The 15-minute referral.

Built by operators, serving major health systems, with a genuine platform behind it — this is not a laggard. But the front door of the whole operation was human: referrals arrive from hospital EHRs, and an operator reads each one, interprets it, verifies it across systems, and re-keys it into the platform. We didn’t take anyone’s word for the pain — we got operator access and walked a live referral ourselves: ~15 minutes. And the time isn’t typing. It’s reading, interpreting, and verifying across four different systems.

Multiply that by every referral, every clinic, every day — while half the industry’s specialty referrals never get scheduled at all — and the front door is the business case.

The obvious pitch was “point AI vision at the documents.” We killed that framing early. The leverage isn’t reading pixels — it’s the mapping layer: one canonical field schema every referral lands in, no matter how it arrives. So intake became a ladder:

Under all three tiers, the same pipeline — exception-based by design:

The structural work was delivered and handed over: the baseline measured, the architecture agreed, a working proof placed in their engineers’ hands, the governance installed, and their own champion teaching the practices onward. The capacity numbers below are the expectations we set with the client — and the measured 15-minute baseline is what they’ll be judged against. Six months on, the client’s own year-end review put the first banked numbers beside them — read those next.

In November the client’s technology team presented its year-end review at the company’s technology all-hands, CEO in the room. None of the numbers below are ours — they are the client’s, from that review. The headline for us wasn’t a metric. It was the slide titled “Our guiding idea”: this engagement’s thesis, AI² — Actual Intelligence × Artificial Intelligence, “it’s times, not plus” — presented as the company’s guiding idea for 2027, three rules intact: help people, don’t replace them; teach before tools; be quick, don’t hurry.

Closer to the front door: the provider-registry lookup we handed over as running code is now a built-in feature of their platform (“fewer tab switches”). Referral intake with notes, attachments and fax automation is partly rolled out, and the team now counts structured data intake as a capability it has proven three times. Phase 1 of the AI education program is complete — employees “moved past basic chat to repeatable AI workflows and agents” — with shared agent practices in place and a second phase planned. A voice-agent proof of concept passed 22 patient-call scenarios on six purpose-built, tenant-scoped tools with 175 passing tests. SOC 2 Type II was issued, a restricted and audited workspace for protected files stands, and platform ownership moved in-house to a newly hired U.S. engineering team.

The 2027 plan reads like the ladder on this page: direct EHR integration first, starting with a major EHR in January; the voice agent to production for one client and one workflow, supervised, with a warm handoff to a person and an off switch; and “agents working with agents” under a registry of every agent, shared memory, and firm rules about which may touch patient data. The expectations above are still expectations. The baseline is still 15 minutes. What changed is that the people judged against it now own the plan.