Agentic AI deployed into regulated healthcare, led from the operating seat.
I spent thirteen years as the customer: territory business development, national BD leadership and portfolio CMO at behavioral health operators. Since January 2026 I have built and deployed the systems I would have bought, starting with Claude scoring admissions calls for a multi-state treatment operator under HIPAA and a BAA.
The operator, its facilities and its people are not named anywhere on this page. The validation below is correlation on historical calls, not a prospective or causal result.
The deployment
A person seeking treatment rarely decides in one call. They call several times, often over days, and the CRM records each contact as a separate event. The system stitches those call transcripts together with CRM stages and notes into one episode per prospect, then scores the episode on three axes: readiness, viability and need. Leadership sees how the scores move across calls, and which opportunities the funnel was hiding.
The original idea was to follow the whole treatment episode, from first click to discharge. I split it at admission and made pre-admission scoring the MVP, because it was bounded, less ambiguous and more directly actionable.
What I owned. The scoring rubric, the modular agentic workflows, the corpus-synthesis layer and the leadership surface. An engineering partner built the stitching mechanism and handed the project to me.
Where it ran. Nearly 12,000 real calls. The reporting was used by the operator’s corporate leadership team, across all of its residential facilities, for rep-performance and trend review. I paused the deployment in September 2026 for commercial reasons unrelated to the product; the surface for admissions reps themselves was not yet finished.
The moment of truth
I built the scoring to read only the calls, with no admission data, so a score could not simply echo an outcome it had already seen. That left the real question open: do the scores track who enters care? After the fact, I matched 452 scored call sequences to CRM outcomes.
The bands ordered the way admissions went. Across the four priority bands, admission rate rose monotonically with the score. The top two bands admitted at more than twice the rate of the middle band.
The arc across calls carried the most signal. The read of how a prospect’s episode was trending separated likely admits from likely losses by a wide margin, wider than any single-call score.
The test showed where I was wrong. Need scores did not predict admission at all; viability carried most of the signal. I rebalanced the rubric toward viability and readiness.
The limits. A temporary phone-number join matched 47% of sequences, the extreme bands had small samples, and the window was five days of calls. This is correlation on historical records. Absolute admission rates are the operator’s performance data and stay private.
Compliance as architecture
Substance-use treatment records carry 42 CFR Part 2 on top of HIPAA, so most of the convenient tooling people use to ship AI products quickly is off limits. The constraints were designed in, not bolted on.
Every service that touches PHI sits under a BAA. One vendor comparison that could not be covered ran on synthetic calls only.
Redaction before storage. Transcripts are redacted and phone numbers and names hashed before anything is written; storage is encrypted with managed keys.
An allow-list in code decides which model routes may carry PHI. When Claude flagged, mid-build, that a model route might not be covered by our BAA, the work stopped until I confirmed it was.
The build was slower and the architecture heavier for it. That was the price of something that can run inside a treatment operator.
Agents that run my company
In my first year running a boutique growth agency, I was buried in reports, data analysis and initiative tracking. The P&L could not support an analyst, let alone a strategist to turn the reports into tasks. So I built the agents instead. Three run in production today.
Agent runtime. A Claude agent for the agency: email triage, KPI monitoring and daily briefings. About 70 tools in 18 categories, with rate-limit and budget hooks and models tiered by task. It flags trends against per-client thresholds and proposes Bets, grouped tasks that a human approves before anything ships. Agent-run monthly reporting reaches a dozen-plus client organizations, checked against source before delivery.
Decision-queue assistant. Turns email and call transcripts into a one-card decision queue in a macOS app. Untrusted work runs under a separate security profile: tool deny-list, read-only replica, propose-only output.
Vera. A persistent memory graph served to every agent through a custom MCP server on Postgres and pgvector, in daily use. Retrieval is a write: memories that prove useful together strengthen their connection, and unused structure decays.
The systems of record underneath: Salesforce, CallTrackingMetrics, GA4, Google Ads and Search Console, joined in TypeScript, Python and SQL. The attribution layer and the search and AI diagnostic are documented separately, with their validation and limits.
Production is where the design gets corrected.
01Figures that did not match the source. Early agent reports carried numbers the underlying data did not support. I rebuilt the pipeline as narrow, single-purpose steps with two checks: every finding looked up against the data, and a second agent reviewing the synthesis with fresh context. The fabricated figures stopped.
02An assistant that could not ask. When the assistant was unsure, it added another card, so uncertainty became a silent guess. I added question cards that route the uncertainty to me instead. Where a decision carries risk, the agent stops and asks.
03Memory that reinforced itself. In Vera, the more a memory surfaced, the more it surfaced, even after I had said it was not relevant. I changed the rule so only retrievals that proved useful in context strengthen a memory. Mechanical traffic no longer counts.
04A queue of hundreds. The first assistant produced hundreds of open items. A prioritization rubric (who is asking, urgency, relationship importance) cut the queue to dozens. Complexity on the front end to remove decisions on the back end.
How this work was made
About 1,950 commits across 12 private repositories since January 2026, written with Claude Code. AI co-authorship is disclosed in the commit trailers. Every design decision is mine to explain, and I would rather be asked about the decisions than the syntax.
Reusable patterns. Skills, hooks and MCP servers so each workflow is faster to build than the last; a writing skill that keeps client-facing AI text in my voice; and a standing build protocol: spec, independent spec-versus-code check, builder agents, tests re-run by the orchestrator. The main systems each carry a README or design document.
I can walk through the scoring rubric, the validation and any of the three production systems. Start a conversation.