When does an AI agent actually learn from experience? →
Current answer · Explainer + follow-up Brief
Agent experience is one mechanism. Organizational learning also requires coordination, authority and evidence across people and decisions.
AI can help a company learn when experience changes later decisions and checks show the changes help. Shared memory can carry lessons across teams; it can also spread a mistake. The crucial capability is revising what the organization treats as true.
Current answer · Explainer
Following covers this question’s published contributions. It does not follow neighbouring questions or subscribe you to email.
AI, organizational learning and the difference between shared information and better collective decisions.
Research shows improvements in specific tasks. Reliable learning across teams and changing conditions remains a stronger, unproven claim.
Article version history → A company that learnsThese narrower questions are not promises of a future article.
Prospective comparisons on new work, changed conditions and handovers, with outcomes, maintenance cost and decision rights kept visible.
See the limits of this answer →Question versions describe its framing, published contributions and place in the map. Article versions keep their own evidence and history.
A new organizational perspective connects experience, decision rights and evidence from later work.
See this change across Atlas →Original article history and evidence
This history comes from saved content versions. A new revision can be an editorial change; it does not by itself mean that the underlying facts changed.
Explains how AI might help organizations learn across teams, and why shared memory and faster execution are insufficient evidence.
No earlier reviewed assessment is available to compare. No change in the underlying facts is being claimed.
Explains how AI might help organizations learn across teams, and why shared memory and faster execution are insufficient evidence.
Compare later outcomes under changed conditions with a well-managed baseline using the same tools and current data.
Prospective comparisons on new work, changed conditions and handovers, with outcomes, maintenance cost and decision rights kept visible.
AI can help a company learn when experience changes later decisions and checks show the changes help. Shared memory can carry lessons across teams; it can also spread a mistake. The crucial capability is revising what the organization treats as true.
Research shows improvements in specific tasks. Reliable learning across teams and changing conditions remains a stronger, unproven claim.
AI assistance increased resolved issues per hour by 15% on average in the support-worker study.
5,172 workers; one company and subcontractors; staggered rollout, heterogeneous effects. This is not a shared-memory experiment.
AI-assisted individuals produced higher-rated proposals than unaided individuals in the P&G experiment.
791 randomized professionals; one-day workshops; no demonstrated long-term organizational learning or replacement of durable teams.
Nubank reports improved card-delivery outcomes after evaluation-driven changes to agent variants.
Company-authored; multiple changes bundled; heterogeneous use cases; no isolated memory effect.
Booking.com documented a repository of hypotheses, experiment iterations and decisions, including failures.
Historical company account; the archive’s causal benefit was not isolated.
Performative prediction formalizes settings in which deploying a prediction changes its target distribution.
Formal framework. Atlas’s inventory example is invented and illustrates a possible feedback, not measured behavior.
Atlas infers that a common AI explanation could reduce the variation needed to challenge a mistaken rule.
March’s model is theory and simulation, not a field estimate of AI groupthink.
Atlas applies the formal-versus-real authority distinction to AI systems that frame options and evidence.
The 1997 paper models authority and information; it does not measure contemporary AI deployment.
The Indian management follow-up found that about half the adopted practices had been dropped, with a substantial treatment–control gap still remaining.
Nine-year follow-up in weaving firms; reported explanations for decay are not randomized mechanisms; not an AI study.
Atlas derives contrasting possibilities from which internal and external coordination costs AI reduces.
Conceptual scenarios based on transaction-cost theory, not a prediction or measured AI effect.
A team concludes that falling sales mean it should hold less stock. Purchasing cuts its order. Sales fall again. An AI assistant retrieves the earlier decision and recommends another cut. The company has shared information and a repeatable process. It may also have created the evidence that seems to justify its decision.
In this fictional bag company, availability limits recorded sales. A customer who cannot buy does not appear as a completed order. The decision changes the data that the next decision will use. Researchers call the broader phenomenon performative prediction: acting on a forecast can change what it predicts. A loop is therefore no guarantee of learning.
A fictional bag company sees fewer orders. Its AI has the sales history. What happens depends on what the organization treats as evidence.
Recorded orders fall.
The team accepts “demand is weakening”.
Purchasing reduces replenishment.
More stockouts mean fewer recorded orders.
The next forecast sees the lower sales
“Order less when sales fall” becomes a shared rule.
The loop cannot distinguish lost demand from customers who could not buy. Agreement across departments can spread the same mistake.
Illustrative mechanism, not observed company data or an Atlas experiment. Selecting a view reveals an authored scenario; it does not run an AI model.
Back to the next observation: The next forecast sees the lower sales.
“Order less when sales fall” becomes a shared rule.
The loop cannot distinguish lost demand from customers who could not buy. Agreement across departments can spread the same mistake.
Back to the next observation: New evidence can change the explanation.
Keep the explanation provisional until the comparison is informative.
A rise after restocking alone is not a causal test. Specify the comparison, expected result and cost limit before acting.
Back to the next observation: Changed conditions reopen the question.
The earlier finding stays in the record; its authority over this decision does not.
Retain what was established under the old conditions. A changed situation calls for a new check, not a rewritten history.
Organizational learning means more than teaching a model or saving a conversation. Here, it means that experience changes how people coordinate and decide, helps beyond the original person or task, and remains open to revision when circumstances change. A company’s practice can change without retraining the model.
Generative AI offers a plausible bridge: retrieve an old investigation, explain it to another function and propose how it applies to a new case. Research already shows narrower benefits. In a study of 5,172 customer-support workers, AI assistance raised issues resolved per hour by 15% on average. Gains were concentrated among less experienced and lower-skilled workers. That is evidence of expertise becoming more accessible within a specific workflow.
A preregistered P&G experiment assigned 791 professionals to innovation work alone or in pairs, with or without AI. AI-assisted individuals produced higher-rated proposals than unaided individuals; AI also helped bridge commercial and technical perspectives. These were one-day workshops, not a test of whether teams retain better judgement over years.
The strategic possibility is that a lesson discovered in customer service could change product design before the product team repeats the same mistake. The missing step is showing that the transfer improves a later decision, rather than merely making the earlier explanation easier to repeat.
Return to the bag company. “Low sales mean low demand” is a conclusion stripped of its conditions. A more useful record separates the observation, rival explanations, the decision, what was expected and what the check actually established. If stockouts explained the earlier result, that finding should travel with its scope and exceptions.
An AI with access to shared records can help find the relevant previous case, surface a conflicting one, compare today’s conditions and point to the evidence. But a persuasive lesson still needs a reason to be reused. Someone must decide when a hypothesis becomes an accepted practice, which decisions it may influence and what would suspend it.
Booking.com described an important precedent in 2017: an experiment repository preserving hypotheses, iterations, results and final decisions, including failures. Teams could consult earlier work rather than inherit only the winning story. The authors also acknowledged difficulty extracting broader lessons across experiments. This is a company’s account of its own system, not an isolated causal estimate of the repository’s value.
AI could make that history easier to question. It could also compress every caveat out of it. The design question is what survives the journey from one team’s evidence to another team’s default.
We distinguish four jobs that a “living organization” would need to connect. They happen at different speeds; they are not four departments to automate. The map below is a conceptual design, not a validated operating model.
In the example, marketing wants more availability, purchasing wants less cash tied up and finance wants better margin. The disagreement may be entirely legitimate. Shared facts help people see the trade-off; they do not decide whose objective should prevail.
Atlas conceptual model. Faster execution does not authorize a change in objectives, spending rights or company policy.
A 2026 paper by Nubank’s team describes an evaluation-led development process for customer-support agents in a Latin American banking context. Failures informed changes to instructions, tools and escalation, followed by checks against live outcomes.
In the card-delivery case, the team reports a 37 percentage-point improvement in transactional Net Promoter Score and a 29-point improvement in self-service rate over earlier variants. Those are different measures, and the changes bundled several interventions. They are company-reported results for that use case, not effects of memory alone.
The mechanism matters more than the headline: experience affected a specified part of the system, and that change was evaluated. People designed and revised the process. This supports a bounded improvement loop; it does not show a bank autonomously rewriting its strategy.
For the bag company, the equivalent would be a visible chain from a failed forecast to a changed decision rule and then to evidence from later decisions. “Our agents reflect after every task” leaves that last link unproven.
If every function retrieves the same approved explanation, agreement can look like independent confirmation. James March’s classic model of organizational learning shows why adopting shared beliefs too quickly can remove the variation needed to improve them. Applying that warning to AI is an inference: common definitions can help coordination while competing explanations remain worth preserving.
Authority creates a second problem. Aghion and Tirole distinguished the formal right to decide from effective control through information. An AI that selects the three options, drafts the forecast and summarizes the objections can shape a decision even if a manager clicks “approve”. That is our application of the theory, not a measured finding about current AI systems.
In the bag company, a purchasing assistant should not quietly turn a cash-saving target into the company’s definition of success. People affected by the decision need a way to challenge its assumptions, see excluded alternatives and escalate a conflict. Better prediction cannot determine the acceptable trade-off between margin, availability and jobs.
Alongside “Which tasks can agents perform?”, leaders need to ask: “Who may turn a local result into a rule everyone else must follow?”
Cleaner data, clear ownership, experiments and accessible decision records could explain much of an apparent transformation. None requires an autonomous company brain. This is a serious alternative to the claim that a new AI architecture made the organization learn.
A follow-up of a management experiment in Indian weaving firms found that about half the adopted practices had been dropped nine years later, while a substantial treatment–control gap remained. Managerial turnover and scarce director time were commonly reported explanations for the losses. This older, non-AI study shows why preserving a practice is an organizational problem; it does not prove that AI fixes it.
A fair test would compare two teams with reliable data, the same AI access and a simple decision log. Give one team curated shared memory and explicit revision rules as well. Use new decisions, include changed conditions and measure errors, business outcomes, review effort and total cost. Check whether useful lessons survive handovers and whether invalid lessons lose influence. This describes the comparison to earn, not a complete trial protocol: credible causal estimates need enough comparable decisions and a design that handles spillovers between teams.
If the simpler arrangement works as well at lower cost, keep it. If the richer one only repeats old answers more consistently, it has demonstrated standardization, not better adaptation. Atlas has not run this company-level comparison; the public research reviewed here does not settle it.
More capable AI does not imply a single shape for the company. Coase’s theory of the firm asks when coordinating inside an organization costs less than using the market. Applied to AI, the direction depends on which costs fall. Cheaper external coordination could support smaller firms and specialist networks. Cheaper internal monitoring could strengthen large, tightly controlled organizations. These are scenarios, not forecasts from the 1937 paper.
Greater adaptability is another possibility: teams form around problems, carry evidence and commitments across boundaries, and revise working rules as results arrive. They would still need lasting expertise, people responsible for consequences and ways to contest decisions. Otherwise, frequent reorganization could erase the knowledge it was meant to spread.
The useful ambition is a company in which a good lesson travels further without acquiring more authority than its evidence deserves. Start with one recurring decision that crosses functions. Find the point where an observation becomes a rule, and the point where that rule can be challenged. Those are more revealing places to invest than another universal memory layer.
Experience becomes useful when it can change a decision—and change it again.
The Indian management follow-up found that about half the adopted practices had been dropped, with a substantial treatment–control gap still remaining.
Scope and conditions: Nine-year follow-up in weaving firms; reported explanations for decay are not randomized mechanisms; not an AI study.
Supports a strong conventional-management comparator, not an AI effect estimate.
Published abstract About the source and its versionA retrospective selection: publications used in this question and its discovery, ordered by their publication dates. This is a small historical overview, not a complete account of the field or evidence of a research trend.
Publisher original consulted. Relative internal organization and market transaction costs, not an AI forecast. The possible organizational futures are Atlas scenarios.
Publisher abstract and research-hosted original consulted. The model concerns mutual learning between people and an organizational code; applying it to AI memory is an editorial inference, not a measured AI effect.
Original opening and introduction checked through the Duke-hosted paper. Formal rights and effective control differ. AI agenda-setting is Atlas’s application of that distinction, not an empirical result of the 1997 model.
Original full text checked, especially the repository of successes and failures. Historical account by company practitioners, not an independently randomized estimate of the archive’s effect.
Original abstract and framework checked. Decisions based on predictions can alter their target distribution. The bag-company example is Atlas’s illustration, not a case from this paper.
Published abstract checked; earlier full World Bank working paper consulted in the research. Indian weaving firms, revisited nine years after the original experiment. Reported reasons for abandoning practices are not randomized causes.
Published 2025; historical GPT-3-based workplace deployment, including 2020 outages. Mainly Philippines-based support workers serving US customers. Section VI.B and Figure VII consulted.
Version 1, Sections 4–6. Latin American banking context; five support use cases. V0–V10 includes baseline plus ten variants. Changes bundle prompts, tools, escalation and models; gains do not isolate memory or a single intervention.
Published version checked, including Table 2 and sample footnotes. 826 attended; 35 nonrandomized excluded; 791 randomized professionals, 776 with complete post-task surveys. One-day virtual workshops at one multinational firm. Company co-authorship and institutional funding disclosed.