Mis-hires don't ramp: the year of salary you can't recover

The standard back-of-envelope on a mis-hire (30% of first-year compensation, rule-of-thumb consultant figure since the early 2010s) was always more useful as a soundbite than as an actual estimate. It packed three different costs together: salary spend during the wrong fit, lost output during ramp, and replacement search cost. In a 2019 hiring environment those three sums averaged out to something defensibly close to one-third of base.

In a 2026 hiring environment they do not. The ramp cost has changed shape, the lost-output cost has compounded with AI amplification, and the replacement cost has come down sharply in some segments while the discovery time has stretched out by months. The 30% rule is now a comforting fiction. It absorbs the question instead of answering it.

This article is the cost case. It is the argument we make to engineering directors, finance leads, and chiefs of staff when the question is why does the hiring process need to change again. The argument is not "AI is exciting, your tests are old." The argument is "the unit economics of getting it wrong have moved."

The three costs, separately#

A mis-hire imposes three distinct costs. Treating them as one composite hides which one is dominating in a given role. They have to be looked at separately.

Cost one: salary during occupancy. This is the easy one to compute. It is the fully-loaded compensation (base, benefits, equity, employer costs) multiplied by the number of months between hire and identification-as-mis-hire. For a $180K base in tech, fully loaded at roughly 1.4×, that is about $21K per month occupied.

Cost two: lost output during occupancy. This is the cost of the work that did not get done because the role was nominally filled. In a 2019 model, this cost rose linearly during a 4-to-6 month ramp and then plateaued at the difference between the mis-hire's eventual stable output and the right-hire's eventual stable output. The integrated cost was on the order of 0.6× of cost one.

In a 2026 model, this is the cost that has changed. The mis-hire who over-trusts AI output is not at zero or even negative. They are producing artifacts at high speed. The artifacts are sometimes wrong, sometimes confidently so, and the cost of those artifacts is not the time to produce them but the blast radius of the wrong artifacts that go downstream before the wrongness is caught. We have seen this cost run 1.5–3× of cost one in cases where the wrong artifact went to clients, into financial filings, or into product launches before being caught.

Cost three: replacement. Recruiter fees, interview-panel time, opportunity cost of an unfilled role, onboarding amortization. This was the cost most consultants meant when they cited the 30% figure. It has come down in many segments (direct sourcing, candidate-side intent surfaces, lower agency fees), but the time to discover the mis-hire has gone up. The clock on cost one and cost two does not stop until the discovery happens.

0

Weeks until AI mis-hire is locked in

Across early Acta customer data, AI mis-hires are typically identified between weeks ten and fourteen post-start. Cost one and cost two compound for that entire window.

Why discovery time has stretched#

In a 2019 workflow, a mis-hire was usually identified by week six. The visible signals were straightforward: the person could not write the code, could not produce the spec, missed deadlines. Output volume was the primary signal, and output volume was easy to read.

In a 2026 workflow, the mis-hire's output volume is often comparable to a right-hire's. They produce drafts on schedule. They send Slack messages with attached documents. They demo work in standup. The signal that something is wrong is no longer output volume; it is output quality under inspection, and inspection takes time.

The first wrong artifact may go to an internal stakeholder, who edits without flagging. The second goes to a client, who notices something is off but is too polite to write back. The third goes to a board pack, where the wrong number is caught. And then the question is whether this was a one-off or a pattern. The diligence required to answer that question is two to four weeks. So we end up at week twelve before the conversation happens.

Twelve weeks of salary at the tech-mid-base example is $252K. Twelve weeks of lost output is $150–450K. Twelve weeks of cleanup, depending on what went out the door, is $0 to seven figures.

Why ramp is not the right framing#

We sometimes hear hiring teams describe an AI mis-hire as "they just need more time to ramp." The framing is reassuring and almost always wrong.

The productivity-dispersion data on AI users is clear that the gap between calibrated and uncalibrated AI users is not a ramp gap. Brynjolfsson, Li & Raymond's field study saw the gap stable across the eight-month observation window, the high-AI-literacy worker did not catch up the low-AI-literacy worker, and the low-AI-literacy worker did not catch up to themselves. The two distributions did not converge.

The implication is that the mis-hire's calibration profile in week three is the calibration profile they will still have in month six. The trait is approximately stable on the hiring-relevant timescale. Ramp is a comforting word that suggests the gap will close on its own. The data say it does not.

There is a small literature on AI-literacy training that suggests some lift is possible, particularly on the Use and Apply axis, which is largely procedural. But Detect AI and Evaluate and Create, the two axes most predictive of calibration, are not procedurally trainable in the time frame the hiring decision is being made over. By the time a training intervention could close the gap, the mis-hire's cost one and cost two have already run.

“

The mis-hire's calibration in week three is the calibration they will still have in month six. Ramp is a comforting word the data do not support.

The asymmetry: false positives vs false negatives#

Hiring tests have always traded false positives (hiring someone who fails) against false negatives (passing on someone who would have succeeded). The 30% rule was an attempt to put a number on the false-positive cost so it could be weighed.

In AI-era work, the cost asymmetry has widened, and not symmetrically. The false-positive cost, hiring the over-trusting AI user, has gone up sharply, because the artifacts they produce can go further into the business before being caught. The false-negative cost, passing on a strong AI-literate candidate, has, in some segments, also gone up, because the supply of high-calibration AI users is thinner than the supply of generally-strong candidates was in 2019. So both directions cost more.

The net is that the test has to be more discriminating, not less. A test that produces a clean true-positive on the calibrated candidate and a clean true-negative on the over-trusting one is worth significantly more than a 2019 test was, because the cost surface it is keeping you off of is steeper.

This is the cost argument for an AI-positive structured assessment, not the moral one. The moral argument is that you should test what the work is. The cost argument is that the work has become expensive to get wrong.

Two concrete model lines#

Two illustrative cost lines, drawn from the kind of mid-cap professional services and financial services hires Acta is most often deployed for.

Mid-cap PS associate, $130K base, fully-loaded $182K:

• Cost one (12 weeks salary): $42K
• Cost two (12 weeks of artifacts with one client-facing draft requiring rework): $60–110K
• Cost three (replacement, recruiter, panel time): $35K
• Total mis-hire cost: $137–187K

This is roughly 105–144% of base, versus the 30% rule's $39K. The gap is the artifact cost, which the 2019 rule did not contain.

Mid-cap FS analyst, $115K base, fully-loaded $161K:

• Cost one (12 weeks salary): $37K
• Cost two (12 weeks of compliance-sensitive artifacts, with one SOX-material miss): $80–250K
• Cost three: $35K
• Total mis-hire cost: $152–322K

This is 132–280% of base. The variance is large because the upper bound of cost two depends on whether the SOX miss propagated to a filing.

These are illustrative, not benchmarks. The point is that the 30% rule has stopped being useful as a planning number in the segments Acta is built for.

The case the test has to make#

The cost case is uncomfortable for hiring teams that have stable processes built around the older numbers. We are not asking anyone to overhaul their hiring process on a vibe. We are asking them to consider, when they next have a senior role open whose first six months will include AI-assisted work, whether the test they are about to run can distinguish a calibrated AI user from an uncalibrated one.

If it cannot (if the test is a take-home, a code review, a behavioral panel) then the cost line above is the line the team is exposed to on every offer they extend. The right test does not eliminate the risk. It reduces it by however much the test is predictive of on-the-job calibration. That predictiveness is exactly what the Acta scoring rubric is anchored on and exactly what the ValidationArtifact pipeline keeps honest.

The methodology page walks through the research the Acta score is anchored on, the competency framework behind it, and how the calibrated-trust composite is measured.

References

1. 01Brynjolfsson, E., Li, D., & Raymond, L. Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942, 2025.Read source
2. 02Markus, J., Carolus, A., & Wienrich, C. Objective Measurement of AI Literacy: Development and Validation of the AI Competency Objective Scale (AICOS). arXiv:2503.12921, 2025.Read source
3. 03Bansal, G., Nushi, B., Kamar, E., Lasecki, W. S., Weld, D. S., & Horvitz, E. Beyond Accuracy: The Role of Mental Models in Human-AI Team Performance. Proceedings of the AAAI HCOMP 2019, 2019.Read source