9 min. read

Public markets have already priced a large AI payoff into the companies that use software, not just the ones that build AI. A September 2026 NBER working paper estimates that AI news from November 2022 through 2025 raised the market’s expected value of software-engineering productivity by the equivalent of a permanent 32.6% increase, concentrated in firms where software engineering is a large share of payroll.

That expectation now sits in those companies’ valuations. For boards and CFOs, the question is no longer whether AI matters. It’s whether the organization is built to deliver what the market already assumes.

Key facts

32.6% Market-implied equivalent of a permanent productivity increase in software engineering, Nov 2022–Dec 2025. A market expectation, not measured output; working paper, not peer-reviewed. Source: NBER Working Paper 35793.

2×+ By mid-2026, the market-implied effect had more than doubled relative to the end of 2025, the same paper finds. Source: NBER.

240% → 30% In a separate NBER study of AI coding tools, a 240% cumulative effect on commits fell to 80% for projects and 30% for actual releases. Source: NBER Working Paper 35275.

5–8¢ AI revenue per dollar of AI infrastructure spending, by two 2026 analyses built on similar figures. Sources: Tom Tunguz; Asymco.

What the market priced, and where

The paper, by economists at UC Berkeley and the London School of Economics (NBER Working Paper 35793), reads AI’s expected payoff from stock prices. It measures how each firm’s returns move with AI news, and how that sensitivity rises with the share of payroll the firm spends on software engineers.

Two design choices make it relevant to a board. The sample excludes companies that produce or sell software and companies in the semiconductor supply chain, and the results are driven by the firms with the most software engineering in their own operations. By design, then, the estimate isolates the payoff the market expects at companies that use software engineering, not the ones that sell it.

The size is large. The authors translate the market-implied gain into a permanent increase of about 3.6% in the level of GDP in their baseline, or about 6.5% if better software engineering also raises R&D productivity. And they find that news in the first half of 2026 exceeded the total over the previous three years, so that by mid-2026 the effect had more than doubled.

Your entry could not be saved. Please try again.

We sent links to your email! You should have received an email from us already. If you did not receive, make sure you check your spam folders and add masterofproject.com to safe senders list to receive our emails.

Read it with its caveats. It’s a working paper, not yet peer-reviewed. It measures what investors expect, not output already delivered. And it covers one channel, software engineering, not every way AI could matter. The authors’ central estimate, 32.6%, falls to 22.5% or 14.0% under alternative assumptions about how readily inputs substitute for one another, and rises under others.

The market moved first

On the builders’ side, the money is flowing out faster than it’s coming back. Tom Tunguz estimated in March 2026 that “for every dollar hyperscalers earn from AI today, they’re spending twelve dollars to build more capacity.” Horace Dediu of Asymco, working from Milk Road AI’s figures, put AI services revenue at about 5% of what’s being spent on infrastructure. Somewhere between a nickel and a dime of revenue for every dollar of capital.

In our reading, investors are underwriting that gap because they expect the return to show up downstream, inside the companies that put AI to work. If your company is listed and heavy on software engineering, some of that expectation is likely already in your price; that’s our inference from the paper, not a firm-level estimate. The awareness gap is closed. The org-chart gap is still open.

Why the payoff isn’t automatic: weak links

In sizing its estimate, the paper notes that “total productivity is constrained by the least-improvable steps, so task-level speed-ups overstate overall gains.” A separate NBER study of more than 500,000 developers shows what that looks like in practice (Demirer, Musolff and Yang). Autonomous coding agents raised commits by a cumulative 240%, but that fell to 80% for the number of projects and 30% for actual releases. Across four major software marketplaces, the authors found “a sharp increase in the number of new apps but no increase in total usage.”

A third NBER paper points to where the durable gains come from. Babina, He and Jiang find that early “productivity gains are driven by AI-skilled jobs that build organization capital”: the processes, know-how and ways of working that outlast any single tool.

Put together, the binding constraint is rarely the step AI already made faster. In our reading, it sits in the steps around it: data readiness, legacy systems, approvals, regulation and how quickly decisions get made. Those are organizational problems, and they need someone whose job is to find and widen the slowest step. We map the five places AI pilots stall in a separate post.

Five questions for the board

These are the questions a prudent board or CFO can ask before an investor, auditor or acquirer asks them first.

Question	Why it matters	What a good answer looks like
1. Which AI systems are in production, and who owns each one by name?	In our reading, the payoff the market expects comes from AI in use, not AI in pilots. A system with no named owner has no one accountable for it.	A one-page list: system, business owner, internal delivery owner, date it went live.
2. What share of the AI use cases we approved in the last 12 months reached production?	Approvals measure intent; production measures delivery. The gap between them is where the expected value sits unrealized.	A ratio, tracked quarterly, with the stalled cases named and the reason for each.
3. For each major initiative, what is the binding constraint, and who is working on it?	The slowest step sets the pace. Speeding up a step AI already made fast doesn’t move the result.	One named constraint per initiative (data, legacy systems, approvals, regulation) and an owner for it.
4. How much of our deployment work do vendor teams do, and what stays with us when they leave?	Renting gets a first rollout moving. When the engagement ends, the learning leaves with them unless someone inside was building it.	For each vendor-led deployment, the internal person who shadows the work and owns the system afterward.
5. Which AI-skilled roles do we have on our own payroll, and is that number growing?	Early AI productivity gains are driven by AI-skilled jobs that build organization capital, the capability that outlasts any single tool.	A count of internal AI delivery roles, with a plan for next year that doesn’t depend on one vendor.

One diagnostic

If your largest shareholder asked which named person owns getting each of your AI systems into production, could you answer on one page?

If not, the market’s expectation is sitting on your valuation with no one on the org chart accountable for it.

Own the muscle

Large companies can rent deployment talent, and for a first rollout many should; vendor teams and implementers are often the right first move. But the research on organization capital suggests the durable gains go to companies that build capability in their own people. When a vendor engagement ends, the learning leaves with them unless someone inside was building it all along. We cover that handoff in who owns your AI after go-live.

The org-chart answer is a role: an AI project manager on your own payroll, accountable for getting each system into production and owning it afterward. Our AI project manager job description template defines it, and build, buy or rent compares the ways to fill it.

Master of Project Academy sponsors the nation’s first registered apprenticeship dedicated to AI Project Management, registered by the Maryland Apprenticeship and Training Council (MATC #2575, September 8, 2026) under the registered occupation Information Technology Project Manager. The registration was announced via Business Wire.

The AI Project Management apprenticeship explains how employers grow the role from their own staff.

Frequently asked questions

Is AI already priced into stock valuations?

For companies that use software engineering heavily, a September 2026 NBER working paper says yes: AI news from November 2022 through 2025 raised the market’s expected value of software-engineering productivity by the equivalent of a permanent 32.6% increase, and the effect more than doubled by mid-2026. It’s a market expectation, not measured output, and the paper isn’t yet peer-reviewed.

Does the market expect AI’s payoff at the companies that use it?

That’s what the paper is designed to measure. It excludes companies that sell software and the semiconductor supply chain, and finds the market-implied gain is driven by firms with the most software engineering in their own operations.

Why don’t AI speed-ups translate one-for-one into company results?

Because the slowest steps set the pace. One NBER study found a 240% cumulative increase in commits from autonomous coding agents fell to 80% for projects and 30% for actual releases, with more new apps but no increase in total usage. In our reading, the gains depend on fixing the organizational steps around the fast one.

What should a board ask about AI deployment?

Five questions: which AI systems are in production and who owns each by name; what share of approved use cases reached production; what the binding constraint is on each major initiative; what stays with the company when vendor teams leave; and how many AI-skilled roles are on the company’s own payroll.

Own the muscle the market already priced.

The AI Project Management apprenticeship, registered under the Information Technology Project Manager occupation (MATC #2575), develops one of your own employees into the role, on your projects, as a full-time employee from day one.

See how employers join the apprenticeship →

About Master of Project Academy

Founded in 2012, Master of Project Academy has served 500,000+ enrollments across 180+ countries in project management training. It sponsors the AI Project Management apprenticeship, registered under the Information Technology Project Manager occupation (MATC #2575). Learn more at masterofproject.com/p/aipmapprenticeship.

Sources

It’s Already in Your Multiple: Markets Priced the AI Payoff Before Most Companies Organized for It