In September 2026, AB Magazine reported a sobering result: across firms, bigger AI budgets did not reliably make operations more efficient, even five years on. That finding, drawn from research by Pantelis Kazakis at the University of Glasgow, puts a name to what many CFOs feel in quarterly reviews — the AI productivity trap.
What AB Magazine found about the AI productivity trap
Kazakis asked a clean question: do companies that invest more in AI become more efficient? According to AB Magazine’s September 2026 write-up, the answer on average was no clear link to efficiency gains, including when the analysis followed results up to five years later. The piece stresses this isn’t an indictment of AI itself; it’s a warning that spend without the right conditions often fails to move core metrics.
The study highlights four conditions where the same capability produces very different outcomes: competitive pressure, managerial ability, stable ownership, and access to long-term finance. AB Magazine argues that firms under sharper competition tended to extract more value, while strong managers, patient capital, and stable governance each raised the odds that AI translated into faster processes, better decisions, and lower costs. The evidence points to context, not tooling, as the hinge.
Zooming out, this echoes decades of work on the so‑called productivity paradox. New general‑purpose technologies can take years to show up in aggregate productivity because organizations must rework processes and intangibles first. Erik Brynjolfsson and coauthors call this the “productivity J‑curve,” where benefits arrive after heavy complementary investment in software, data, and skills; their research at the National Bureau of Economic Research helps explain why fast spend can precede slow payoff. It’s a frame that fits AB Magazine’s findings.
Why outcomes diverge: the four conditions in practice
Competitive pressure changes behavior. When margins are thin and rivals move quickly, firms have fewer places to hide. According to AB Magazine, those conditions correlate with bigger efficiency gains from AI. The likely reason is simple: leaders in tough markets are forced to redesign work, not just add a tool on top of it.
Managerial ability matters because AI is a force multiplier for judgment. High‑quality managers set crisp objectives, pick tractable use cases, and insist on process change. Weak managers spray licenses across teams and hope. The study’s link between leadership quality and outcomes tracks with broader evidence that management practices predict productivity far more than technology spend alone. Readers can find background on this linkage in policy work on AI and productivity from the OECD.
Stable ownership reduces whiplash. Frequent leadership turnover or short‑term incentives can kill multi‑year programs. AI‑enabled reengineering often crosses functions and quarters; you need continuity to change how claims get processed, how a forecast is produced, or how quality checks run on the line.
Long‑term finance buys learning time. Cash that can be committed for several years lets teams build clean data pipelines, retrain staff, and replatform brittle workflows. Without patient funding, pilots stall in year two right as the organization begins to improve. That timing gap explains why many firms report busy AI teams yet flat throughput. As economists have noted since the early debates on the paradox, returns surface when companies line up investment with complementary change — not before. For context on the wider debate over whether AI is visible in productivity data yet, see analysis from Brookings.
From study to scoreboard: KPIs to escape the trap
The AB Magazine study focuses on conditions. Leaders need a scoreboard. Here is how to translate those four conditions into metrics you can track and steer — a practical path out of the AI productivity trap.
- Competitive pressure → outcome velocity: Track time from use‑case idea to first production impact. Set a hard cap (for example, 120 days) and publish it across teams. Falling velocity signals process bloat.
- Managerial ability → rework ratio: For each AI deployment, measure the share of steps removed or automated versus steps added. A ratio under 1.0 means you created work rather than cut it.
- Stable ownership → initiative half‑life: Record how long cross‑functional AI programs run before sponsorship changes. A half‑life under 12 months predicts attrition and weak compounding.
- Long‑term finance → complement share: In annual budgets, tag spend on data quality, workflow redesign, and training. Aim for at least 50% of AI spend on these complements, not just models and licenses.
None of these require a new dashboard. They require naming the operational levers that convert models into results and making them visible at the same level as cost and revenue. As those KPIs move, so will outcomes.
Where the AI efficiency gains tend to appear
The study talks about efficiency, not headlines. Expect early gains in places where judgment is repetitive and data is structured: invoice matching, reconciliation, forecasting, quality checks, and first‑line support. In these zones, small drops in exception rates or handling time stack up fast. Leaders should resist trophy projects that look impressive but dodge the unit economics.
Another place to look is decision latency. Many AI deployments don’t cut headcount; they reduce waiting — for a report, a quote, a price, or a risk score. Measure cycle time before you measure output. If median cycle time barely moves, you haven’t redesigned the workflow yet, and the AI productivity trap is probably still open.
Finally, watch error surfaces. Adding a model can shrink one class of errors while growing another. Track net error cost, not just accuracy. That pushes teams to fix upstream data and downstream handoffs, rather than chasing a single metric that flatters the pilot.
Commit for five years — and make each year count
AB Magazine reports that the study looked for benefits up to five years out. Leaders should plan across that same horizon, but make each year carry real weight. A simple structure works.
- Year 1: Pick three high‑volume processes and redesign them end‑to‑end. Put outcome velocity and rework ratio on the CEO scorecard.
- Year 2: Standardize the operating model for data quality, access controls, and post‑deployment monitoring. Tie bonuses to cycle‑time cuts in two functions.
- Year 3: Expand shared components (prompt libraries, retrieval patterns, review workflows) so teams stop rebuilding the same pieces. Raise the complement share of spend to at least half.
- Years 4–5: Move up the stack to judgment‑heavy work, backed by better data. Retire legacy reports and manual steps as a rule, not a case‑by‑case exception.
The order matters less than the discipline. The AI productivity trap is a governance failure long before it’s a technology problem. With a scoreboard that rewards finished outcomes over busy pilots, the odds improve quickly.
AB Magazine’s takeaway is blunt: the average firm buying AI will not get average gains. Context sets the spread. If leaders wire competitive urgency, capable management, stable sponsorship, and patient finance into their plan — and they measure the right things — the AI productivity trap stops being a cliff and starts looking like a ramp. For more on this, see reuters.com and bloomberg.com and nytimes.com.