Skip to content

Goals and Targets for AI Work

Someone has told you that next quarter's objectives must include AI. You did not choose this and cannot opt out of it. The advice comes first on this page, and the evidence behind it follows.

What to write this week

Begin with one question about any target you are handed or asked to draft.

If this number came in exactly at target, what decision would change?

An answer of "nothing" means the target exists to signal, usually to a board or an analyst. An answer of "we would know whom to let go" means it exists to enforce. Only an answer that names a specific resourcing or design decision describes a target doing measurement work. All three read identically on a slide, and only the third gets better when the number is accurate.

Four moves remain available even when the objective is not yours to write.

Pair any usage number with a quality number. This is the practice the most careful measurement programme in the field follows, set out below. Choose the measure that would deteriorate if people used the tool carelessly, and report it beside the adoption figure in the same review.

Put the expected shape in writing before the quarter starts. Where the returns lag, say so in the objective, and state what an early decline would and would not mean. A dip predicted in week one leads to a very different conversation from a dip explained in week twelve.

Decline self-reported time savings as the outcome measure. A government evaluation and a controlled study both found that instrument wanting, for reasons given below.

Ask who will be graded on the number, and whether they were in the room when it was set. That question is not about metrics. It determines whether the number ends up describing the work or defending the person reporting it.

The difference shows up clearly in a concrete case. This illustration is ours, and no published example of a written AI objective surfaced during research for this page.

A weak version names the tool and a usage share:

Objective: become an AI-first engineering organization. Key result: half of all engineers using AI tools weekly.

A stronger version names the outcome and carries its own guardrail:

Objective: shorten the time from customer request to shipped fix. Key result 1: median cycle time falls from nine days to six. Key result 2: change failure rate stays at or below today's baseline. Key result 3: every team can name one task where the tool made things worse.

A login satisfies the first version. The second requires shipping faster without shipping worse, and its third key result rewards exactly the honest reporting that the first punishes.

Objectives and key results are not one thing

The definition is unstable, and that instability explains much of what follows.

A 2024 systematic mapping study surveyed the academic literature on Objectives and Key Results (OKRs) and found 47 studies. Its verdict was that OKR use is "under-documented from a theoretical point of view" (Silva & Santos, 2024). That describes the state of academic grounding, not a measurement of whether the practice works inside companies.

Nine of those studies cite no source at all for what an OKR is. The reviewers write that this "may indicate a lack of rigor and care" (Silva & Santos, 2024).

The studies that do cite a source disagree with each other. The review identified nineteen reference works. Fifteen studies cite a widely read trade book, fifteen cite a practitioner handbook, and one cites Grove's management book (Silva & Santos, 2024). Those three texts describe different objects.

One is a goal-setting discipline. One is an operational recipe with a fixed number of objectives per quarter. One is management by objectives with milestones attached.

One of the reviewed studies lists the consequence among its own reported difficulties, as "Subjectivity in the OKR definition" (Silva & Santos, 2024).

There is a practical edge to this. Measuring the effect of a practice requires agreement about where the practice starts and stops. When someone tells you that OKRs work, a useful reply asks which of these objects they have in mind.

A key result is not a performance indicator

A key performance indicator (KPI) tracks something you already care about, continuously, with no end state. Revenue per customer remains a KPI whether or not anyone sets a goal for it.

A key result is a milestone that paces an objective, and it is built to expire. Grove designed his system, management by objectives (MBO), around two questions (Grove, 1983). Where do I want to go, and how will I pace myself to see whether I am getting there? The key result answers the second question. His book never uses the acronym now attached to it, and calls the practice management by objectives throughout.

Promoting a KPI to a key result converts a measurement into a target, and targets behave differently. Most of this page concerns how.

Grove drew one further line that the modern practice ignores. The system, he wrote, "is not a legal document upon which to base a performance review". It should be "just one input used to determine how well an individual is doing" (Grove, 1983).

He also expected teams to hit key results while missing objectives. His worked example is Columbus, who achieved every milestone and "most certainly did not find a new trade route to China" (Grove, 1983).

For building the underlying metric stack, see the companion guide to standing up AI measurement.

What the evidence supports, and what it does not

A search of the academic literature for studies measuring OKR use against organizational outcomes surfaced none. That describes what such a search found. It falls well short of a finding that the practice fails, and it should not be read as one.

Goal-setting theory is the foundation usually offered, and it is narrower than its reputation.

Its authors attach three conditions to the core finding. A person must be committed to the goal, able to attain it, and free of conflicting goals. Only under those conditions does "a positive, linear relationship between goal difficulty and task performance" hold (Locke & Latham, 2006).

Task complexity weakens the effect without reversing it. The authors report goal-difficulty effect sizes "of .48 for the most complex tasks versus .67 for the least complex tasks" (Locke & Latham, 2002). Smaller, and still positive, which is their framing and worth preserving.

On genuinely novel work the sign can flip. Where a task is complex for the person doing it, "urging them to do their best sometimes leads to better strategies" (Locke & Latham, 2002). The authors name the failure mode "tunnel vision", meaning attention to the goal crowding out the skills needed to reach it (Locke & Latham, 2006). An air traffic control simulation produced a performance goal that "actually interfered with acquiring the knowledge necessary to perform the task" (Locke & Latham, 2002).

Their prescribed fix is a learning goal in place of a performance goal. Critics accept the distinction and doubt anyone can apply it. Managers "may have trouble determining when a task is complex enough" to warrant one over the other. In their phrase, "The goal of setting the right goals is itself a challenging affair" (Ordonez et al., 2009).

Those critics argue more broadly that "the beneficial effects of goal setting have been overstated" (Ordonez et al., 2009).

Both camps overstate what the critics actually recommend.

They stop short of saying abandon goals. Their prescription is that managers treat goal setting as "a prescription-strength medication that requires careful dosing". They add one further note. After working through their checklist, "many managers may conclude that goals are not the best way to motivate their employees" (Ordonez et al., 2009).

What a target does to work you cannot measure well

The evidence in this section comes from Soviet industrial planning, English hospitals and American schools. The mechanism travels to AI work more readily than any particular magnitude does.

Attaching an incentive to one activity does more than raise effort. Incentive pay "serves to direct the allocation of the agents' attention among their various duties" (Holmstrom & Milgrom, 1991).

One consequence of that model matters more than anything else on this page. The case for incentivizing any one activity "decreases with the difficulty of measuring performance in any other activities" (Holmstrom & Milgrom, 1991).

Apply it to an AI usage target.

Usage is the cheapest thing in the workflow to count. Recognizing when the output is wrong costs far more to observe. So does restraint about when to leave the tool alone. So does the verification that makes AI output safe to ship. The model predicts that a usage target draws attention toward the countable thing and away from those complements.

Correlation with value does not make a metric safe to target. Contracts built on such measures "will not in general provide first-best incentives, even when the agent is risk neutral" (Baker, 1992). What matters is the behaviour a target produces once someone sets out to satisfy it.

That behaviour has been documented. A study of targets in English public health care describes gaming as including "hitting the target and missing the point" (Bevan & Hood, 2006).

Two distortions recur under target regimes. Threshold effects act on any fixed target. They create "a perverse incentive for those doing better than the target" to let performance decay toward it (Bevan & Hood, 2006).

Ratchet effects need one more condition, that the target is set from last period's result. The same authors quote a 1958 study of Soviet planning. A wise director there "fulfils the plan 105 per cent, but never 125 per cent" (Nove, 1958, via Bevan & Hood). Overshooting a target buys a harder number next period and nothing else.

Misreporting forms a second channel, distinct from doing the wrong work. Goal setting can motivate people "to misrepresent their performance level" (Ordonez et al., 2009).

At scale that becomes measurable. Serious cheating by teachers or administrators on standardized tests occurred in "a minimum of 4-5 percent of elementary school classrooms annually". Its frequency responded "strongly to relatively minor changes in incentives" (Jacob & Levitt, 2002).

Together these results explain the diagnostic at the top of this page. Once people understand a number as the basis for decisions about them, that number begins reporting on their position alongside the work (Ordonez et al., 2009).

The review clock

AI work has a timing problem, and a quarterly grade handles it badly.

Returns to this class of investment arrive late. United States manufacturing data shows "causal evidence of J-curve-shaped returns, where short-term performance losses precede longer-term gains" (McElheran et al., 2025).

None of this began with AI. Computing's measured productivity contribution ran "up to five times greater over long periods" than over single years (Brynjolfsson & Hitt, 2003). The complementary investments that make technology pay are "often intangible and poorly measured" during the years teams build them (Brynjolfsson et al., 2020).

Research on incentives for exploratory work points the same way. One study compared two funding regimes. The first "tolerates early failure, rewards long-term success", while the second put grantees on "short review cycles, pre-defined deliverables, and renewal policies unforgiving of failure". The tolerant regime "increases overall publication output by 39%", and by 96% for work in the top percentile of citations (Azoulay et al., 2011).

Patience bought that gain at a stated price. Those same scientists produced 35% more papers falling below the citation bar of their least well cited earlier work (Azoulay et al., 2011). More hits arrived together with more duds, and somebody has to absorb the duds.

A controlled experiment found the same direction. Subjects paid for performance "explore less than subjects under the fixed-wage contract". The authors add that "the threat of termination can undermine incentives for innovation" (Ederer & Manso, 2013). The clearest positive prescription names three conditions. They are "commitment to a long-term compensation plan, job security, and timely feedback on performance" (Manso, 2011).

Running both kinds of work under one regime carries its own cost. Combining "hard to measure activities (innovation) with easy to measure activities (routine) is particularly costly" (Holmstrom, 1989).

One caveat belongs in the open. These are life-science grant cycles and a laboratory experiment, not a company deciding whether to continue an AI pilot.

The practical consequence is narrow. A quarterly review suits looking at AI work. It suits concluding that the work has failed much less well, because an early decline is a documented shape for this class of investment.

What the most careful measurement programme refuses to do

DevOps Research and Assessment (DORA) runs a long-running research programme on AI in software development, and it declines to measure adoption as a usage count. Adoption, it states, "is a psychological process involving attitudes, intentions, and actions", which it measures through reliance, trust and reflexive use (DORA, 2025).

Its caution rests on its own data. The 2024 research found "an estimated 1.5% reduction in software delivery throughput and an estimated 7.2% increase in software delivery instability" for every 25% increase in AI adoption (DORA, 2025). Read that as edition-bound, drawn from one survey wave.

The stake runs in both directions here. Google Cloud publishes this research and also sells AI coding tools, so a finding of negative delivery effects cuts against its commercial interest. Its stronger causal language elsewhere still deserves the scepticism owed any vendor-adjacent source.

DORA frames the overall result as amplification. "AI's primary role in software development is that of an amplifier". It magnifies "the strengths of high-performing organizations and the dysfunctions of struggling ones" (DORA, 2025).

Two design lessons follow, and they are the basis for the paired-number advice above. DORA measures delivery as a pair, throughput together with instability (DORA, 2025). Reporting both is what stops a gain in one concealing a loss in the other. It also reports that a clear and communicated organizational stance on AI use "determines AI's impact on organizational performance" (DORA, 2025).

Self-reported time savings fail as an outcome measure, and two results show why. A United Kingdom government department evaluated a 1,000-licence Copilot pilot. It reported that "The evaluation did not find evidence that time savings have led to improved productivity". Even so, "many pilot participants reported noticing time savings in their own roles" (UK Department for Business and Trade, 2025).

Model Evaluation and Threat Research (METR) measured the same gap under controlled conditions. Developers forecast that AI would "reduce completion time by 24%". Measured across 16 developers and 246 tasks, it "increases completion time by 19%" instead (METR, 2025).

Those authors have since qualified the result themselves. In February 2026 they wrote that "we believe it is likely that developers are more sped up from AI tools now". They also called the newer data "an unreliable signal". Participants had declined to take part when "they do not wish to work without AI" (METR, 2026).

The durable finding is not a figure for how much AI helps. People overstate their own time savings, so a target graded on perceived time saved is measuring perception.

What AI targets looked like in practice

These examples carry dates deliberately. Companies revise public mandates quietly, so read each as a statement made at a moment rather than as current policy.

In August 2025, Coinbase's chief executive spoke on a podcast about the company's AI use. He claimed about a third of its code came from AI, "with a goal for 50% by the end of the quarter" (Fortune, 2025). He described mandating that engineers onboard to AI tools within a week. He held a weekend meeting with those who had not, and concluded "And they got fired" (TechCrunch, 2025).

Colleagues had warned him that adoption would take months. Both accounts are press reporting of spoken remarks, and the recording itself was not reachable.

In March 2025, Shopify's chief executive circulated a memo, published the following month. It stated that "Using AI effectively is now a fundamental expectation of everyone at Shopify" (BetaKit, 2025). It also set a hiring test. Teams must show "why they cannot get what they want done using AI" before asking for more headcount (TechCrunch, 2025).

The same memo added "AI usage questions to its performance and peer review questionnaires" (BetaKit, 2025). These accounts quote a memo its author published; we did not retrieve the memo itself.

In November 2025, Meta's head of people told employees in an internal memo that "AI-driven impact" would become a "core expectation" from 2026. The same memo excluded individual usage metrics from the 2025 review cycle (Business Insider, 2025). Meta has not published it.

Survey evidence suggests the pressure falls unevenly. The Society for Human Resource Management (SHRM) surveyed 5,875 United States workers. Entry-level and early-career respondents reported the most demand, with 45% reporting pressure to use AI (SHRM, 2026).

The same survey suggests the effective levers are not the punitive ones. Among workers whose employers used a given tactic, monetary incentives rated most effective at 64%. Only 15% reported any exposure to negative consequences (SHRM, 2026). That body sells membership and training.

One finding sits oddly beside the rest. A survey of almost 6,000 executives across four countries asked about their own AI use. Of those, 28% reported none at all in a working week, and a further 41% reported up to one hour (Yotzov et al., 2026).

Most do use it, but briefly. The pressure to adopt often comes from people whose own use is slight.

One correction is worth carrying, since the sentence appears constantly in these debates. "When a measure becomes a target, it ceases to be a good measure" is not Goodhart's law. The anthropologist Marilyn Strathern wrote it in 1997, about British university examination grades (Strathern, 1997).

Key Takeaways

  • Ask what decision a target would change. An answer of "nothing" means it signals, and "we would know whom to let go" means it enforces. Only an answer naming a resourcing or design decision means it measures. Grove separated the same system from appraisal in 1983 (Grove, 1983).
  • Pair every usage number with a quality number. DORA reports throughput and instability together (DORA, 2025).
  • The definition is unstable, and that is the first finding. Nine of 47 academic studies cite no source at all for the term, and the commonly cited sources describe different practices (Silva & Santos, 2024).
  • The costs of targets are better evidenced than the benefits of this one. No study measuring OKR use against organizational outcomes surfaced in the published record. That describes the literature rather than the practice. The effects of targets on hard-to-measure work appear across economics, public administration and education research.
  • An early decline is an expected shape rather than a verdict. Returns here are J-curved (McElheran et al., 2025), so a quarterly grade suits review far better than cancellation.
  • Never grade AI work on self-reported time saved. Perceived savings ran ahead of measured ones in both studies (UK Department for Business and Trade, 2025; METR, 2026).

Sources

  • Silva & Santos — Surveying the Academic Literature on the Use of OKR, an Update (iSys 17(1)). OKR use is under-documented from a theoretical point of view. View source · verified 2026-09-16 · primary
  • Silva & Santos — Surveying the Academic Literature on the Use of OKR, an Update (iSys 17(1)). nine identified studies do not cite any bibliographic reference for OKR. This may indicate a lack of rigor and care. View source · verified 2026-09-16 · primary
  • Silva & Santos — Surveying the Academic Literature on the Use of OKR, an Update (iSys 17(1)). nineteen sources were identified, six books, six scientific articles, and seven websites. View source · verified 2026-09-16 · primary
  • Silva & Santos — Surveying the Academic Literature on the Use of OKR, an Update (iSys 17(1)). Subjectivity in the OKR definition. View source · verified 2026-09-16 · primary
  • Andrew S. Grove — High Output Management (Random House, 1983), pp. 110-114. A successful MBO system needs only to answer two questions. View source · verified 2026-09-16 · primary
  • Andrew S. Grove — High Output Management (Random House, 1983), pp. 110-114. It is not a legal document upon which to base a performance review, but should be just one input used to determine how well an individual is doing. View source · verified 2026-09-16 · primary
  • Andrew S. Grove — High Output Management (Random House, 1983), pp. 112-113. he most certainly did not find a new trade route to China, and therefore failed to meet his objective. View source · verified 2026-09-16 · primary
  • Locke & Latham — New Directions in Goal-Setting Theory (Current Directions in Psychological Science 15(5)). there is a positive, linear relationship between goal difficulty and task performance. View source · verified 2026-09-16 · primary
  • Locke & Latham — Building a Practically Useful Theory of Goal Setting and Task Motivation (American Psychologist 57(9)). of .48 for the most complex tasks versus .67 for the least complex tasks. View source · verified 2026-09-16 · primary
  • Locke & Latham — Building a Practically Useful Theory of Goal Setting and Task Motivation (American Psychologist 57(9)). urging them to do their best sometimes leads to better strategies. View source · verified 2026-09-16 · primary
  • Locke & Latham — New Directions in Goal-Setting Theory (Current Directions in Psychological Science 15(5)). tunnel vision. View source · verified 2026-09-16 · primary
  • Locke & Latham — Building a Practically Useful Theory of Goal Setting and Task Motivation (American Psychologist 57(9)). having a performance-outcome goal actually interfered with acquiring the knowledge necessary to perform the task. View source · verified 2026-09-16 · primary
  • Ordonez, Schweitzer, Galinsky & Bazerman — Goals Gone Wild (HBS NOM Working Paper 09-083). The goal of setting the right goals is itself a challenging affair. View source · verified 2026-09-16 · primary
  • Ordonez, Schweitzer, Galinsky & Bazerman — Goals Gone Wild (HBS NOM Working Paper 09-083). the beneficial effects of goal setting have been overstated. View source · verified 2026-09-16 · primary
  • Ordonez, Schweitzer, Galinsky & Bazerman — Goals Gone Wild (HBS NOM Working Paper 09-083). a prescription-strength medication that requires careful dosing. View source · verified 2026-09-16 · primary
  • Holmstrom & Milgrom — Multitask Principal-Agent Analyses (Journal of Law, Economics, & Organization 7). it also serves to direct the allocation of the agents' attention among their various duties. View source · verified 2026-09-16 · primary
  • Holmstrom & Milgrom — Multitask Principal-Agent Analyses (Journal of Law, Economics, & Organization 7). the desirability of providing incentives for any one activity decreases with the difficulty of measuring performance in any other activities. View source · verified 2026-09-16 · primary
  • George P. Baker — Incentive Contracts and Performance Measurement (Journal of Political Economy 100(3)). will not in general provide first-best incentives, even when the agent is risk neutral. View source · verified 2026-09-16 · primary
  • Bevan & Hood — What's Measured Is What Matters (Public Administration 84(3)). hitting the target and missing the point. View source · verified 2026-09-16 · primary
  • Bevan & Hood — What's Measured Is What Matters (Public Administration 84(3)). perverse incentive for those doing better than the target to allow their performance to deteriorate to the standard. View source · verified 2026-09-16 · primary
  • Nove (1958), quoted in Bevan & Hood — What's Measured Is What Matters (Public Administration 84(3)). a wise director fulfils the plan 105 per cent, but never 125 per cent. View source · verified 2026-09-16 · ⚠ secondary mirror
  • Ordonez, Schweitzer, Galinsky & Bazerman — Goals Gone Wild (HBS NOM Working Paper 09-083). goal setting can motivate people to misrepresent their performance level. View source · verified 2026-09-16 · primary
  • Jacob & Levitt — Rotten Apples: An Investigation of the Prevalence and Predictors of Teacher Cheating. a minimum of 4-5 percent of elementary school classrooms annually. View source · verified 2026-09-16 · primary
  • McElheran, Yang, Kroff & Brynjolfsson — The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s) (US Census Bureau CES Working Paper 25-27). We find causal evidence of J-curve-shaped returns, where short-term performance losses precede longer-term gains. View source · verified 2026-09-16 · primary
  • Brynjolfsson & Hitt — Computing Productivity: Firm-Level Evidence (Review of Economics and Statistics 85(4)). the productivity and output contributions associated with computerization are up to five times greater over long periods ... accompanied by relatively large and time-consuming investments in complementary inputs. View source · verified 2026-09-16 · primary
  • Brynjolfsson, Rock & Syverson — The Productivity J-Curve: How Intangibles Complement General Purpose Technologies (NBER Working Paper 25148, revised January 2020; published AEJ: Macroeconomics 13(1), 2021), 2020. Our adjusted measure TFP is 11.3% higher than official measures at the end of 2004 ... these complementary investments are often intangible and poorly measured in the national accounts. View source · verified 2026-09-16 · primary
  • Azoulay, Graff Zivin & Manso — Incentives and Creativity: Evidence from the Academic Life Sciences (NBER WP 15466). increases overall publication output by 39%. View source · verified 2026-09-16 · primary
  • Azoulay, Graff Zivin & Manso — Incentives and Creativity: Evidence from the Academic Life Sciences (NBER WP 15466). they publish 35% more articles that fail to clear the. View source · verified 2026-09-16 · primary
  • Ederer & Manso — Is Pay-for-Performance Detrimental to Innovation? (Management Science 59(7)). explore less than subjects under the fixed-wage contract. View source · verified 2026-09-16 · primary
  • Gustavo Manso — Motivating Innovation (Journal of Finance 66). commitment to a long-term compensation plan, job security, and timely feedback on performance. View source · verified 2026-09-16 · primary
  • Bengt Holmstrom — Agency Costs and Innovation (Journal of Economic Behavior and Organization 12). Mixing hard to measure activities (innovation) with easy to measure activities (routine) is particularly cost. View source · verified 2026-09-16 · primary
  • DORA (Google Cloud) — State of AI-assisted Software Development 2025. Adoption is a psychological process involving attitudes, intentions, and actions. View source · verified 2026-09-16 · primary
  • DORA (Google Cloud) — State of AI-assisted Software Development 2025. an estimated 1.5% reduction in software delivery throughput and an estimated 7.2% increase in software delivery instability for every 25% increase in AI adoption. View source · verified 2026-09-16 · primary
  • DORA (Google Cloud) — State of AI-assisted Software Development 2025. AI's primary role in software development is that of an amplifier. View source · verified 2026-09-16 · primary
  • DORA (Google Cloud) — State of AI-assisted Software Development 2025. focus on two key factors: throughput and instability. View source · verified 2026-09-16 · primary
  • DORA (Google Cloud) — State of AI-assisted Software Development 2025. A clear and communicated AI stance determines AI's impact on organizational performance. View source · verified 2026-09-16 · primary
  • UK Department for Business and Trade — The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade (September 2025). The evaluation did not find evidence that time savings have led to improved productivity. View source · verified 2026-09-16 · primary
  • Becker, Rush, Barnes & Rein (METR) — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (arXiv:2507.09089). we find that allowing AI actually increases completion time by 19%. View source · verified 2026-09-16 · primary
  • METR — We are Changing our Developer Productivity Experiment Design (24 February 2026). we believe it is likely that developers are more sped up from AI tools now. View source · verified 2026-09-16 · primary
  • Fortune (Marco Quiroz-Gutierrez) — Coinbase CEO on AI coding assistants (25 August 2025). with a goal for 50% by the end of the quarter. View source · verified 2026-09-16 · ⚠ secondary mirror
  • TechCrunch (Julie Bort) — Coinbase CEO explains why he fired engineers who didn't try AI immediately (22 August 2025). And they got fired. View source · verified 2026-09-16 · ⚠ secondary mirror
  • BetaKit (Alex Riehl) — Shopify CEO Tobi Lutke tells employees to prove AI cannot do the job before asking for resources (7 April 2025). Using AI effectively is now a fundamental expectation of everyone at Shopify ... AI usage questions to its performance and peer review questionnaires. View source · verified 2026-09-16 · ⚠ secondary mirror
  • TechCrunch — Shopify CEO tells teams to consider using AI before growing headcount (7 April 2025). Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI. View source · verified 2026-09-16 · ⚠ secondary mirror
  • Business Insider (Jyoti Mann), via Yahoo Finance — Meta to start grading workers on AI (14 November 2025). core expectation. View source · verified 2026-09-16 · ⚠ secondary mirror
  • SHRM — Navigating AI in the Workplace: 2026. 45% reporting pressure to use AI. View source · verified 2026-09-16 · primary
  • SHRM — Navigating AI in the Workplace: 2026. the most effective methods were offering monetary incentives (64%). View source · verified 2026-09-16 · primary
  • Yotzov, Barrero, Bloom et al. — Firm Data on AI (Federal Reserve Bank of Atlanta Working Paper 2026-3). only 28% of respondents do not use AI at all during the working week. The modal response is up to 1 hour a week (41%). View source · verified 2026-09-16 · primary
  • Marilyn Strathern — 'Improving ratings': audit in the British University system (European Review 5(3)). When a measure becomes a target, it ceases to be a good measure. View source · verified 2026-09-16 · primary