AI Productivity Isn't AI Value

Most AI initiatives ship faster work. Far fewer ship more valuable work. That gap is the real story of AI adoption in 2026 — and most of the rollouts celebrated in the trade press are productivity, not value.

Three operators compared notes on what makes the question urgent: workflows that compress weeks into hours, a prioritization problem that flips when execution gets cheap, and the structural reasons large companies struggle to turn token spend into revenue.

What follows are notes from this week's Executive AI Roundtable discussion, shared under the Chatham House Rule.

When AI flips prioritization

The session opened with a concrete operator example. One founder runs a custom data-and-reporting business — Asana dashboards, API integrations, AI-driven automations for clients. His recurring problem was the one most teams have: people working on the wrong things. "Their intuition is often opposite of what I want."

He'd never tackled the problem head-on because the cost was too high — "it would take, like, 6 months of work to do." With AI, the build took an afternoon.

I chatted with Claude... and I built a prioritization hierarchy, and then Claude helped me build a pointing system for it, and then I started layering in all of the different other priority things that come into play. I ended up with something very complex, 12 criteria that all have different point values and weights.

The architecture splits the work cleanly. AI handles the subjective categorization — which value bands does this task fall into? A Python script handles the deterministic scoring — apply weights, compute urgency, write it back. The script runs hourly across every task in the system.

Every single person and every single meeting and every single board is sorted by urgency, and it updates live throughout the day, so we have this live, filtered view. Everybody knows the top 5 things they should work on. If we're in a meeting, the right things get discussed, because it floats to the top.

The pyramid at the base of the system reveals what the company actually values, in order: customers in onboarding (top, ~200 points), then in-flight work that's nearly done, then internal efficiency — "if there's anything that we can do that's gonna spend an hour today to save two hours tomorrow." Multiple bands compound: a task that's both onboarding and internal efficiency gets a proportional bump.

AI applies the company's value model in a system that runs continuously and surfaces value-aligned work to every employee. The team isn't choosing what to work on by gut anymore. They're seeing what creates the most value, encoded in the weights, surfaced in the dashboard, refreshed every hour.

The team loved it. The system started — the founder admitted — "in a fit of rage" after a few too many cases where someone tackled work in "the completely backwards order. What are you thinking?" What began as an attempt to spell out the obvious kept layering until it was running the whole company's review queue. The unexpected benefit landed at the top of the org: when his own review queue piled up to 27 items, the system surfaced the right three — "I can just grab the top 3 off the list, and those are the right 3." Institutional knowledge, crystallized into an AI-plus-software workflow.

He didn't stop at the internal tool. The same system got productized for a client with the same problem. Tokens turned into revenue, irrefutable evidence of real value.

A second operator extended the principle from the other end. Once execution is cheap enough to run overnight, the supply-side prioritization problem changes shape:

Prioritization only matters when there's a supply constraint and the AI goes so fast, relative to what human beings can do, that... if it just runs overnight and takes care of whatever it can take care of, maybe it's not all the most important stuff. But the field's been cleared, and the only thing I need to prioritize is whatever's left.

The whole prioritization apparatus — quarterly planning, scoring rubrics, capacity allocation — is built around scarcity. AI doesn't repeal scarcity; it just moves it. The new scarce resource is human judgment about what the AI produced, not human capacity to produce.

This comes back to AI as an opportunity engine. "If it turns out that it messed up something, well, whatever. It was a roll of the dice. And now I can roll the dice again, and it costs very little."

As another founder put it:

Tokens are incredibly cheap compared to my time.

Treat AI execution as expensive, and you spend too much human attention on prioritization and planning. Treat tokens as effectively free, and you can skip a decision by letting AI build a version featuring each possible answer in parallel. Then the human makes a decision based on the work product, not a prediction.

Alternatively, you do what the first founder did and encode the value model into the system so the right work surfaces automatically.

The two-phase workflow

The most interesting operational pattern of the session was a clean separation between thinking with the model and working with the model.

Don't start with a big plan. The early back-and-forth is prototyping — figuring out what the idea actually is.

a lot of what's happening early on is prototyping the idea. This could be a technical thing, this could be a writing thing, this could be a business thing, whatever.

Once the direction is clear, the work shifts:

I got an idea like, okay, well this is, I've got a clear sense of where this thing should go. We've described it. Now it's just about doing the work.

That's where the handoff happens. Auto modes in Claude Code and Codex get used aggressively at this point — ambiguous calls (yellow lights) go to the machine; blocking decisions (red lights) still come back to a human. The price is the occasional permission call you wouldn't have made yourself. On the other hand, you don't have to babysit your AI.

In one example of deep AI delegation, the clerical layer is given to another agent:

One thing I don't want to do is work for the AI. So I have the AI work for the AI.

In practice, that looks like Claude calling Codex as a sub-agent, or a planning agent orchestrating execution agents, with no human approval in the middle. If a human is the API between two models, the human is the bottleneck.

A practical note worth considering when kicking off a project: AI is bad at estimating its own work. The first operator's prioritization system flags "quick win" tasks for a point boost — but the AI can't identify which tasks qualify. "Every time you ask Claude to do anything, and it estimates one to two weeks of dev time to do this, and then it's done, like, an hour later." Don't ask AI to predict effort, and don't let it divide a project into phases based on inflated time estimates. Just run it in one go.

Productivity is not the same as value

Most large companies running AI initiatives haven't fully named the disconnect between productivity and value:

It's easy to take AI inside of a big company and apply it to something that, you know, is productive. But doesn't actually end up making value for the company, because the definition of productivity is actually not linked at all to value.

In practice, an enterprise can ship the AI initiative, hit every KPI on the rollout dashboard, automate a workflow that used to be painful — and still not see a dime of customer impact. Sometimes productivity is just internal busy-work made faster.

The test for every AI workstream: does this turn tokens into revenue? When the answer is "I'm not sure," the AI work is almost certainly productivity, not value.

The older form of the same lesson came up via Kodak — not the version most leaders learn (Kodak missed digital), but a sharper version:

Kodak was in the chemical business. And there was nothing about digital photography that was going to help them with their chemical business. What they missed was [there would] no longer be a demand for the [chemicals].

Are you in the transportation business or are you in the bridge maintenance business? Are you in the photography business or are you in the chemicals that make film business?

There's an AI-shaped version of this phenomenon where the activity is mistaken for the value. Companies trying to adopt AI today are full of software engineers — and many of them refuse to add AI to their idea of how great software is made. They're stuck in an occupational mindset, much like Kodak's chemists didn't reframe themselves as imaging engineers. The org becomes resistant to AI in any meaningful sense because the people closest to the work don't see it as their work.

Another value disconnect masked by productivity happens when metrics like AI token usage, email processing speed, or lines of code written fail to pull in more revenue. Streamlining a cost center might result in an AI win, but the value is invisible to customers, board, and revenue.

Why mid-market and below is winning AI right now

Larger enterprises have inertia from the organization they accumulated as they grew:

They built the structure because the structure was working, and now they have a hard time turning. And also like the individual pieces of the company that start to pick up this tech, right, are locally optimizing... local optimization, to your point, can be further away from customer value.

This is the constraint problem from The Goal — Eliyahu Goldratt's Theory of Constraints classic, and a useful book on this dynamic right now. Optimizing a part in isolation can make the whole system worse. A division gets praised for making their unit more efficient. The unit gets faster. Tokens get burned. Customer outcomes don't move, and meanwhile, other parts of the business are strained.

Smaller companies have an easier time. The CEO is close enough to the work to see whether their token budget multiplied a revenue outcome. The historical pattern: transformative change has always needed capital plus CEO-level operators in the room — and mid-market is where that combination still fits on one desk. McKinsey can sell strategy to a Fortune 500 board, but "selling transformative value to large companies is a hill that many have died on."

Among companies still figuring out their AI ROI, the ones that close the gap fastest are small enough that productivity and value decisions land on the same desk.

The labor market is already pricing this in

AI is reshaping the labor market for software work. Layoffs at Webflow and Wix hit that week — both consistent with a pattern of blocking-and-tackling roles thinning out first. The people staying in seats either have deep domain knowledge that the model can't replicate, or have moved into AI-native workflows that make them force multipliers. Roles that generated productivity without value, or through rote action, are the first to go.

The middle is the worst place to sit:

the people who have like 3 to 5 years of experience, they're the ones who... Because they are neither you nor are they native. And they're the ones who need to make the transition.

The advice for engineers in that zone:

You need to be thinking about how can you become an AI engineer? Or at least be on that trajectory within your organization by the time you're ready.

"AI engineer" means orchestration:

It's a lot more about systems thinking, having good judgment about the output, building the tools and scaffolding around the models.

The leverage point for engineers who don't want to be commodified by the model:

They need to be studying things the AI is dumb about.

Depth in something the model doesn't know is what gives you something to feed the model and judge model output against. Generalist skill at writing code is now table stakes. The asymmetry is in your domain knowledge — what's inside your head, not what's in the model's training data.

Where to start

Corrective moves:

• Audit your AI initiatives against revenue, not productivity. For every workflow you've automated, ask whether a customer would pay more or churn less because of the change. If the answer is no, reclassify or stop — the work was productive, not valuable.
• Separate prototyping from execution. Spend interactive time crystallizing the idea, then hand the execution to auto mode or an agent loop. Don't sit in the middle approving every step.
• Run parallel options. Tokens are cheap; your judgment is the bottleneck. When you can't tell which direction is right, run both and read the output.
• Resist local optimization. Before optimizing a unit's process, check whether the value of that optimization shows up at customer level. Theory of Constraints, applied to AI.
• If you're a 3-to-5-year engineer, move now. The mid-career zone is where AI compression hits hardest. Pick a domain the model is bad at — your industry, your codebase, your operational craft — and become the person who feeds the model rather than the one it replaces.

The fastest test: pick one AI initiative on your roadmap this quarter and answer the revenue question out loud, in front of someone who will push back.

For related takes, see the AI adoption arbitrage on why most companies are stuck, and where AI quietly goes wrong in small companies on local-optimization traps.

Resources From the Roundtable

• Claude Code — Anthropic's CLI coding assistant. Discussed extensively; auto mode preferred over per-step permission gates.
• OpenAI Codex — Recently added an auto mode of its own. One attendee uses Codex as a sub-agent called from Claude.
• Cursor — AI-first editor; mentioned alongside Claude Code.
• Webflow and Wix — both announced engineering layoffs the same week the roundtable met; cited as evidence of AI-driven labor compression.
• Base44 — Wix's AI app-builder acquisition; came up as an example of platform consolidation.
• Asana — One attendee uses it as the substrate for a custom AI priority dashboard.
• The Goal by Eliyahu Goldratt — Theory of Constraints classic; recommended in the session as the most useful book on this dynamic right now.
• Reengineering the Corporation by Hammer & Champy — historical reference for the consulting-driven optimization-and-rightsizing wave of the 1990s.