AI Doesn’t Fix Operational Debt. It Compounds It.

Every enterprise carries operational debt: the undocumented workflow that lives in one person’s head, the data nobody fully trusts, the approval step nobody remembers designing, the report that gets hand-fixed every month before anyone sees it. And every enterprise is currently being sold the same comforting story — that AI is the thing that will finally clean it up. Point the machine at the mess; the mess resolves.

The opposite is what actually happens, and we now have large-scale evidence for it. AI is a multiplier, and a multiplier doesn’t care what it multiplies.

The J-curve is now measured, not theorized

Researchers from Toronto, Stanford, Oklahoma and the U.S. Census Bureau studied AI adoption across tens of thousands of American manufacturing firms, using Census survey waves from 2017 and 2021. The finding: adoption produces a J-curve — a real, significant productivity decline before the gains arrive. Firms that weather the dip go on to outgrow their non-adopting peers in productivity and market share. But the dip is not a rounding error, and it is not evenly distributed: older, established firms take the deepest losses, while younger ones adjust faster.

Two details in the data should stop every operator cold. First, the adjustment costs are visible as rising work-in-progress inventory — the factories literally accumulated half-finished work while absorbing the technology. That is the last essay’s argument showing up in physical goods: accelerate one step of a constrained system and you don’t get output, you get inventory. Second — and this is the mechanism nobody budgets for — among older establishments, roughly one-third of the losses trace to firms abandoning the structured management practices they had been running on. Adoption didn’t just strain their operations. It knocked out the instruments they were using to manage their operations.

Read that again the way a pilot would: the turbulence hits, and the gauges go dark at the same time.

Why the dip is deepest where the debt is highest

Automation consumes structure. An agent needs what a new hire needs on day one, except it can’t ask around: clean data, a defined workflow, a clear owner, an explicit boundary on what it may do. Where those exist, the agent compounds them. Where they don’t, the agent doesn’t fail loudly — it improvises, encoding the tribal workaround at machine speed. Operational debt compounds through four mechanisms at once:

Volume. A process that produced errors at human pace now produces them at machine pace. The defect rate didn’t change; the denominator exploded.

Opacity. The human workaround was at least visible — someone knew it was a workaround. The automated version disappears into logs and becomes, functionally, the process.

Authority ambiguity. Nobody defined what the agent may do because nobody had ever written down what the humans could — the first essay in this series in its most expensive form. Undefined authority is operational debt with an execution engine attached.

Verification overload. The review stations that debt already made slow are exactly where the new volume lands. The queue doesn’t grow linearly. It compounds.

Measure all three families, or the J-curve will kill a good program

Boardrooms have already made one correct move this year: shifting from usage metrics — seats, tokens, adoption leaderboards — to the returns question. But during the dip, a rigid focus on returns is its own trap: the J-curve research builds on a decade of evidence that early gains from a general-purpose technology are systematically understated, because the value being created is intangible — redesigned processes, cleaned data, new skills — and none of it shows up in the quarter’s output numbers. Judge the program on outcomes alone in month six and you will cancel it at the bottom of the J, right after paying the tuition and right before collecting the return.

The operating answer is three families of measures, run simultaneously. Input measures — adoption, usage — tell you the experiment is actually running; they say nothing about value. Outcome measures — productivity, cycle time, quality, customer results — are the point, but they lag, and they must be designed to catch the negative alongside the positive: more output is not more value, as one health dataset’s 190 formulaic papers demonstrated. Organization measures are the leading indicators the dip demands: share of workflows documented with a defined outcome and a named owner, data readiness in the processes being automated, skills actually built, and — per the manufacturing evidence — whether your management instruments survived the transition. During the J, the third family is the only honest progress report you have.

What to do now

Pay the debt where the agent will run — first. Document the process, define the outcome, name the owner before deployment. A pilot’s first deliverable is a process spec, not a demo. The sequence is the strategy: data, workflow, accountability, authority, evidence — then scale.

Budget the dip out loud. Put the J-curve in the board deck before the program starts: expected depth, expected duration, and the organization measures that count as progress while outcomes lag. A dip you predicted is an investment. A dip you didn’t is a scandal.

The failure mode here has an address: Pilot City. A pilot becomes a bridge only when its learning changes roles, decisions, workflows, incentives, controls, budgets, and performance measures. If nothing about the operating model changes, it wasn’t a bridge — it was a suburb, and programs retire there.

Re-instrument in the same sprint you deploy. If AI changes how the work flows, your old dashboards are measuring a process that no longer exists — the one-third finding is what happens when nobody rebuilds them. Treat management instrumentation as part of the deployment, not an afterthought.

Define authority before automation. Scope what the agent may do — which actions, which data elements, which limits — before it runs. Ambiguity that was survivable at human speed is a liability at machine speed.

Capture evidence from day one. Decision evidence — what was authorized, what was done, what was verified — is what turns the dip into a measured investment rather than a leap of faith. It’s the same record I’ve been working on in the open in a proposed OCSF extension: one ledger that serves the auditor, the QBR, and the board’s J-curve review alike.

Compounding runs both directions

The J-curve is not an argument against AI. It is the shape of every general-purpose technology paying down the complementary investments on the way to the payoff — electricity did it, computing did it, and the firms that came out ahead were the ones that rebuilt the factory around the technology instead of bolting it onto the old floor plan. Pay the operational debt deliberately and you get a shallow, short J. Scale on top of the debt and you get a deep one — or a canyon you never climb out of.

Because the multiplier doesn’t care what it multiplies. Fix the process, and AI compounds the fix.

Sources

• McElheran, K., Yang, M., Kroff, Z., Brynjolfsson, E. — The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s), U.S. Census Bureau Working Paper CES-WP-25-27 (2025): census.gov
• MIT Sloan, Ideas Made to Matter — The ‘productivity paradox’ of AI adoption in manufacturing firms: mitsloan.mit.edu
• Brynjolfsson, E., Rock, D., Syverson, C. — The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, American Economic Journal: Macroeconomics 13(1), 2021: aeaweb.org

This is the fourth essay in Field Notes on the Agentic Enterprise. Previous: Verification Is the New Bottleneck in Enterprise AI. Next: “When AI Agents Manage Your Money, Who Controls the Agent?”