The week the accountability stayed put

The signals reshaping how organizations deploy AI arrive from outside the room — from the labs, the payroll data, the policy drafts, the earnings calls. This week they arrived with the same finding five times over.

Everyone is trying to hand something to the machine. This week showed, five different ways, which part doesn't go.

Five signals.

1. The return was never in the tool. It was in the wait.

An investor deck published this week did something the ROI debate has mostly avoided: it put the task-level, firm-level, and industry-level evidence side by side. Read that way, the pattern is plain. AI reliably raises output wherever it lands. It does not reliably raise return, because the return sits in whatever step the AI didn't touch. Across more than half a million developers, commits rose about 240% with coding agents, while shipped releases rose about 30% — review and release didn't speed up. Scoring of S&P 500 annual reports puts non-tech firms stuck at pilot stage two to three margin points below firms with no AI at all. The handful that rebuilt their work around AI run well ahead, though that figure rests on 19 firms and is an association, not a proof.

The deck's oldest example is the sharpest. An insurance application worked on for 17 minutes sat in process for 22 days. Double the speed of every step and it still takes about 22 days. The return comes from removing the wait, not from speeding up the work.

→ Before you fund the next AI project, time where the work waits, not how long each step takes. On IBM i the slowest step is rarely writing the code. It's review, change control, promotion, and the human signature. Replace "hours saved" with "cost per accepted result," count the failed attempts and the checking time, and name the budget line that changes if it works. If you've added AI beside an unchanged process, the evidence says you may be behind a shop that never started.

Source: Social Capital / Learn With Me, "AI ROI: Why So Few Firms See a Return on AI Spend," synthesizing NBER, METR, MIT FutureTech, PwC, and Ramp research, September 2026.

2. The mandate produced volume. The responsibility didn't come with it.

The CEO of a major commerce platform made AI a baseline expectation last year: prove you can't do it with AI before you ask for headcount. This month he named the side effect himself. Unvetted AI output lands on a colleague's desk, and the colleague does the review and the repair. The canonical case is a round trip — a long AI-written email that the recipient runs through another model to shorten. Output goes up, net work goes up, and understanding stays flat.

Researchers have a name and a number for it. "Workslop" is polished-looking AI output that needs rework to be usable. More than half of 962 surveyed desk workers had sent it. Recipients spend about three and a half hours a month cleaning it up, which comes to roughly $9 million a year for a 10,000-person company. This isn't a skeptic talking — his own company runs an internal agent that may handle about half of production code changes. And a second AI-first CEO moved to demanding human oversight the same month.

→ A volume mandate assumes responsibility travels with the output. It doesn't; it pools on the next desk. Built-in AI keeps a named human as the accountability layer for anything that leaves their hands. Measure what your people receivefrom AI, not only what they send.

Source: Fortune and Business Insider on a commerce-platform CEO's podcast remarks; BetterUp Labs and Stanford Social Media Lab "workslop" research, September 2026.

3. "The model did it" is not an answer.

A Wall Street Journal op-ed argued that the labs' loud warnings about their own systems double as reputation repair, and that talking about models as if they had intent lets builders shift accountability onto "the system." The motive claim is arguable. The accountability claim holds. Every sentence that starts "the model decided to…" moves the subject away from the people who configured, authorized, and deployed it.

The op-ed overreaches in the other direction too. "It's just math" doesn't make an agent safe; goal-directed systems produce outcomes nobody asked for without any intent at all, as this summer's eval escapes showed. Intent is the wrong axis either way. The question that survives is simpler: whose authority was that action taken under, and can you prove it?

→ Build so the question never has to be argued. Every agent carries a persistent identity tied to an accountable person or team, acts inside a bounded scope, and leaves a record you own. On IBM i, object authority and the audit journal already answer "who, under what authority, and when." When the incident review starts, "the agent did it" should be an incomplete sentence.

Source: WSJ Opinion, "The Hidden Agenda Behind the AI Panic," September 2026.

4. The market priced the customers who never bother.

A consumer AI agent that will actually cancel the forgotten subscription, shop the insurance renewal, and move idle cash to a better rate went to the top of the app charts this month. On Tuesday the market repriced the industries that depend on customers not doing those things. Financial stocks fell as much as 2.4% in a day; big banks dropped more than 2.5%, and an insurer and a brokerage more than 5%. A trading-desk note named the mechanism: businesses built on recurring bills, add-ons, and customer passivity are exposed once an agent does the shopping.

The access fight is the part to watch. One major retailer blocked the agent at the door, citing unauthorized access. A commerce platform opened a governed checkout lane for it instead.

→ Banking and insurance are the heart of the IBM i install base, and this is the admission question with a price on it: which agents get in, on whose authority, with what audit trail. You can block them at the door or admit them on your terms. Either way, decide on purpose — before an agent shows up acting for one of your customers.

Source: Bloomberg, Investing.com, and CNBC market coverage, September 22, 2026.

5. The auditor moved onsite.

In one week, oversight of frontier AI moved from "tell us what you did" toward "let someone check before you ship." California's governor signed an order directing work on an emergency shutoff for frontier models, with independent verification organizations embedded onsite at the labs and loss-of-control events counted as critical incidents. It cites this summer's lab-agent intrusion by name. Six days later, three senators introduced a bill for a permanent federal AI safety board, with model access 45 days before release and penalties of up to $250,000 per violation per day.

Neither is an operating control yet. The order sets up recommendations and implementation work, and the bill is introduced, not law, with federal preemption still in play.

→ The direction is what matters for your shop. Verifying a vendor's model is becoming someone else's documented job. Verifying how that model behaves inside your estate — on your data, under your authorities — is still yours, and no onsite auditor at a lab does it for you. Plan to the enforcement date, not the announcement.

Source: Office of the Governor of California, Executive Order N-9-26, September 18, 2026; U.S. Senate, AI Risk Management and Security Act of 2026, September 24, 2026.

The pattern.

Every one of these is an attempt to delegate, and every one ran into the same limit. The work speeds up, but the wait stays with the process. The output ships, but the responsibility stays with a person. The agent acts, but the authority stays with whoever granted it. The customer's agent shops, but the admission decision stays with you. And the lab gets audited, but checking what happens in your own systems stays in-house.

That's not bad news. It's the most useful map of the quarter. The parts that won't transfer — process design, named accountability, bounded authority, admission, in-estate verification — are the parts a well-run shop already knows how to own. The organizations that get a return from AI won't be the ones that handed the most to the machine. They'll be the ones that knew exactly what they were keeping.

— Reggie

Orientation is drawn from the Signal Stack — 638 signals across 22 categories, tracking the gap between what AI can do and what organizations are ready to do with it. New signals land daily; the week's five are the ones worth your four minutes.