Can AI make a company learn? →
Current answer · Explainer
Organizational learning needs leadership that can allocate authority and invest in checks and feedback.
As AI takes on more work, a CTO must help decide which judgements can be delegated, what evidence justifies that authority and how the company can recover when a system fails. The role combines technical architecture with investment, expertise and decision rights; the business still owns its goals.
Current answer · Explainer
Following covers this question’s published contributions. It does not follow neighbouring questions or subscribe you to email.
AI strategy for technology leaders: decisions, investment, expertise and three possible paths to 2030.
An argument about leadership grounded in research, not a validated CTO model or a forecast. The launch example is fictional.
Article version history → The CTO’s AI mandateThese narrower questions are not promises of a future article.
Longitudinal comparisons of alternative decision arrangements, with business outcomes, human skill, costs and reversals recorded.
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 leadership question connects AI capability with decision rights, investment and the expertise required to supervise it.
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 the CTO’s changing AI responsibilities through a fictional product launch, original research and conditional operating choices for 2030.
No earlier reviewed assessment is available to compare. No change in the underlying facts is being claimed.
Explains the CTO’s changing AI responsibilities through a fictional product launch, original research and conditional operating choices for 2030.
Compare complete workflows on new work, including review, failure recovery and the cost of keeping alternatives viable.
Longitudinal comparisons of alternative decision arrangements, with business outcomes, human skill, costs and reversals recorded.
As AI takes on more work, a CTO must help decide which judgements can be delegated, what evidence justifies that authority and how the company can recover when a system fails. The role combines technical architecture with investment, expertise and decision rights; the business still owns its goals.
An argument about leadership grounded in research, not a validated CTO model or a forecast. The launch example is fictional.
In the BCG analytical arm, unaided consultants more often selected the correct answer than either AI group.
One deliberately challenging task; June 2023 GPT-4; not current-model performance.
AI assistance helped commercial and R&D participants produce less functionally divided proposals.
One-day P&G workshops; no established effect on actual product launches or long-term expertise.
Inspection staff favoured their own priorities over algorithmic recommendations, limiting potential gains.
One US city pilot; authority arrangements not randomized; non-generative algorithms.
The productivity J-curve model accounts for intangible investments alongside new technology.
Historical US analysis and economic modelling; not a forecast of project profitability.
NIST’s 2024 profile recommends fallbacks, third-party incident plans and criteria for deactivation.
Voluntary institutional guidance; not a certification or measured performance result.
Bainbridge describes how automation can leave operators with difficult exceptions while removing routine practice.
1983 industrial automation analysis; application to AI knowledge work is an Atlas analogy.
Csaszar emphasizes verifiable computational strategy; Felin emphasizes causal theory and actor-specific judgement.
Attributed positions in a scholarly debate; neither is presented as settled fact.
METR judged its newer productivity data unreliable because of selection and measurement problems.
February 2026 update; no reliable present-day speedup is inferred.
Singapore reported 28.5% had started adoption and 3.8% were integrating AI into core processes.
Both shares concern all covered private-sector establishments with at least ten employees; employer reporting, not causal evidence.
A fictional lighting company wants to launch a rechargeable desk lamp. Its AI assembles customer complaints, supplier specifications and margin estimates, then proposes a design, a price and a launch plan. The presentation takes minutes. The executive team still approves the launch.
But suppose the system ranked designs by margin on the initial sale, excluded a more repairable option and treated a supplier’s battery-life claim as established. The team can approve every slide without ever approving those choices. The person at the end of the workflow has formal authority; the process has already shaped what that person can see.
For a chief technology officer, these are architectural choices with commercial consequences. Which objectives enter the system? What evidence is required before a proposal becomes a commitment? Who may change either? The comparison below holds the model and source material constant while changing who makes the decisions.
“Prepare a launch plan for a rechargeable desk lamp.”
Fictional company. Same model and available documents in both routes; only decision ownership and checkpoints change.
The team reviews a finished recommendation after key choices have already narrowed it.
AI defaults
Margin becomes the ranking objective; repairability drops out.
AI summary
Supplier claims and measured tests appear in one recommendation.
Executive approval
The team sees the shortlist, then purchasing receives an approved plan.
Access boundarySales report
Launch sales are recorded; rejected alternatives and assumptions are not.
A human approved the plan. The company still cannot tell which earlier choices deserved that approval.
Authored illustration, not a company case study, tested workflow or live AI run. It demonstrates a design choice, not a measured performance improvement.
The team reviews a finished recommendation after key choices have already narrowed it.
A human approved the plan. The company still cannot tell which earlier choices deserved that approval.
Objectives, evidence and spending authority each have an identifiable owner.
More decisions are visible and contestable. Whether the extra review earns its cost still needs measurement.
In a BCG experiment using June 2023 GPT-4, AI improved speed and quality on product-development tasks. In a separate analytical task, AI-assisted consultants were less likely to reach the correct answer. The published 2026 paper makes the distinction unusually visible: fluent advice and correct judgement can move apart.
Calling an entire department “AI-ready” tells us too little about what to delegate. In the lamp launch, generating designs, checking battery tests and authorizing a purchase need different evidence. A strong result on the first does not establish reliability on the others.
Imagine the lamp team produces ten times as many proposals while its engineers can still validate only the same number. More drafts could create a longer queue, rushed checks or more discarded work. The constraint has shifted from producing material to deciding what merits attention. This is a capacity argument, not a forecast of AI’s productivity effect.
The opportunity is larger than automating the existing queue. In the P&G field experiment, 791 professionals in Europe and the Americas took part in bounded innovation workshops. AI helped assisted participants bridge commercial and technical perspectives. That suggests a way to explore designs earlier across functions; it does not establish better launches or durable team learning.
A CTO should therefore ask where earlier access to expertise could remove a handoff, which checks can be automated reliably, and which need more human capacity. Measure the whole process: elapsed time to a useful release, defects, rework, customer outcomes and total cost. Drafts per employee cannot show whether it has improved.
“Keep a human in the loop” is an incomplete answer. In a US restaurant-inspection pilot, algorithms identified more violation-prone restaurants than inspectors’ rankings. Yet inspectors disproportionately followed their own priorities, limiting the potential gains. Hyunjin Kim and colleagues examined the pilot and interviewed 55 departments; they did not randomly assign alternative authority structures.
The result concerns predictive algorithms in one municipal setting, not generative agents. It offers a useful counterweight: discretion can correct a model or weaken a better recommendation. Neither automatic obedience nor automatic override is a sound default.
For the lamp company, record why a reviewer rejects an option and examine later outcomes where observable. Give reviewers original evidence, time and permission to stop a commitment. Give customers and employees a way to challenge decisions. A check that never changes anything deserves investigation; it does not prove that the system is flawless.
Buying a capable model does not require outsourcing the definition of good work. The lamp company’s valuable assets may include reliable product facts, test methods, customer access and evidence connecting design choices to returns. A CTO can buy generation while retaining the means to assess a result, switch suppliers and recover from failure.
That requires technical work: reliable integrations, access controls, versioned evaluations, records of tool actions and recovery procedures. A spending limit in a prompt is not an enforced permission. Replacing a model can change behaviour even when the interface stays the same, so important workflows need representative checks again.
NIST’s 2024 generative-AI profile recommends documented third-party dependencies, fallback arrangements and criteria for disabling systems. It is voluntary US guidance, not proof that any particular architecture works. Our practical implication is to test an exit before critical operations depend on a system: can the company retrieve its records, run a viable alternative and reconcile unfinished actions?
Self-hosting everything does not automatically create independence. It can exchange supplier dependence for maintenance work and reliance on scarce internal expertise. Keep the parts whose ownership improves differentiation or control; buy the parts whose replacement you can realistically manage.
The same launch connects three investment agendas. Operational AI could reduce work in supplier reconciliation and support. AI-assisted innovation could uncover a better design. AI-mediated commerce could change how an assistant discovers, compares and recommends the lamp. These opportunities need different success measures: reliable throughput, better validated products and qualified demand.
Shared product facts and a trustworthy test record could support all three. A universal company-wide agent platform may be unnecessary. Start from the commitments each workflow needs, then build the smallest common foundation that removes repeated work.
Brynjolfsson, Rock and Syverson’s productivity J-curve explains how investment in processes, skills and other intangible assets can precede visible returns. Their economic model and historical US analysis help explain a measurement problem; they do not make every slow AI project a future success.
Separate the cost of creating reusable capability from the cost and benefit of each use. Agree with finance how time saved becomes capacity, service quality or cash savings. Reserve room for experiments that could change the offer itself, and specify what evidence would stop them. Otherwise, efficiency projects with easily counted benefits can consume the entire innovation budget.
Lisanne Bainbridge’s 1983 analysis of industrial automation exposed a persistent tension: routine automation can remove the practice needed when a person must handle an abnormal situation. That is an analogy for AI work, not a measured generative-AI effect.
In the lamp company, senior engineers may catch a dubious battery claim because they have tested many products. A junior colleague who only approves generated summaries may never acquire that judgement. Training should include independent predictions, direct examination of evidence and unusual cases, alongside prompting skills. Which practice to retain depends on the decisions people will still own.
Knowledge sharing also has an incentive problem. If every documented improvement is treated as a reason to eliminate the contributor’s role, asking people to share failures and tacit expertise may produce silence. This is a design risk to investigate, not an assumption about employees. Leadership and HR need a credible account of how roles, recognition and gains will change.
A 2026 Strategy Science debate makes the disagreement explicit. Felipe Csaszar, at Michigan, argues that computational search and evaluation could expand strategic capability if outcomes become sufficiently verifiable. Teppo Felin, at Utah, argues that distinctive strategy depends on causal theories and actor-specific judgement that pattern-based systems cannot supply. These are competing scholarly positions, not settled boundaries.
The lamp company need not resolve that debate to act. Let AI surface options and objections; require an explanation of why this company could win, what would disprove it and which real-world observation comes next. Synthetic customer approval is a hypothesis about customers, not evidence from them.
For the CTO, this calls for several kinds of judgement: technical depth to investigate failures, experimental understanding to question a result, commercial understanding to recognize differentiation, and organizational skill to negotiate ownership. The title does not grant authority over every decision. Business leaders own objectives, finance owns agreed capital limits, and the CTO makes the technical commitments inspectable and enforceable.
There is no evidence-based date when a company should become autonomous. Singapore’s April 2026 official release reported that 28.5% of covered private-sector firms had started adopting AI, while 3.8% were integrating it into core processes. Coverage was establishments with at least ten employees. This is a local, employer-reported snapshot, not a global maturity curve.
The paths below are conditional choices from a November 2026 starting point. A company could use all three at once in different workflows. More delegation earns its place when outcomes, recoverability and economics support it; a change in conditions can justify moving back.
Atlas scenarios, not predicted adoption rates. No path authorizes an AI to redefine business goals, spending rights or acceptable harm.
A strong objection is that much of this is simply good technology leadership: clear goals, reliable systems and thoughtful investment. That is partly right. AI makes these disciplines urgent when proposals become cheap, outputs vary and systems can move from recommending work to executing it. It does not make the CTO the company’s sole strategist.
Begin with one consequential workflow and a baseline that already has clear ownership and decent data. Compare the AI arrangement with that baseline on new work, tracking outcomes, review, failures and cost. Revisit the evaluation when the model or workflow changes. METR’s February 2026 update illustrates why: recruitment and task-selection problems made its newer developer-productivity data an unreliable estimate of the current effect.
For the first investment cycle, agree who may change workflows, who bears the downside, how spending is released and what triggers a stop. Use the next cycle to expand only where the evidence warrants it. Keep open the longer-term possibility of a different product or route to market, without treating the 2030 story as a spending commitment.
The CTO’s durable contribution is a company that can adopt a better capability without losing the ability to question it, change it or continue without it.
Deciding what may become a commitment is part of the architecture.
Inspection staff favoured their own priorities over algorithmic recommendations, limiting potential gains.
Scope and conditions: One US city pilot; authority arrangements not randomized; non-generative algorithms.
Published pilot and follow-up interviews distinguish prediction quality from implementation.
Sections 5.1–5.2 and 6 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.
Original published paper, pp. 775–779. Industrial control and aviation examples; application to generative AI is an analogy, not a measured effect.
Published journal abstract and authors’ working-paper description checked. Concerns complementary investment and measured productivity; no guarantee of returns to an individual AI project.
First online 5 December 2023; April 2024 journal issue. Published version after 31 January 2024 appendix correction. Pilot in 2016; later inspection outcomes to spring 2022. Predictive algorithms, not generative AI.
July 2024 profile checked, including GV-6.2 and MG-2.4. NIST says AI RMF 1.0 is being revised; this is the named 2024 document, not a legal requirement or proof of effectiveness.
Update distinguishes early-2025 results from later data affected by recruitment, task-selection and timing problems. No reliable current speedup inferred.
Published in 2026; experiment used June 2023 GPT-4. 758 BCG consultants in two separate task arms; outside-frontier arm n=373. Figure 5 and section 4.2 checked.
Contributions developed from an August 2025 conference; published March 2026. Csaszar and Felin sections checked. Their opposing positions are attributed arguments, not settled limits of AI.
Private-sector establishments with at least 10 employees in Singapore. 28.5% had started adoption; 3.8% of all covered firms integrating AI into core processes. Survey reporting, not causal evaluation or a global estimate.
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.