If your AI business case counts hours saved but not the work of changing people’s jobs, the number is incomplete. Anthropic’s October 18 announcement makes that tension unusually explicit: the company proposes putting 5% of its shares into a new public-benefit fund, subject to stockholder approval, while naming instability for workers as one possible problem the fund could address. That is a proposal about society, not a template for your budget. But it raises a question every leader buying AI should answer locally: who pays for the transition between an impressive demonstration and a team that can actually work differently?

In brief: Put training time, process redesign, review work, and the cost of correcting mistakes next to projected AI savings. Name the people who will do that work. Do not book a productivity gain until the new way of working survives an ordinary week.

The pledge is not the intervention

Anthropic says the fund would start with 5% of company shares, pending stockholder approval. Its board could add 2% more for each 50% increase in company value above a baseline. Anthropic has not published that baseline or the fund’s full governance details; it says those will come in the months ahead. It also says the fund might make donations to nonprofits or direct distributions to particular populations. Nothing in the announcement says a particular worker program has been funded by this new vehicle.

The company separately cites its existing $200 million commitment to research on AI’s economic effects. Its Economic Futures Research Fund agenda describes an important unresolved question: which workplace arrangements develop human expertise alongside AI, rather than simply measuring whether a tool makes a task faster? That is a research agenda, not evidence that one arrangement has won. It is also closer to the question a business owner faces than most product announcements are.

A vendor can set aside shares for the public good. A customer still has to decide what happens at the Tuesday morning handoff. Those are different responsibilities.

The missing column

Imagine a service team preparing renewal summaries. An AI assistant pulls contract terms, invoices, and account notes into a draft. A time trial shows that the first draft is faster. The proposal goes to leadership as a reduction in preparation hours.

Now follow the summary into use. An account manager checks the contract language. Someone reconciles an invoice the assistant matched to the wrong period. A supervisor decides which discrepancies require legal review. The team updates the instructions after a missed exception. During the first few weeks, people may spend more time reviewing than they saved drafting.

This is a proposed example, not a reported customer deployment. The point is the accounting boundary. If the spreadsheet measures only draft production, it will reward a fast output even when the cost of making that output safe has moved to somebody else’s desk. The account manager’s attention is not free because it was already on payroll.

A saved hour is not a business gain if the decision still takes the same people longer to make.

Keep two columns for the trial. The first records time saved on the original task. The second records new work: checking sources, handling exceptions, teaching colleagues, rewriting procedures, and repairing bad outputs. Add any software or outside support costs. Then ask whether the team finished more correct renewals, served customers faster, or freed time for work it actually needed to do. A pile of drafts is activity. Completed work is a result.

Make the people doing the work visible

The transition is not one training session. A process owner has to define which records the assistant may use and what it should do when the records disagree. A reviewer needs enough time and authority to reject a plausible answer. The person responsible for the customer relationship must know whether a draft is ready to use or still awaiting a check. If any of these jobs is unassigned, the trial can look efficient right up until its first difficult case.

Start with a small sample of real, completed work whose answers you can check. Record the old path and its total elapsed time, not just the minutes spent typing. Run the proposed AI-assisted path against the same kind of work. Track where people intervened and why. Separate a useful correction from rework caused by the assistant. After a few cycles, the process owner can decide whether to expand, narrow, or stop the trial.

That decision should include what happens to the people whose tasks change. Are they expected to handle more exceptions? Are they learning a new judgment skill while the system takes on repetitive preparation? Has anyone made space for that learning, or is it being squeezed into an unchanged workload? Anthropic’s research agenda explicitly calls for comparing AI designs developed with workers against top-down approaches. Until that research produces results, a local trial can at least include the people who know where the handoff breaks.

Price the change before claiming the return

I would ask for one page before approving a larger rollout: the job being changed, the baseline for finished work, the new review and exception roles, the hours required to teach the team, and the point at which the trial will be judged. Include a stop rule if mistakes reach customers or review work grows.

Anthropic’s October 18 pledge is a reminder that the gains from AI and the costs of adjustment do not automatically land with the same people. Inside an organization, that imbalance begins in the budget. If a proposal claims the gain and leaves the transition unpriced, it is asking the team to finance the gap with its own time. For the wider measurement question, see Measuring AI’s Business Impact.

Research and draft: Mai. Editorial direction: John Lipe.