Artificial intelligence is already in use at 92% of Brazilian capital-market institutions surveyed by ANBIMA. Only 3% of the AI users say the technology creates a clear competitive advantage. Another 3% have corporate measurements tied to strategy.
Those figures describe more than an adoption gap. They reveal an instrumentation gap. Institutions have put AI into work, but few can connect that work to a repeatable business result. Adoption has become the easy number.
ANBIMA and Datafolha received responses from 177 institutions in the association’s ecosystem between January 21 and April 6, 2026. The questionnaire was self-administered online. The reported confidence level was 95%, with a maximum margin of error of 7 percentage points for the full sample. C-level executives, directors and superintendents made up 72% of respondents, which ANBIMA summarizes as seven in ten. These are institutional self-assessments, not audited operating data, according to the report’s methodology.
Adoption has become a weak signal
The headline rate comes from 162 of 177 respondents saying they use at least one form of AI, which rounds to 92%. ANBIMA then uses those 162 AI users as the denominator for its maturity questions. Among that subgroup, 48% report applications in key areas with process improvements. The share falls to 11% for integration into critical processes and 3% for a clear competitive advantage, the full report shows.
The same denominator applies to governance and measurement. Only 29% describe data governance as structured or advanced, combining the reported 24% and 5% categories. Just 12% place responsible-use policy in the consolidated or advanced categories, the sum of 9% and 3%. On results, 43% rely on informal perception without metrics or records. Only 3% report standardized corporate measurement linked to strategic goals and senior-management reporting, according to ANBIMA’s findings.
The survey cannot establish that weak measurement caused weak strategic results. It captures institutional posture at one point in time, and estimates for the user subgroup carry more sampling uncertainty than the full-sample headline. The pattern is still internally consistent: tactical use outruns critical integration, while informal perception outruns strategy-linked reporting. That sequence is what an institution should expect when deployment precedes instrumentation.
This does not prove that 97% of AI programs have failed. Competitive advantage is a relative outcome, and widely available models can improve many firms without creating a durable edge for any one of them. The survey also asks a separate multiple-response question in which 14% cite competitive advantage as a positive impact. The stricter maturity ladder finds only 3% claiming a clear advantage. The wording matters.
The measurement result is harder to explain away. An institution that records productivity informally may have a useful tool. It does not yet have evidence that the tool improved a unit of work after model fees, review time, errors and rework. It also cannot compare that deployment cleanly with the next one competing for budget.
The scarce capability is no longer access to a model. It is the ability to produce evidence about a workflow.
BSM shows the shape of a measurable case
BSM, the self-regulatory organization for markets operated by B3, offers a narrower example. B3 says BSM runs 41 alert and ranking processes, with 11 using machine learning or AI. It also reports 60 algorithms for process validation and automation, plus 63 standardized queries for generating insights. Those figures describe a defined supervisory system, not the private institutions in ANBIMA’s sample, according to B3’s account.
B3 reports that a supervised learning model cut false positives by 50% in transaction monitoring. It also says a conversational data room reduced one type of human analysis from four hours to one hour. The release does not disclose the observation period, alert volume, staffing, comparison design, sample size or changes in missed detections. The linked official BSM paper discusses the architecture but does not supply those validation details. The results should therefore be treated as reported outcomes, not independently validated estimates.
The useful distinction is the unit of analysis. A false-positive rate belongs to a particular alert process. A four-hour review belongs to a defined analyst task. Both can be measured against a baseline. “Uses AI” belongs to an institution and says almost nothing about operating performance.
BSM’s case also shows why one speed metric is insufficient. Cutting false positives can reduce review load, but supervisors also need recall, escalation quality and detection stability across market regimes. Reducing analysis time matters only if the resulting judgment remains accurate and auditable. A faster wrong answer is still wrong, only more punctual.
Measurement has to start before deployment
For capital-market operators, the next maturity step is not another inventory of tools. It is a measurement contract for each workflow. The contract should define the unit of work, pre-deployment baseline, model version, human review requirement, cost, latency, error or exception rate, override rate and downstream outcome. Governance then becomes part of measurement rather than a policy file beside it.
This changes procurement. Buyers can require vendors and internal teams to demonstrate improvement on the institution’s own workload, not a benchmark detached from it. Compliance leaders can see which model produced a decision, which reviewer changed it and whether performance drifted. Executives can compare two projects without translating enthusiasm into a fictional common currency.
ANBIMA’s 92% confirms that access is no longer the constraint. Its 3% strategy-linked measurement rate identifies the harder bottleneck. Until institutions instrument the work around the model, reported adoption will keep rising faster than demonstrated advantage.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.