The AI productivity paradox: what the speed dashboard leaves out
AI productivity metrics measure throughput, so they keep rising even when decision quality on hard cases is sliding. Here is what to track instead.
Author
Head of Product & Delivery
Topics
Key takeaways
- AI productivity metrics measure throughput, so they keep rising whether decision quality is holding or sliding
- In one claims team, assessor agreement on complex cases fell from around 80% to the mid-60s while every speed metric improved
- MIT and Stanford research found a 14% average productivity lift driven by convergence on top-performer behaviour, which trades some diversity of approach for speed
- Convergence is efficient on the average case and often brittle at the edges, and the edges are where regulated industries make or lose money
- Track AI dependence instead of AI usage: rate a monthly sample of edge-case decisions blind and watch the gap against your throughput dashboard
Almost every AI rollout I’ve seen over the past year gets measured the same way. Prompts per day, average handle time, tickets closed, all of it pointing up and to the right. Those numbers are usually accurate, the work is faster, so the rollout gets written up as a win and everyone moves on.
What none of it tells you is what’s happening to the quality of the hard decisions, meaning the non-standard cases where reasonable people could disagree, and those tend to be where regulated and high-trust businesses make or lose their money. In more than one team I’ve worked with, that quality has been sliding at the same time as every number on the dashboard kept improving.
Here’s what put me onto it.
The number that didn't fit the dashboard
A claims team we worked with rolled out an AI tool, and within two quarters, agreement between assessors on complex, non-standard cases had fallen from around 80% to somewhere in the mid-60s. Their speed metrics over the same period looked excellent, which is part of why nobody went looking for a while.
The tool itself was giving perfectly sensible suggestions, so that wasn’t the cause. What had happened is that the team split into two camps without anyone deciding to. Some assessors started anchoring on the AI’s first suggestion by default, and others overrode it on instinct nearly every time. That inconsistency had probably been sitting there all along, and the tool reduced the visible friction without touching any of it, which pushed the problem somewhere the speed-to-decision chart couldn’t reach.
Agreement between assessors on complex cases fell from around 80% to the mid-60s, while every speed metric on the dashboard kept improving.
Why that was fairly predictable
There’s a 2023 study by economists from MIT and Stanford that tracked generative AI across more than 5,000 customer support agents, and the headline finding got celebrated everywhere. Productivity rose 14% on average, with the biggest gains going to novice and lower-skilled workers, who improved by roughly a third. The researchers described the tool as a knowledge equaliser, spreading the working habits of top performers to everyone else.
That’s a good outcome and worth sitting with, because of what else it implies. If part of the tool’s job is getting everyone to converge on the same best answer, you’re trading some diversity of approach for speed. Nobody has directly measured that trade-off yet, including this study, but the mechanism it found makes the risk obvious enough to name. A team that starts solving problems the same way is efficient on average and often brittle at the edges, and complex, unpredictable industries tend to make their money, or lose it, at the edges rather than the average.
Which is roughly what the claims team ran into. Convergence for some, over-correction from others, and a quality signal that nobody was watching because every other number was moving the right way.
Why AI productivity metrics keep missing this
The measurement problem is structural. Throughput is cheap to instrument, because the tool logs every prompt and every ticket, so those numbers arrive by default. Decision quality on ambiguous cases has to be assessed by a person, usually weeks later, and often against an outcome that hasn’t landed yet. One set of numbers shows up in the dashboard on day one and the other has to be deliberately built.
So the AI productivity metrics that get reported to a steering committee are the ones that were easiest to collect, and they’re all measuring the same underlying thing. Faster. When a program is reported on entirely through speed, the incentive to go looking for a countervailing signal quietly disappears, because every conversation about the rollout is already a positive one. This is the same pattern sitting underneath AI projects that struggle to scale past the pilot, where the early numbers look excellent and the operating reality underneath them was never measured.
It also makes the board conversation harder than it needs to be. If the only evidence you have is throughput, you can’t answer a question about risk, and talking to a board about AI without overpromising depends on having a second number to put next to the first one.
Speed hides the erosion of judgement
Give a skilled underwriter, engineer or claims assessor an AI copilot and the immediate gain is obvious enough. Decisions get faster and backlogs shrink. The more interesting question is what happens to the decisions that were never straightforward in the first place.
There’s a useful term from aviation human factors research called automation surprise. It describes what happens when an automated system behaves in a way the operator didn’t expect, at the point where the operator’s situational awareness has already eroded too far to recover quickly.
Product teams are flirting with a version of this. When people can prompt an answer in seconds, they stop wrestling with ambiguity, and the skill of sitting with an unstructured problem starts to atrophy quietly, in much the same way a pilot’s manual flying ability fades on a heavily automated flight deck. Invisibly, until the one moment it’s needed. You won’t see any of that on a dashboard until a novel crisis lands and nobody in the room knows how to think without a prompt.
None of this is an argument against the tool. The same research that raises the convergence risk also shows upside in faster onboarding, narrower skill gaps and better retention. The question worth asking is whether you’re building a workforce that’s stronger because of the tool, or one that’s only capable with it, and most organisations can’t currently tell the difference because they aren’t measuring for it.
You won’t see judgement erode on any dashboard until a novel crisis lands and nobody in the room knows how to think without a prompt.
What to observe instead
The product leaders getting this right have shifted from tracking AI usage to tracking AI dependence.
One way to do it in practice. Each month, pull a sample of 15 to 20 decisions from the edge cases, meaning the ones outside the AI’s confidence band where reasonable people could disagree, then:
- have a senior reviewer rate outcome quality blind, without knowing whether the person followed the AI’s suggestion or overrode it
- track that score over time, kept separate from your throughput metrics
- watch the gap between the two
If quality drifts down while the speed dashboard stays flat or keeps improving, you’ve found the thing your current metrics were hiding.
The sample is deliberately small, because you’re after a number that moves for a different reason than throughput does, reviewed by someone senior enough that the rating carries weight. Teams that already run this kind of cadence, usually the ones with mature delivery and product practices, can bolt it onto an existing quality forum rather than standing something new up.
Pair that with one uncomfortable question every quarter. If we turned the AI off tomorrow, would our people still be able to do the job?
That’s a resilience question, and in regulated, safety-critical or high-trust industries, it might be the only metric that matters.
Where this leaves product leaders
Nothing here argues for slowing a rollout down. The upside in these tools is well evidenced and the organisations capturing it are pulling ahead. The argument is for a second number, because a program measured on one axis will keep reporting success right up until something breaks that the axis was never watching.
Most of the organisations that struggle with this aren’t short on ambition. They’ve moved from experiment to rollout and stopped short of the operating discipline that makes it hold, which is the same shape as the AI implementation gap. Measurement is usually where that gap becomes visible first.
We spend a lot of time with product and engineering leaders working through exactly this, usually somewhere between the rollout that looked like a win and the first hard case that tests it. If that’s a conversation worth having, let’s talk.
Frequently asked questions
What are AI productivity metrics and why are they misleading?
AI productivity metrics are throughput measures such as prompts per day, average handle time, tickets closed and time-to-decision. They’re misleading on their own because they only measure speed. A team can process work faster while the quality of its hardest, most ambiguous decisions declines, and no throughput measure will surface that. The metric rises either way.
How do you measure AI dependence rather than AI usage?
Sample 15 to 20 edge-case decisions each month, meaning the ones outside the AI’s confidence band where reasonable people could disagree. Have a senior reviewer rate outcome quality blind, without knowing whether the person followed the AI suggestion or overrode it. Track that score separately from throughput and watch the gap between the two lines. If quality drifts down while speed holds or improves, you’ve found what the dashboard was hiding.
What did the MIT and Stanford study on generative AI find?
The 2023 study tracked generative AI across more than 5,000 customer support agents and found productivity rose 14% on average, with the largest gains going to novice and lower-skilled workers, who improved by roughly a third. The researchers described the tool as a knowledge equaliser because it spread the working habits of top performers across the team. The study did not measure whether that convergence reduced diversity of approach on non-standard cases.
What is automation surprise and how does it apply to AI tools?
Automation surprise is a term from aviation human factors research. It describes what happens when an automated system behaves in a way the operator didn’t expect, at a point where the operator’s situational awareness has already eroded too far to recover quickly. Applied to AI copilots, it describes teams who stop wrestling with ambiguity because an answer is always a prompt away, and then meet a novel case with the judgement muscle weakened.
Should regulated industries slow down AI adoption because of this?
No. The evidence on upside is strong, including faster onboarding, narrower skill gaps and better retention. The recommendation is to add a decision-quality measure alongside the speed dashboard, so a rollout in a regulated, safety-critical or high-trust environment can show it’s building capability rather than substituting for it.
Related articles
Real outcomes, for real Australian organisations.
How we built AI-powered personalisation and retail media for one of Australia’s largest supermarket chains.