In my experience, leaders often start the AI conversation with a reasonable question: what will this save us?
As people begin using it, the conversation can change. The work becomes more consistent. People have more room to engage with customers because they spend less attention on the administration surrounding them.
Those improvements can become more compelling than the original time-saving estimate. They also deserve more scrutiny than a few enthusiastic anecdotes.
In Can AI give people their attention back?, I explored how AI can reduce production time while leaving unfinished thinking for someone else. This piece follows that question through: when we get the workflow right, how do we establish whether the improvement is worth the investment?
Consider an illustrative insurance claims team. AI helps staff assemble case histories and prepare correspondence for review. Usage grows, but revenue hasn't moved.
I wouldn't call that a failure without looking at what has changed in the work. Equally, I wouldn't approve another quarter of spending because everyone seems keen.
I would start with consistency and quality.
Are staff missing fewer relevant details? Do customers understand the explanation and what happens next? Does an experienced reviewer spend less time repairing the work?
A claims team's better performance may show up in service quality or delivery cost before it affects revenue. The measure needs to fit the job.
Measure AI's effect on work quality
For this team, I would agree what a good case file and customer explanation look like, then sample them through the existing review process.
The standard should cover factual accuracy and whether the reasoning supports the decision. For customer correspondence, it should also cover whether the explanation is understandable and useful.
I would track the proportion meeting that standard without substantial correction. An improving average can conceal a few cases that still require extensive repair. The reasons for those corrections tell us where to improve the workflow.
Then I would ask the person receiving the work.
A customer might understand an unfavourable decision perfectly well and still dislike it. Asking whether they understood the explanation and their next step tells us something different from asking whether they were satisfied with the outcome.
The same principle applies inside a business. A report is useful when it helps its recipient make a decision. Its author feeling productive is only part of the evidence.
Track employee experience alongside output
Now suppose the claims team produces more consistent work, but staff spend their evenings checking what AI generated. The quality gain has come at a cost we need to see.
I would add a short pulse to an existing team check-in. Ask whether workloads are manageable and whether people have enough room for work requiring their judgement. Invite one example of where AI removed effort or introduced more checking.
Keep the questions stable enough to compare over time. Check how many people respond and compare their answers with evidence from normal reviews.
There is support for assessing work this way. Microsoft's EngThrive framework combines measures of speed, ease and quality, with wellbeing as a guardrail. It was developed for engineering, so I wouldn't lift it unchanged into a claims team. But I would borrow the discipline of measuring the experience alongside the output.
In Gallup's October 2026 research, 47% of US employees who used AI reported more time on interesting work. Some 46% reported higher-quality work. These were employees' assessments, not independently verified business returns. They give us useful questions to investigate.
I would treat Glassdoor ratings as background context. An overall employer rating reflects too much beyond a particular change in working practice to serve as its scorecard.
More fulfilling work is a legitimate benefit. It doesn't need a speculative dollar value attached to make it worth discussing.
Measure without creating another job
I would establish a baseline from comparable recent cases, then sample work as the team uses the revised process. Account for case difficulty. An easier month shouldn't become an AI success story.
A short AI forum can bring the evidence together. Someone shows an improvement. Someone else brings an example that still needed repair. The team agrees what to change and checks whether it helped.
For innovation, I would look beyond the number of ideas submitted. Which suggestions did colleagues find useful enough to test? Which improved the work enough to keep?
AI can help group recurring feedback for review. The person responsible still needs to check that those patterns make sense.
The forum should lead to a decision. Otherwise, we have automated some admin and replaced it with a meeting.
Keep the investment decision honest
If the original business case promised revenue growth and that hasn't happened, I would say so. New evidence can justify revising the case. It shouldn't quietly replace the original promise.
For the claims team, I would weigh demonstrated improvements against the full cost of delivering them, including ongoing human review and maintaining useful context.
Released capacity also needs a destination. It might allow staff to work through complex cases more carefully or provide a better service to customers. If we expect it to generate additional revenue, there needs to be demand and a practical way to serve it.
I wouldn't count the same hour once as a salary saving and again as additional output.
If quality improves consistently and staff experience less strain, that can justify continued investment within an agreed budget and review period. If only usage and enthusiasm rise, I would investigate before expanding.
This is how I would build stakeholder confidence: agree what improvement would look like, then let people inspect the work and its effects. Make room for evidence that the approach needs changing.
The timesheet can tell us where the hours went. It cannot tell us whether they were well spent.