The AI paradox at Agibank: when execution becomes a commodity and decision-making becomes the expensive asset

This piece is branded content from Agibank, a Brazilian digital bank, on Neofeed, a Brazilian business news outlet, so it's worth starting with the obvious caveat: the numbers come from the bank itself, not from an independent audit, and I treat them as what they are, claims from a party with a stake in the narrative. That said, the thesis behind the piece deserves to be taken seriously, because it isn't empty marketing: it's the cleanest formulation I've read of a movement that's already on the radar of anyone hiring developers in Brazil.

Marciano Testa, the bank's founder, sums it up like this:

The more machines execute, the scarcer and more expensive what only humans do becomes. It's people's talent, judgment, and execution capacity that turn artificial intelligence into real value.

>

-- Marciano Testa, founder and controlling shareholder of Agibank

It's a nice line, and like every nice line delivered on stage, it hides an asymmetry. It's true for whoever is already on the judgment side. For whoever is on the pure execution side, it's a sentence. My job here is to separate the two.

The numbers the bank puts on the table

Before the thesis, the data. Agibank says it operates with 7.6 million active customers at the end of Q2 2026 and has brought AI to 100% of its areas. The gains cited, all from the company's own sources:

| Area | Claimed metric | |---|---| | Customer service | ~60% fewer positions for the same volume; resolution time cut to less than 1/6 | | Complaints to the BC (Brazil's central bank) | 75% fewer upheld complaints | | Development | nearly 80% faster; some stages in less than half the time | | Marketing | 98% lower cost vs. outsourced agency | | Document validation | accuracy from ~80% to +99%; cost 98% lower | | Security | incident investigation from 3.5 days to ~8 minutes |

The number that matters most directly to this piece's reader is the development one: nearly 80% faster, with "AI solutions that support code production and software preparation stages." Translating from press-release speak to the shop floor: copilots generating code and pipeline automation. Nothing a team that uses Copilot, Cursor, or Claude Code day to day wouldn't recognize. What changes is the declared scale of adoption, not the technology.

Where the thesis gets it right

There's a genuinely interesting technical piece hidden in the middle of the marketing: Complexity Points. It's a model trained on the bank's own history of demands and requirements to estimate the complexity of each new initiative and help prioritize and size effort.

This is more relevant than generating code, and I'll explain why. Generating boilerplate faster is a marginal gain, everyone already has that. Estimating complexity from your own real history attacks the problem that was always human and always bad: guessed planning poker, blown deadlines, badly sized scope. If it works as described (and that's a big if), the value isn't in writing the line, it's in deciding what to write and how much it will cost. It's exactly the "judgment" from Testa's thesis, just instrumented.

The same logic repeats in security and anti-money-laundering: AI handles collection, organization, and volume triage, and the human is left with the cases that require a decision. Here the thesis holds up. When the machine eats the mechanical part of the SOC, the analyst who's left isn't the one who used to open tickets, it's the one who decides whether that's an attack or a false positive. The work doesn't disappear, it moves up a layer.

The counterpoint the press release doesn't make

Now the problem. The piece sells the scarcity of human labor as universally good news, and it isn't.

A 60% cut in customer service positions isn't "people moved up to higher-value tasks." It's 60% fewer people on that function's payroll, with a minority repositioned and the majority out. Calling that a valorization of the human is choosing the sunny half of the sentence. The same dynamic looms over the development team: if you speed up development by 80%, you either deliver 80% more product with the same team, or you deliver the same with fewer people. The two aren't equally likely at a bank that measures success by "economic return."

And there's the point every developer should internalize: the boundary between executing and deciding isn't fixed, it rises along with the machine. What was a "decision" yesterday (estimating complexity, writing an integration test, reviewing a simple PR) becomes automatable execution tomorrow. Complexity Points is proof of this within the release itself: a senior-level judgment task being pushed inside a model. Anyone betting that "just becoming an architect" is enough to stay safe is aiming at a target that's also moving.

What changes for those who build in Brazil

The practical takeaway isn't "learn to prompt," that's already a commodity. It's more uncomfortable and more lasting:

• Value migrated from the how to the what and the why. Knowing how to write the function matters less and less than knowing which function solves the right problem, how much it costs to maintain, and why it exists. Business context, which developers historically outsourced to the PM, has become part of the job's defense.
• Judgment over AI output is the new scarce skill. Reviewing code the machine spits out, knowing when the model hallucinated a dependency, when the generated test checks the wrong thing. The bottleneck stopped being production and became validation at scale. That's where humans still win, for now.
• Mastery of unstructured data as a signal. Agibank's credit case (turning text, audio, and images into risk signals via LLM, improving the KS statistic) shows where engineering adds value: not in the off-the-shelf model, but in the pipeline that turns raw data into a reliable feature. That's applied data engineering, and it's hard to automate because it depends on understanding the business.

The honest reading of Agibank's piece isn't the one the bank wants to sell. It isn't "AI frees humans for what matters." It's: AI is redrawing the line that separates who's expensive from who's fungible, and that line rises every quarter. Whoever only executes will be squeezed, and no talk of valuing talent changes that math. Whoever positions themselves where there's decision-making, context, and accountability for outcomes gains ground, knowing they'll need to move up again next cycle. The paradox is real. It just isn't as comfortable as the stage suggests.