Sometimes the right AI decision is to ship no agent

The most useful decision on one recent platform build was to leave AI out of the first release. The service reviews diagnostic reports on behalf of hospitals across several healthcare networks, and what shipped contains no agent anywhere in it. It went live in about six weeks and cut audit turnaround roughly in half.

The manual workflow could not keep pace with case volume. Written down plainly, that decomposed into four problems: cases were assigned by hand, quality risk was invisible, reviewer performance was unmeasurable, and taking on another hospital chain meant adding people in proportion to cases.

Read that list again for the word “judgment”. It is not there. Every one of those is coordination, measurement or throughput: the domain of routing rules, a database, role-based views and a dashboard. The reviewers were already expert; nothing needed to reason on their behalf. What they lacked was a system that put the right case in front of the right person and left a defensible record of what happened next.

That distinction is worth making early, because the alternative is expensive in a way that is hard to see. An agent-first release here would have spent its first month on model selection, prompt design and evaluation for a task nobody had yet defined, on top of a case-handling process that still ran on email and spreadsheets. The model would then have been asked to reason about cases whose routing, ownership and status were not recorded anywhere, which is to say, to guess at the very facts the platform was supposed to establish.

1. Case intake. Reports arrive from hospital chains in the formats they already send.
2. Intelligent distribution. Routed and assigned to the right reviewer, by rules the operations lead can read and change.
3. A reviewer portal. On web, iOS and Android, because reviewing happens wherever the reviewer is.
4. An audit trail. A defensible record of every decision, which is the actual product in a regulated review business.

Six hospital chains onboarded and more than 1,500 cases a month now run through it, with audit turnaround roughly 50% faster. None of that came from a model. It came from removing hand-offs and making the queue visible.

It is worth being precise about which constraint that addressed. The one the engagement existed to remove was structural: taking on another hospital chain meant adding people in proportion to cases, because assignment and oversight were manual. That is a coordination ceiling. A model that read reports beautifully would not have moved it, because the limit was never how fast a reviewer reads.

The platform behind those results contains no agent, because the constraint was case routing and visibility rather than reasoning.

An agent layer is now being introduced, and the interesting part is where: sampling around 10% of completed cases, marking likely errors, and setting up blind second review. That is a genuine judgment task: it is graded, statistical and multi-stage, and no rule expresses it. It arrived second because the evidence for it arrived second, and it will be measured against the audit trail the first release produced.

This is the ordering that matters. Ship the system that removes the coordination cost, and it generates exactly the labelled record you need to justify and evaluate the model work later. Start with the agent, and you are reasoning over a process nobody has yet made observable.

1. Is the bottleneck judgment, or coordination? If the work is slow because it waits in queues, moves between people, or is invisible to a manager, a model will make it a faster version of the same mess.
2. Does the correct output vary with the instance, or is it a fixed transform? A fixed transform is code. Code is faster, cheaper, reproducible, and it can be unit-tested.
3. Would a wrong answer be visible? If nobody downstream would notice a subtly wrong output, you do not yet have the review capacity to run a model safely, whatever its accuracy.

Two yes answers and a model is probably the right instrument. Three no answers and you are looking at a software problem that has been described in AI vocabulary, usually because that is the vocabulary the board asked the question in. The honest version of that conversation takes an afternoon and saves a quarter.

None of this is an argument against AI. It is an argument for spending it where the constraint is. The platform above is on the site under featured work precisely because a first release that shipped in weeks, without a model, is a better outcome than a more impressive one that did not ship, and it is the shape most V1 product engagements take once the constraint is named honestly.