Nassim Dehouche's PRISMA 2020 systematic review, published in Frontiers in Human Dynamics on May 6, 2026, screened 1,847 records and synthesized 94 studies — 42 of them quantitative — into the first peer-reviewed map of observed, not predicted, AI labor-market displacement (Frontiers in Human Dynamics, 2026). The headline result is sharp: entry- and mid-level software-development and content-creation postings in high-income economies fell between −14% and −41% (median −23%) between 2022 and 2024. Most coverage stops there. The number that should change how a 200-FTE operations function builds its Q3 plan is the one nested inside — a +26% expansion in AI-adjacent infrastructure, security, and quality-assurance roles in the same period (Berkes et al., 18-country LinkedIn difference-in-differences), and a 15–22% wage premium for AI-augmented workers in the roles that did not contract.
For a Head of Operations finalizing Q3 backfill in the next three weeks, the operational read on those two numbers is the one most mid-market plans get wrong: the displacement is not uniform, and the planning posture that treats it as uniform — a flat freeze across requisitions, or a flat 10% headcount reduction across teams — is funding the wrong half of the curve. The data does not say slow down. It says redirect.
What Dehouche Actually Measured — and Why "Observed" Beats "Predicted"
The methodology matters here because the AI-displacement literature is mostly model-based. The McKinsey, Goldman, and OpenAI/UPenn studies that defined the public conversation in 2023–2024 estimated exposure — what fraction of tasks a model could plausibly perform — and extrapolated to displacement. Dehouche's review does something different: it pools studies that measure actual job-posting flows, hiring rates, and employment levels against the AI deployment timeline, then applies PRISMA 2020 screening to filter for methodological rigor. The 94 studies that survive the screen are doing observation, not forecasting.
The instrument is unusually well-suited to mid-market planning input. The 1,847-record initial pool reflects the literature's actual density; the 94-study synthesis filters out single-firm anecdotes and unpublished consultancy decks; the 42 quantitative subset gives effect sizes that can be compared across geographies and role types. The methodology is the reason the −23% median is load-bearing. It is the central tendency of measured displacement, not the average of guesses.
The bifurcation finding inside the synthesis is the part that mid-market operations functions need to read against their Q3 plans. The −14% to −41% range across contracting role-types is not a tight band — it is a sectoral signal, with software development and content creation clustered toward the upper end of decline, and adjacent roles inside the same companies expanding. Berkes et al.'s 18-country difference-in-differences on LinkedIn postings adds the comparison: AI-adjacent infrastructure, security, and quality-assurance roles grew +26% in the measurement window, and the workers performing AI-augmented versions of the surviving roles earned 15–22% more than non-augmented peers (World Bank, Jobs and Development; OECD.AI Policy Observatory).
The headline that "AI is displacing jobs" describes one tail of the distribution. The data describes the distribution.
Why Uniform Hiring Postures Miss the 2026 Curve
Mid-market operations functions running 200-FTE rosters typically build Q3 backfill plans from two inputs: attrition forecasts by team, and a top-down headcount envelope set in Q1. When the macro narrative is "AI is displacing jobs," the natural translation is a flat-percentage trim — reduce reqs across the board by 8–12%, prioritize freezes in roles the function reads as AI-exposed, and defer the augmentation question to 2027. The logic feels disciplined and produces the worse outcome.
The mismatch is twofold. First, the contracting and expanding role-types frequently sit inside the same company. A flat 10% headcount cut applied to a SaaS-tooled mid-market firm trims content-marketing and junior-engineering reqs (the contracting tail) at exactly the moment it should be expanding QA-engineering, devops, and security-engineering reqs (the +26% expanding tail). The function reads its own attrition pattern as a market signal and freezes against it, when the market is actually telling it to reshape the requisition mix.
Second, the wage-premium signal is a market validation of where augmentation pays back, and most mid-market plans treat it as a cost rather than a signal. A 15–22% wage premium for AI-augmented workers in surviving roles means external candidates with credible augmentation skills are clearing the market at meaningfully higher prices than non-augmented peers. A function that responds by capping comp bands to "pre-AI" benchmarks is bidding for the workers the data says are not differentiating, and losing the candidates whose presence on the team would shift productivity. The premium is not the problem the function should solve. It is the price tag on the part of the role mix it should be growing.
The function that runs Q3 against the uniform-displacement narrative is not running a planning exercise. It is running a slow-motion rebalancing it neither intends nor controls.
The Redirection Play — What It Actually Looks Like for 200-FTE Ops
The lever is structural and leaner than the published frameworks make it look. Three pieces matter, and they are sequenceable inside the next two to three weeks.
Reclassify the req mix before approving the Q3 envelope
The first piece: split the open and pending requisition list into three buckets — contracting (entry- and mid-level roles the Dehouche review names as the displacement tail), expanding (AI-adjacent infrastructure, security, QA, devops, and data-engineering roles in the +26% tail), and adjacent (roles whose exposure is unclear but which the function is hiring against historical assumptions). The reclassification is one working session of HR + ops leadership; the output is a one-page matrix that replaces the historical req list as the planning input.
The output of this step is rarely the count change the function expects. Most 200-FTE operations functions discover that their open contracting-tail reqs outnumber their expanding-tail reqs by something like 3:1, and that the imbalance is the artifact of last year's plan being copied forward. The function that closes the imbalance — freezes part of the contracting tail and opens net-new reqs in the expanding tail at the same total headcount envelope — has just executed the redirection the data argues for without spending an additional dollar of comp budget.
Fund augmentation training where the wage premium is already priced
The second piece: for the surviving roles in the contracting tail (the 60–80% of the headcount that the data does not predict will disappear), fund AI-augmentation training against the wage-premium signal. The 15–22% premium external candidates are commanding is the market's published price for the skill — internal training that closes that gap inside 90 days is high-ROI by inspection. The cost is bounded: most enterprise augmentation curricula run $800–$1,500 per FTE for the foundational tier, and the breakeven is roughly 4–6 months against the productivity uplift documented in the same literature.
The piece most functions skip is naming the augmented role explicitly. The training works structurally when the post-training role is rewritten — new title, new comp band that captures part of the market premium, new performance criteria that name the augmentation tasks as part of the role rather than as discretionary side-effort. Without the role rewrite, the training is professional development that the function pays for and the market captures. With the rewrite, the function captures the wage premium internally and the team's published comp page tells the candidate pool that augmentation has been priced in (NBER Working Papers).
Open the cross-profile screen on the contracting-tail incumbents
The third piece: before any contracting-tail role is closed via attrition or restructure, run a psychometric and skill-adjacency screen on the incumbents to surface which ones can credibly cross into the expanding-tail roles. The Scovai lens here is the operational one — the candidates the data says could cross are not always the ones the line manager would have nominated, because line-manager nomination is biased toward the legacy role's success criteria, not the destination role's. The screen surfaces the cross-eligible subset against the new criteria, and the function makes the redeployment decision on data rather than narrative.
The cross-profile economics are straightforward. The fully-loaded cost of a senior external hire into an expanding-tail role is typically $25,000–$45,000 in agency fees and ramp friction for a mid-market function; the cost of redeploying an internal incumbent who clears the cross-profile screen and completes a 90-day augmentation track is a fraction of that, and the time-to-productivity is roughly half. The screen pays for itself on the first successful cross. The functions that skip it default to external hiring against a candidate pool that — by the wage-premium evidence — is the most expensive cohort in the labor market right now.
The Counter-Argument and Why the Data Closes It
The natural counter from a budget-pressed mid-market COO: the Dehouche review is one paper, the bifurcation may not generalize to our sector, and the disciplined move is to wait for two more quarters of evidence before reshaping reqs. The logic sounds patient and produces the wrong outcome.
The Dehouche data is direct on the timing. The −14% to −41% contraction was measured across 2022–2024 — the bifurcation is not a forward forecast the function might be early to. It is a backward measurement the function is currently late to. A 2026 Q3 plan that backfills against 2022 role-mix assumptions is staffing for a labor market that already finished shifting. The +26% expansion in adjacent roles and the 15–22% augmentation premium have already been clearing the market at scale for 18–24 months. The function that waits two more quarters before reshaping is not being conservative. It is being two years behind.
A sharper version of the counter: even if the bifurcation is real at the macro level, our specific sector may not show it, and the cost of a mistimed reshape is high. The Dehouche synthesis closes this too. The 94-study pool spans software, professional services, financial services, healthcare administration, and customer operations — the methodological reason for the systematic review is precisely to test for sectoral variation, and the bifurcation pattern survives the screening across all five clusters. Mid-market sectors that think they are insulated are the ones the review names as already inside the −23% median, not outside it.
The Q3 Decision Compressed to One Action
For a Head of Operations finalizing 2026 mid-market workforce plans in the next two to three weeks, the implication compresses to one rule:
Before the Q3 hiring envelope is approved, reclassify every open and pending requisition into contracting, expanding, or adjacent — and reshape the mix at the same total headcount, fund augmentation training against the published wage premium, and run the cross-profile screen on contracting-tail incumbents before closing any role to attrition.
The triage cost is one cross-functional working session, one curriculum decision, and one psychometric pass on the incumbent population. The downside cost of not triaging — at the −23% median, the +26% expansion, and the 15–22% wage-premium signal Dehouche and Berkes have now placed in the peer-reviewed record — is a Q3 plan that hires the contracting tail at last year's rate, freezes the expanding tail by oversight, and pays the cross-profile cost twice in 2027 when the function rebuilds the augmented roles it should have opened in May.
The −23% number is not the displacement story. The +26% nested inside it is. The Q3 question is whether this cycle's requisitions go out against the role mix the headline narrative describes, or against the role mix the peer-reviewed evidence now names.
