personal accounts, so shadow usage ended by preference rather than policy.
Results
| Area | Before | After |
|---|---|---|
| AI tools awaiting security review | 8 in queue | 1 platform, approved |
| Shadow-AI accounts in use | 11 found | 0 at follow-up |
| Time to approve a new workflow | 6-8 weeks | 3 days |
| PM hours per week on manual note synthesis | ~6 | 1.5 |
| PHI incidents attributable to AI tooling | 0 | 0, maintained |
What came next
The review queue is gone, and so is the standoff between product velocity and the security perimeter. The security lead now cites the single-review model as the template for evaluating any new category of tooling.
Company details anonymized at the customer's request.