Why don’t AI investments show up in the results?
The return on an investment comes from the workflow, not the tool. Distributing a tool is a purchase, not a change: AI layered on top of an old process produces, at best, time savings that leak away into emails and meetings. Returns only materialize when work is redesigned around AI.
Nearly every AI investment is justified by returns: faster, cheaper, fewer errors. Yet most of the leadership teams we work with cannot demonstrate those returns when asked. The fault is not in the technology. It is in how the technology was rolled out.
The pattern repeats in Astu Labs’ client work so often that it is almost predictable. An organization picks a tool, distributes licenses widely, and watches with satisfaction as usage climbs. The demos look convincing. Then the quarter ends, someone asks where the returns are, and no answer comes. The work was done exactly as before. AI was glued on top of the old workflow, and the workflow itself was left untouched.
Why do 95% of AI pilots fail?
AI pilots most often fail because the work does not change around the tool — not because the technology falls short. This is not an opinion. The most widely cited figure comes from MIT Project NANDA’s 2025 report, which estimated that roughly 95 percent of enterprise GenAI pilots produced no measurable impact on the bottom line. One can debate the report’s methodology. What matters, however, is not the number itself but the reason behind it. Pilots do not fail because the technology is immature — they fail because the work does not change.
The most precise evidence comes from a randomized field experiment. In a study published by Dillon et al. in 2025 (NBER), half of 7,137 knowledge workers across 66 companies were randomly given access to an AI assistant integrated directly into email, document work, and meetings. Usage was tracked for six months. The result was revealing. Those who used the tool saved a couple of hours a week on email and did less overtime, but the tasks themselves did not change. In practice, the benefit appeared only where an employee could change their work on their own — not where the change would have required organizational support and coordination.
The same logic was written into economics before AI. In 2021, Brynjolfsson et al. described a phenomenon they call the productivity J-curve. General-purpose technologies — electricity, the computer, now AI — require complementary, often intangible investments alongside them: new processes, skills, and ways of working. While those are being built, measured productivity can even dip, and only once the investments start to pay off does the curve turn upward. Anyone expecting AI to deliver a straight climb without the dip has misunderstood the nature of the technology.
Back in 2002, Bresnahan et al. showed with firm-level data that information technology, the reorganization of work, and new products or services are complementary innovations. The link showed up both in productivity and in the demand for production inputs, and IT’s effects were greatest where it was combined with organizational change. Two decades, three technology waves, the same result. Technology pays for itself through organizational change, not instead of it.
“But we did roll out AI across the whole company”
That is the most common objection, and it deserves a direct answer. Isn’t an organization-wide rollout a workflow change? It is not. Distributing a tool is a purchase, not a change. The process is still the same sequence of steps, in the same order, with the same responsibilities — just with faster tools.
The strategic mistake is graver still. When a tool is distributed without redesign, usage scatters into isolated experiments that no one leads, and the budget that was meant to prove the returns has already been spent. The next investment then becomes harder to justify.
Three questions worth asking right away
We encourage leadership teams to run one exercise at their very next meeting:
- Name one workflow you have redesigned around AI — rather than merely added AI to. If the answer does not come immediately, you have tools, not returns.
- Who owns the workflow and its outcome — not the tool, but the outcome? If the owner is IT, the change has not been made yet. The workflow is owned by whoever answers for its result in the business.
- Which single process would you change if you could change only one this year? Returns come from depth in one place, not from a thin layer everywhere.
Changing workflows is one of four reasons why every strategy question is ultimately an AI question. Alongside it run skills, cost structure, and governance — and behind them all, one thing no tool can replace: people, and how fast they learn.
Sources
- Dillon, E. W., Jaffe, S., Immorlica, N. & Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. Randomized field experiment: 7,137 knowledge workers in 66 companies over 6 months; the tool studied was Microsoft 365 Copilot. Working paper, not peer-reviewed; three of the four authors worked at Microsoft at the time of the study (openly disclosed in the paper). https://www.nber.org/papers/w33795
- Brynjolfsson, E., Rock, D. & Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics 13(1), 333–372. https://www.aeaweb.org/articles?id=10.1257/mac.20180386
- Bresnahan, T. F., Brynjolfsson, E. & Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics 117(1), 339–376. https://doi.org/10.1162/003355302753399526
- Challapally, A., Pease, C., Raskar, R. & Chari, P. (2025). The GenAI Divide: State of AI in Business 2025. MIT Project NANDA, July 2025.
