A common way to justify an enterprise AI deployment is a headcount model: minutes saved per task, multiplied by task volume, divided by the working hours in a year, and reported as full-time equivalents. The model measures the hands-on part of a process. A customer experiences the elapsed part, and where most of the elapsed time is spent waiting between steps, the two numbers can move independently.
Take an illustrative onboarding request with three hours of hands-on work spread across four people and eight business days of waiting between them: for access, for the current version of a document, for the one person who knows how the exception was handled last time. Halve every hands-on step and the model books 90 minutes per request. The customer still waits about eight days.
A large field study shows both kinds of gain side by side. In an NBER working paper revised in November 2023, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents at a Fortune 500 software company as access to an AI assistant that suggests replies during chats was rolled out in stages. The sample includes agents who never received the tool during the study; 1,636 agents were observed after gaining access. Their difference-in-differences estimates show chats getting 3.8 minutes shorter, a 9% fall, and chats handled per hour rising about 14% from a baseline of 2.6. In live chat the customer waits inside the task, so here shorter chats are shorter waits as well. The same paper found that customers asked to speak to a manager almost a quarter less often, from a baseline of about 6% of chats, with the largest declines among less skilled or less experienced agents. The authors counted requests, partly because they had no data on actual escalations, and they say the study is not designed to show effects on aggregate employment or wages.
My hypothesis is that fewer requests meant fewer handoffs. If the assistant spread the practices of stronger agents, as the authors’ evidence suggests, more conversations could end without passing the customer up the line, sparing the customer a wait for a second person and a supervisor some time. The study did not measure that. It had no data on actual escalations and reports no measure of handoff waiting or supervisor time, so it cannot show whether any such saving falls outside the 14% estimate. Finding out takes the measurements proposed at the end of this piece.
The headcount framing also changes what employees are willing to show. Ethan Mollick argues that AI gains for individuals do not become gains for the organization without changes to process and incentives, and that many workers hide their AI use, partly because they suspect the company will treat productivity gains as a reason to cut staff. If he is right, a business case built on removable heads gives the people who know where work stalls a reason to say nothing. A case built on elapsed time asks them to point at the stall.
The financial objection is the serious one. Headcount can be booked; shorter elapsed time sounds soft, and budgets go to what can be booked. The answer turns on whether money waits for the process to finish. When subscription revenue starts at go-live, an implementation that finishes three weeks sooner brings three weeks of revenue forward. When a contract is held up by a security review, the close date moves with the review. Where nothing is gated, as with an internal report that has no customer and no deadline, the objection is right, and the honest benefit is the hands-on hours actually freed, stated as hours rather than as fractions of people who will not leave.
The two numbers worth collecting before and after any deployment are elapsed time from request to result and handoffs per case. If a vendor’s business case reports hours saved and neither number moved, the hours were real and the customer waited just as long.