The pilot worked. The demo impressed the steering committee — someone used the word “transformational” in the follow-up email. That was eighteen months ago. Today the old process still runs in parallel with the new one, because turning it off requires a signature, and nobody will put their name on the line that says the model’s output can be trusted without the shadow process behind it.
That signature is the last mile. The industry keeps misdiagnosing it as a change-management problem — workshops and adoption curves — when it is an engineering deliverable with a definable spec.
Change management assumes the missing ingredient is willingness. Walk the stalled deployments and you find something else: the people closest to the workflow are often the most enthusiastic, and the most senior signature is the one that never comes. Enthusiasm is abundant. Indemnity is not.
BCG’s framing of the gap is the most useful one in circulation. Getting a model to roughly 70 percent performance on a task is fast. Getting the workflow to the roughly 95 percent an operation can actually run on is where impact becomes structural — and BCG estimates that 10 to 20 percent of anticipated value erodes before it ever reaches the P&L when no one owns the value logic end to end. Read that estimate carefully: the erosion is not in the model. It happens in the seam between “the model can do the task” and “the organization has stopped doing the task the old way.” McKinsey’s “Rewired” thesis lands on the adjacent point from the consulting side: transformation is domain-by-domain rewiring of how work happens, not tool adoption — and a workflow is only rewired when the old version of it is off.
Why regulated industries move first
Regulated industries feel the seam earliest, which is why — counterintuitively — that is where the serious deployment work is concentrating. Anthropic’s own hiring language says its forward-deployed engineers focus on financial services, healthcare, legal, and government first. Not because those sectors are easy. Because in a regulated operation the last mile is explicit. Nobody gets to hand-wave the signature. The compliance officer’s question — who approved this output, and where is the evidence — is the last mile written down as a job requirement.
Inside a CMS-regulated health plan we operate in, we have watched what actually closes the gap, and it is none of the items on the change-management slide. It is accountability infrastructure: a named human gate on every privileged action, decision traceability a reviewer can walk backwards step by step, evidence retained in a form a compliance officer will sign. Once that infrastructure existed, the parallel process became killable — not because the model got better that month, but because trusting the system stopped requiring anyone to be brave.
Regulation also converts “someday” into budget. Mandates in the CMS-0057-F class arrive with dates attached, and a dated requirement does what no internal AI strategy deck can do: it forces the organization to decide who owns the workflow this fiscal year. Some of the fastest-moving enterprises on AI right now are the ones with a regulator’s clock on the wall.
The waiting argument
The comfortable objection is to wait. Models improve monthly; the 70 will drift toward 95 on its own, and the last mile will close itself. Some of it will. But run the thought experiment honestly: hand that steering committee a model with zero errors. Who turns off the old process? On whose authority? Where does the audit trail live when the examiner asks for it? A perfect model answers none of those questions. Capability moves the floor; it does not produce a signature. The stalled pilots were never really stuck on the missing 25 points.
Enterprises are not failing at AI. The models clear the bar and the pilots prove it — and then the work stalls at the precise point where capability has to become accountability. The last mile is not a gap in the model. It is a gap in who is answerable for the output, and that gap gets built, not waited out.
LockedIn Labs FDE treats that gap as the product surface: the accountability layer — named gates, traceability, retained evidence — ships as a platform, engineered once instead of improvised inside every deployment.
LockedIn Labs FDE is the platform forward-deployed engineers carry into the enterprise — the operating reality held as a governed asset, workflows and agents as reviewable definitions, and every privileged action stopped at a named human.