How to Measure AI Value Outside Engineering

A software team can set its AI use beside a count of merged pull requests, which are the units in which finished code is reviewed and accepted. Most other functions have no count that is as tidy. In a Deloitte survey of 3,500 leaders published on October 29, 2026, 51% said their company tracks token volume, while 28% reported revenue gains from AI. A token is the unit AI vendors bill by, roughly a piece of a word. A token count tells a company how much it used and leaves open what the use produced. Uber’s chief operating officer, asked on a podcast whether the company’s AI spending was paying off, said of the connection between use and results: “That link is not there yet,” as Fortune reported on August 7, 2026.

This article is for the finance or HR leader who has to make that link for support, finance, legal and sales. It pairs each function with a measure and with the published research that used it, then covers the ways the comparison goes wrong.

Why is engineering easier to measure?

Code leaves a record. Each change is reviewed and merged in one system, with a time stamp, and the AI coding tools report use per developer. GitHub’s own documentation describes a dashboard that connects depth of Copilot adoption with pull request output, and it adds a warning that such comparisons “show associations, not proof” that adoption caused a change in productivity.

Even with that record, the engineering evidence does not agree with itself. A field study of 4,867 developers at Microsoft, Accenture and a third large company found that those given an AI coding tool completed 26.08% more tasks. A METR trial of 16 experienced developers working on public software projects found they took 19% longer when AI was allowed, and afterward they still believed it had made them about 20% faster. The 2025 DORA report, a survey of nearly 5,000 technology professionals, found that AI adoption went with higher delivery throughput, meaning more software shipped, and with lower delivery stability, meaning more releases that failed or had to be fixed.

Those results carry two lessons for every other function. The developers in the METR trial misjudged their own speed, so the measure should come from a system and not from a survey of how much time people think they saved. The DORA finding shows that a measure of volume, read alone, can rise while the quality of the work falls.

What is the matching measure in each function?

The table below pairs each function with an output measure and a quality check. The pairings are our suggestions, built from what the cited studies measured. The system named in the last column is the kind of system, and your own may differ.

Two things make an output measure usable here. The team already records it for its own reasons, so nobody has to fill in a new form, and it counts finished work that someone else accepted, such as a resolved ticket or a closed month. It also helps if the measure moves within weeks, so that a reading follows a change in AI use soon enough to act on.

What does the research show for each occupation?

Customer support has the strongest evidence. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered introduction of an AI assistant to 5,179 support agents. Issues resolved per hour rose 14% on average. The gain was 34% for novice and low-skilled agents and minimal for experienced ones, and the abstract reports that customer sentiment improved. A planner should expect the same unevenness inside a single role: the average hides who gained.

In accounting, a Stanford working paper by Jung Ho Choi and Chloe Xie combined field data from 79 small and mid-sized firms with a survey of 277 accountants. For accountants using AI, weekly client support rose 55%, the monthly close came 7.5 days sooner and the detail of the general ledger increased 12%. The authors call this early evidence. Firms chose whether to adopt, so the adopters may have differed from the others before any tool arrived.

In legal work, a randomized trial led by Daniel Schwarcz gave upper-level law students six legal tasks, with a legal research tool that retrieves source documents, with a general AI model that has no such retrieval, or with no AI. Both tools raised productivity in five of the six tasks, with gains of roughly 50% to 130%, and both improved the quality of the work. The abstract says the retrieval tool did this without introducing more hallucinations, which are confident statements that are false. The participants were students working on set exercises, so a law firm should test the size of the gain on its own matters before relying on it.

Sales is the gap. The studies gathered for this article include no controlled trial of sales roles, so a sales leader has less outside evidence to compare against and more reason to measure carefully in house.

How can an output measure be gamed?

When a count becomes a target, people find ways to raise the count. Fortune’s AI editor, Jeremy Kahn, applied Goodhart’s Law to token leaderboards on May 28, 2026: “any measure that becomes a target, ceases to be a good measure.” The same article relays a Financial Times report that some Amazon staff ran AI agents on unnecessary tasks to raise their token numbers. Output counts can be pushed in the same way. A support agent can close tickets that are not solved, and a finance team can close the month sooner by moving corrections into the next one.

Two habits limit this. Read the quality check in the same view as the output measure, so that a rise in one with a fall in the other is visible at once. And do not turn either measure into a rating or a ranking for a person. In a performance review they are questions for the manager to ask, with the role’s task score beside them. Amazon’s own guidance to its engineers, published October 27, 2026, calls input measures such as token use “deeply flawed” and tells leaders to check whether delivery rises where AI use is heaviest.

Does a link between AI use and output prove value?

A link between use and output is the start of the analysis. The teams that use a tool most are often the ones with the most suitable work or the keenest managers, and they might have produced more without it. Microsoft’s Copilot Business Impact report, read on November 11, 2026, lets a customer upload its own outcome figures and compare them with Copilot use. The documentation asks for at least 65 users of the tool and 65 others. It uses deals closed and win rate as examples for a sales team. It also states that the differences shown “have not been evaluated for statistical significance”, which means a gap between the groups could be chance, and that role responsibilities and seasonality may influence the results.

The stronger designs compare like with like. Give the tool to some teams before others doing the same work, as in the support study, and compare the two groups over the same weeks. Record the measure for a period before rollout so that each team is compared with its own past.

It also matters which tasks the tool is used for. In a study of 758 consultants, those using AI did better on tasks the model handled well, and on one task outside its ability they were 19 percentage points less likely to reach the correct answer. A role is a mix of both kinds of task, so heavy use in a role where few tasks suit AI deserves a question before it earns praise. JobRoute publishes how it scores each task in a role for this purpose.

How should a company set this up?

Choose one output measure and one quality check for each role, and write down where each is read from. Report them by role and team. A token count is not a performance measure, and in our view one person’s output over a few weeks depends on too many other things to support a judgment about the tool. Amazon’s memo makes a related point about low use. Where work does not benefit from current AI tools, it says, “that is a finding, not a failing.”

Then read three things together for each role: how much AI the role uses, which of its tasks suit AI, and whether its output measure moved while quality held. JobRoute joins those three for each role, as described on the AI spend by role page.