Every Decision You Automate Is a Loan

In June 2026, Ford ranked highest among mass-market brands in J.D. Power’s Initial Quality Study. The story behind the result should unsettle every executive currently celebrating what AI is saving them.

Ford’s turnaround was not built by deploying better algorithms alone. Ford’s remedy included bringing back experienced engineers, in a quality organization that had leaned harder on automated systems while shedding people. Over three years the company hired, promoted, or rehired 350 veteran engineers — many of them former employees, others recruited from suppliers — into its quality organization. Ford’s own word for them is “gray beard” engineers. Their assignment was telling: mentor junior engineers, rebuild the data pipelines feeding Ford’s machine-learning tools, and repair the automated systems themselves.

Ford’s quality problems had more than one parent, and we won’t pretend automation acted alone. But look past the press releases to the remedy, because a company’s remedy is its real diagnosis. Ford had automated large portions of quality decision-making while shedding thousands of salaried positions. The automated systems worked, in the sense that they ran, produced outputs, and were trusted. Its chief operating officer later described a period of relying more and more on those systems for results that came in below expectations.

The company has been unusually direct about what its automation strategy had missed. Its vice-president of vehicle hardware engineering said Ford had mistakenly assumed that introducing AI and ingesting its existing design requirements would by itself produce a high-quality product, and acknowledged that the company “didn’t pay as much attention as we should have” to the experience of its most knowledgeable engineers. Those engineers knew why a tolerance sat where it did, which supplier’s parts failed in which weather, what a particular squeak meant. The remedy suggests that important parts of that knowledge had left before they were captured. The knowledge wasn’t documented. It was employed. Then it wasn’t.

Note what Ford did not do: abandon the automation. It built a dedicated software quality-assurance team and deployed more than a hundred thousand automated tests. The fix was more machinery and restored judgment. What it had lost was not the technology but the capacity to tell when the technology was wrong.

So here is the question worth sitting with. On what financial statement, in any year from 2020 to 2023, did that exposure appear? The savings were booked immediately — headcount reduced, cycle times cut, efficiency gained. The obligation accumulating on the other side appeared nowhere. Ford’s leadership wasn’t reckless. Their conventional financial and operating instruments did not expose this particular dependency. That is precisely the problem: their normal management instruments did not appear to test whether the decisions they had delegated remained recoverable.

We call that exposure model debt — the obligation an organization takes on when it delegates a consequential decision to a system or external provider. That obligation comes due when the organization needs the decision back.

The phrase is already used in technical circles for something else: stale, undeployed or poorly governed models. That usage concerns the condition of the artifact. Ours concerns the recoverability of the delegated decision — whether it has been delegated to a system or an external provider.

The anatomy is worth stating precisely, because loose use of financial language is how this kind of argument gets dismissed. The principal is the judgment displaced: what you borrowed, and what reconstruction will cost when you repay. The interest is damage accruing while you cannot detect or correct errors, the wrong decisions made while blind, plus the value that expired because you could not act. The collateral is what still stands behind the decision: people who could rebuild the judgment, current documentation, working challenge mechanisms. And a call is any change that exposes the inability to retake the decision. For automated systems, drift — the widening gap between the world a system learned and the world it operates in — is a common cause of a call, not the interest itself.

Not every delegation qualifies. Nobody needs to rate the spam filter. What matters are decisions whose consequences outlive the moment: what a customer pays, who gets hired, which roles are eliminated in a restructuring, what ships, what a supplier is promised, and who gets credit.

To be clear about the accounting: model debt is not a liability in the reporting sense. No standard requires you to recognize it, and we are not suggesting one does. It is an operational exposure that conventional financial statements do not reveal. We use balance-sheet language because it gives managers a practical way to identify, measure, and manage that exposure.

The first test is simple enough to run this week. List each consequential decision delegated to a system or external provider, and name the people who could reconstruct it. If nobody can, the debt is already unsecured.

If Ford were an isolated stumble, this would be a curiosity. It is not. And the useful observation isn’t that reversals keep happening — it’s how sharply they differ in cost.

Commonwealth Bank of Australia moved in July 2025 to cut 45 customer-service roles, on the strength of a voice bot it said had reduced call volumes by two thousand a week. But the bank appears to have treated that one measure as evidence that the roles were no longer needed. Its frontline staff reported a different operational reality: total call volumes were rising, overtime offered, team leaders pulled onto the phones. The Finance Sector Union took the bank to the workplace relations tribunal, and within weeks CBA reversed the decision, apologized, and conceded that its initial assessment “did not adequately consider all relevant business considerations.”

But notice how the error surfaced. The people who knew otherwise were on the phones, with no way of being heard until a union gave them standing to force a review. That is not detection — it is being caught. The bank’s fast, cheap repayment was a function of having an organized workforce able to compel a second look, which many organizations do not have and no dashboard supplies.

Klarna took fifteen months. In February 2024 the company announced, alongside OpenAI, that its assistant had handled 2.3 million conversations in a single month — work it calculated as equivalent to roughly 700 full-time agents. By May 2025 its chief executive was telling Bloomberg that cost had been too dominant a factor and that “what you end up having is lower quality.” Recruiting resumed — though not like-for-like. The roles lost were full-time; the roles refilled were flexible and remote. Klarna settled a judgment loan with contract labor: different from what it borrowed, and thinner.

Three companies, three industries, one structure — and wildly different bills. Interest on this kind of loan does not accrue at a stated rate. It accrues at the speed of the borrower’s own blindness. One finding is clearer: none discovered the problem through an instrument designed to test whether delegated decisions remained recoverable. CBA was alerted by its union, Klarna by its customers, and Ford by its warranty claims.

In 2021 Zillow shut down its home-flipping business after buying homes at prices its own models could not sustain. Its filings put the cost plainly: a $408 million inventory write-down and $491 million in total wind-down charges for the year, with roughly 2,000 people cut. Zillow’s own explanation is the mechanism, admitted by the borrower: “the unpredictability in forecasting home prices far exceeds what we anticipated.” The decision was pricing, the failure was drift, and the purchases were houses. Nothing about that was reversible in weeks.

Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from zero in 2024. The next vintage is credit, pricing, supply commitments, hiring, engineering judgment — higher consequence, slower feedback, harder to undo. The early margin calls were a mercy. They arrived in decisions cheap enough to teach the lesson affordably.

If model debt behaves like credit, it can be organized using disciplines borrowed from credit analysis. We call the result a Model Debt Rating, and what it measures is decision recoverability. It relies primarily on verifiable evidence — names, dates, counts, and documented tests — rather than management confidence alone.

Start by listing the loans. Work through every area of the business and record what has been delegated: approve, price, rank, route, flag, reject. Include decisions embedded in your vendors’ software and decisions made by external providers on your behalf, since you carry the debt even though someone else makes or enables the decision. And include every workflow where a human reviews system output but has approved more than 95% of it for a year. Any entry without a nameable owner gets flagged immediately: unowned debt is already distressed.

Then audit the collateral, in four questions that rest on observable facts rather than confidence. Who could rebuild this decision’s logic from scratch — not operate the system, rebuild the judgment? When did someone last successfully challenge the system or external provider? When was the logic last tested against this year’s world rather than the one it was built for? And could a competent new hire reconstruct why the delegated decision is made as it is from what is written down, or does the reasoning live only in people’s heads, some of whom have left?

Then assess how serious the exposure is. Severity reflects consequence per error, decision volume, judgment displaced, rebuild time, reversibility, detection time, and opportunity expiry. Crossing a likelihood band with a severity band produces a provisional grade on a letter scale familiar to boards and finance teams. The enterprise grade is based on its highest-risk delegated decision, not an average — because averaging would let twenty trivial decisions conceal one potentially catastrophic dependency.

A rating is a forward-looking opinion, not a guarantee or a point estimate of loss. The grades below are retrospective illustrations, not validated calibration. They show that the method can distinguish among recognizable situations; they do not show that it would have produced those grades at the time.

Scored on what is now known, Ford in 2021 comes out at C — collateral severely weakened, revalidation absent, consequences severe, nothing reversible. The debt was called in 2023, and D records that event rather than forecasting it. Klarna in 2023 comes out at CCC: the same collateral failures, but small and reversible severity. CBA in 2025 comes out at B, caught in weeks rather than quarters.

And IBM illustrates what a stronger collateral position looks like. Its internal HR agent resolves roughly 94% of employee requests without a person touching them, and helped cut HR operating costs by about 40%. What it did not do was let the residual go unstaffed, or assume that a function running at 94% containment no longer needed people who could do the work. That is a borrower who read the terms.

The predictable objection: we have this covered — an AI governance program, a responsible-AI committee, model audits, perhaps ISO 42001 underway. Good. Keep all of it. And understand that none of it measures model debt, for a structural reason rather than a maturity one.

Model audits examine the system: is it accurate, biased, documented, secure. Those are questions about a piece of property. ISO 42001 obliges an organization to document the competences behind an AI system — but the requirement is that people be competent to run and oversee the system, not that the organization could make the delegated decision if the system were withdrawn. A reviewer empowered to halt an output is not a team that could rebuild the judgment behind it.

Banking provides the closest mature precedent. After the financial crisis, US regulators issued joint guidance on model risk management, known as SR 11-7. What it does not require is a demonstration that the bank could remake a delegated decision without the model — which is the gap this article is about. Its central principle is what it calls “effective challenge,” and it warns that validation matters especially during benign conditions. A regulator wrote down, in 2011, that the dangerous moment is the one when everything appears to be working.

Three moves this quarter, none of which requires a consultant.

Build the ledger: one ledger covering every consequential decision delegated to a system or external provider, an owner’s name beside each, with a one-page summary for the board. Put the CFO in the room alongside the CIO, COO and CHRO — the point is to see the list the way finance sees liabilities.

Review how staff reductions would affect the collateral behind each delegated decision. Before any restructuring takes effect, take the list of people leaving and check it against the names recorded behind each delegated decision. Where a departure would remove the last person who could rebuild a decision, that stops being a staffing matter and becomes a resilience decision.

Change how you borrow. From now on, no consequential decision gets automated or delegated externally without booking the loan: what judgment is being displaced, who owns the exposure, what a rebuild would cost, and how you would know if it went wrong.

Ford’s story will be told for years as a comeback. Read it instead as a disclosure: a company confessing, in public and in headcount, what it had owed all along. Your organization holds the same paper. The savings are on your income statement; the loan is not on your balance sheet; and somewhere, the world your systems learned is quietly ceasing to exist. Every decision you delegate or automate is a loan. The only question is whether you are the kind of borrower who knows the balance — before the margin call tells you.