How would you phase out the human as model quality improves?

Wrenfield Virtual Care runs an AI symptom checker that recommends how urgently a patient should be seen: emergency care now, a same-day video visit, or self-care at home. Nnamdi Okafor supervises the triage nursing team and, under today's policy, reviews every single AI recommendation over a headset before it reaches a patient.

How to answer this, stage by stage

Nobody is grading whether you can say "phase it out gradually." They're grading whether you can show the actual numbers behind where the line gets drawn, and why one tier never moves.

Let's learn

Picture four thousand symptom checks a day, and one nurse team reviewing every single one before a patient hears back.

Wrenfield's model reads a patient's described symptoms and recommends emergency care, a same-day visit, or self-care, in seconds. Before the model, a nurse triaged every call by phone, slower, but nobody's recommendation reached a patient unreviewed.

Confidence alone doesn't decide the order. The severe tier sits high on harm no matter how confident the model gets.

Now Nnamdi's team still reviews every recommendation, at every severity level, because the policy never distinguished between a wrong self-care suggestion and a wrong emergency call. Both cost the same nurse-minute today, even though they're nowhere near the same risk.

Here's the turn: the real question was never "is the model good enough overall." It's "which specific tier is good enough, proven over time, to earn less review, and which tier's harm ceiling makes that trade a bad one no matter what the number says."

At its worst, review effort keeps getting spent evenly across all three tiers forever, wasting a team's attention on thousands of low-risk self-care calls while giving no more scrutiny to the handful of severe ones that actually deserve it.

Three different clocks for three different tiers. Only one of them never finishes.

What I would leave alone: nothing about the severe tier's policy changes, ever, on this plan. That's the one place where "leave it alone" means never touching the review requirement at all, not a temporary caution.

The lesson: a single blended accuracy number hides exactly the tier where a miss matters most. Split the number before you trust it.

Now here is the same thing as a story

The short version above is what you'd say pitching this plan to Wrenfield's chief medical officer. Read this one for how the tiered thinking actually got built.

Nnamdi Okafor has supervised triage nurses for seven years, and before the symptom checker, his team fielded every call by phone, sorting genuine emergencies from routine worry with nothing but a script and a lot of practice.

When the phase-out conversation first came up, the instinct in the room was to pick one accuracy threshold, say 97 percent, and reduce review across the board once the model crossed it. Nnamdi pushed back: "a 97 percent average tells me almost nothing about the one call where someone's having a heart attack."

Three branches, three very different endings. The third one never actually ends.

The team split the accuracy number by tier instead, and the picture changed completely. Self-care recommendations were already extremely reliable. Moderate cases needed more evidence. Severe cases, even at a respectable 1.2 percent error rate, meant roughly one missed serious case for every eighty-three reviewed, out of a small but real daily volume.

With the tiered plan running, self-care review drops to a small sample after two straight months under the error floor, moderate cases follow later on a longer proof, and severe cases stay fully reviewed, forever, on principle backed by the sanity check: no labor savings there is worth trading against even one preventable miss.

The old framing asked "is the model good enough yet." The new one asks "good enough for which tier, and never for which one."

I proposed the single blended threshold first because it was the simplest policy to write down. It took Nnamdi's one sentence about a heart attack to see that simple and safe were never the same policy.

BOUND, in one screenNot a lecture on rolling out automation. BOUND is what makes the phase-out plan a real estimate instead of a guess with a confident tone.

Phasing out review is not a one-way door. Any of these three flips a tier straight back to full review.

The recap, one line per letter: break it down is the review-cost and harm-cost equation, own numbers is the measured error rate per tier, use a range is a two-to-four and eight-to-fourteen month window rather than one date, nail the sanity check is comparing savings against the cost of one missed severe case, and direction is the severe tier's own error rate as the number that swings everything.

None of these four assume the model stays good forever. They assume it might slip, and build the catch in from day one.

And if you want to be sure it really works, try it somewhere elseSame five letters, a content moderation queue instead of a telehealth line. A completely different kind of harm, the same tiered logic.

Bracklowe Media runs an AI model that scores posts for policy violations before a human moderator acts. Solenne Kadiri leads the trust and safety team and, under the current policy, has every flagged post reviewed by a person before any action, removal or otherwise, is taken.

Mapped onto BOUND: break it down is moderator-minutes times volume times review percentage, weighed against the harm cost of a wrongly removed post or a missed genuine violation; own numbers means splitting posts into low-harm categories like spam, at 62 percent of volume with a 0.3 percent error rate, and high-harm categories like credible threats of violence, at 2 percent of volume with a 4 percent error rate; use a range means spam review can drop to a 15 percent sample within two months of sustained accuracy, while anything flagged as a credible threat stays at full human review indefinitely, the same logic as the severe telehealth tier, just wearing a different industry's numbers.

Swap "symptom" for "post," and the same tiered arithmetic decides which category earns less review, and which never does.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "split the phase-out by harm tier, and never fully automate the tier where a miss is catastrophic, no matter the accuracy number," and stop.
Cost: there's no budget to build a full tiered rollback system this quarter. Say so honestly, and start with the single highest-leverage move, holding the highest-harm tier at full review while the rest of the system is still being built.
The model gets better, for real: even a genuinely excellent model doesn't change the sanity check on the highest-harm tier, because the argument was never about today's accuracy, it was about what a single miss costs no matter how rare it becomes.

Where people run it wrong.
They set one blended accuracy threshold for the whole system, letting a good overall number hide a dangerous error rate on the one tier that matters most.
They treat a phase-out as a one-way door, with no rollback trigger if a tier's error rate creeps back up after review was already reduced.
They pick a phase-out date off a launch deadline instead of a sustained accuracy window measured over real months.

How to use it live. When someone asks how you'd phase out a human reviewer, ask one question first: what does a single missed case in the worst tier actually cost, and does any plausible labor saving really outweigh it? Build the plan around that answer, not around a single confidence number.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0