Where AI actually pays off in a mid-size company

A Brazilian occupational-health services provider, 260 people, serving 900 corporate clients. Last year the company bought generative AI licenses for the whole back office, trained the team over two mornings, and waited. Twelve months later the internal survey said 71% of people used the tool regularly, and not one operating indicator had moved.

This is not a failure of execution or of tooling. It is the pattern productivity economics has described for decades, and it has a name.

01The J-curve #

Erik Brynjolfsson, Daniel Rock and Chad Syverson published in 2021, in the American Economic Journal: Macroeconomics, the work known as 'The Productivity J-Curve'. The thesis: general purpose technologies require complementary investment in intangible assets — redesigned process, organized data, people's skills, new ways of coordinating. That investment is expensive and doesn't show up in the accounts as investment; it shows up as cost and as lost time.

The result is the J shape: measured productivity dips or stays flat while the complements are being built, and only rises afterward. Companies that buy the technology and skip the expensive part sit permanently at the bottom of the J, with high usage and no result.

Brynjolfsson and Andrew McAfee had already laid the groundwork in The Second Machine Age (2014), arguing that the economic effect of digital technologies depends on recombination — pairing the technology with organizational change — not on the machine's capability in isolation. The difference between 2014 and now is the speed of the technology, not the nature of the problem.

02Where there is evidence of gain #

There are controlled studies, and it is only honest to use what they actually show. Shakked Noy and Whitney Zhang published an experiment in Science in 2023 on half-hour professional writing tasks: participants with access to the tool finished faster and with quality rated higher, and the gain was larger for those who started out performing worse.

Brynjolfsson, Danielle Li and Lindsey Raymond studied customer support in a real setting, in work circulated as an NBER working paper in 2023 and published in the Quarterly Journal of Economics in 2025. The pattern repeated: gains concentrated among less experienced agents, small effects among the most experienced, and improved satisfaction for the customers being served.

Two limits have to be stated alongside that. First: these are specific tasks, not whole companies — extrapolating from task to organizational result is exactly the leap the J-curve says doesn't happen on its own. Second: the 'bigger gains for those who know less' pattern has a direct and uncomfortable management implication — the return tends to show up as leveling of performance, not as multiplying the performance of people who are already good.

03What pays off at 50 to 1,000 people #

1. Repetitive coordination work. Chasing open items, routing requests, consolidating status. High volume, stable rule, reversible error. It is the safest territory and the least glamorous.
2. Intake triage and classification. Requests, résumés, invoices, incident reports. The gain comes from cutting the time until the thing reaches the right desk, not from being right 100% of the time.
3. Structured, recurring writing. Standard replies, minutes, meeting summaries, first drafts of proposals. This is where the experimental evidence is strongest — and the effect is largest on the less experienced team.
4. Reading long documents. Contracts, tenders, technical reports. Extracting fields and flagging discrepancies, with human review wherever the consequence is irreversible.
5. Onboarding new people. Answering 'how do we do it here' from what the company has already written down. Entirely dependent on the company having written it down.

Want to see how this looks inside a real operation? Explore the platform.

And what usually doesn't improve: prioritizing between workstreams, negotiation, evaluating people, diagnosing root cause in an operation with no data, and anything whose bottleneck is political. If two departments disagree about who owns a process, no agent settles that — the disagreement will simply get expressed faster.

04The math, with assumptions stated #

The scale of that occupational-health provider — a mid-size Brazilian company (figures in Brazilian reais) — looking only at triage of scheduling requests. Assumptions explicit; swap in your own numbers before using this in any decision:

On those assumptions, the first wave addresses roughly 243 hours a month. At an internal cost of R$ 62 per hour, that is approximately R$ 15 thousand a month in freed capacity — against an investment that, in year one, includes licenses, the work of mapping the process and two people's time on rollout. The point of the exercise isn't the result: it is that the exercise is only possible because somebody measured those 4.5 minutes per request first.

A company without that number cannot calculate AI return for a simple reason: it doesn't know the denominator. It is the same trap of measuring only what is easy to count, described in when the measure becomes a target, the metric dies.

05How to climb out of the J #

At Relevanti, measuring where the time goes and putting the agent inside the process are the same thing, because they happen in the same place — see the platform, the department view under solutions, or bring your own numbers to a conversation.

The J-curve is not an argument against investing in AI. It is an argument against expecting a result without paying for the expensive part — and a warning that, with no baseline, you won't even know where on the curve you are.