Fintech with AI: more 'tech', less 'fin' | Akua + Común

The audio and video are in Spanish. These notes are a translation.

Today I brought together two CTOs who are rebuilding fintech from scratch with artificial intelligence.

Abiel Gutiérrez was a founding engineer at Brex and is now cofounder and CTO of Común, the neobank for Latino immigrants in the United States. They do tens of millions in revenue with 35 people.

Juan José comes from dLocal and Pomelo, and is now cofounder and CTO of Akua, a payments infrastructure that went to production in 9 months. His estimate: what they built would have taken five years and between 200 and 300 people.

The conversation starts from one observation: a lot of fintechs ended up being more fin than tech, with huge teams. These two are doing the opposite, with fewer people, less SaaS, and the CTO back at the center.

Structure and Key Highlights

How Many Fewer People Their Previous Companies Would Have

• Abiel’s caveat: it depends on how much you can buy from vendors. Brex built everything in house, and Ramp, its competitor with more scale, is known for leaning on vendors, Stripe included
• At Común they build more because they serve a user with no access to the financial industry, and there is less to take off the shelf
• His estimate: in early stage the difference is enormous, on the order of 10x, because you are in prototypes and validating the market hypothesis
• His nuance: when the codebase grows and there are lots of dependencies, that multiplier comes down
• Juan José lived both cultures: dLocal looked for partners for everything, for speed, and Pomelo built in house so it would scale
• At Akua he says he is living the intersection, and his data point: industry experts told them it was impossible to go to production in 9 months
• They shipped in 9 months and did more than they had promised
• Today they are about 70 people, of which 40 are technology

Where the Leverage Shows Up Most

• At Común: operations and support, by far
• The explicit decision at Brex that Abiel remembers: engineering would never build internal tools for support, so the team used low code
• The consequence: large operations teams or outsourcing. Común got up to 50 human agents in Mexico
• The other front in fintech: compliance, with a lot of manual review work that now gets automated with confidence
• At Akua the answer is different: every team, because from day zero they automated operations
• The advantage of their model: they are the client and the provider of their own product
• Their daily metric: ship one feature a day, and they hit it
• What surprises Juan José the most: the team that builds the most tools internally is marketing
• The context that makes it hard: they are PCI compliance, ISO compliance, SOC 2, with an acquiring license in several markets

What to Measure

• Juan José measures output, not lines of code or tokens: features shipped, implementation time
• His goal: bring customer implementations down from three to six months to one or two weeks
• His record: a customer integrated in 6 hours from the contract signature to production
• Abiel tracks lines per engineer and PRs at the leadership level, without exposing it as a leaderboard
• Why not a leaderboard: he has heard of companies where people burn tokens reviewing Slack just to climb the table
• The metric he does find useful: the average time from when a PR is opened until the merge, which should compress by using AI for the first pass
• His warning: those metrics spike when you start a codebase from zero, so in the end you go back to the business metrics
• Akua’s odd metric: the number of Slack messages, which goes up with productivity because there are more threads and more coordination
• His way of sizing it: each person leads at least 10 agents, and an engineer who used to carry one thread now carries ten

How Their Own Role Changed

• Abiel no longer codes for the customer day to day, except in intense stretches with deadlines
• What he does do with AI: internal tools and personal things
• Juan José is in meetings all day, but with five Claude Code agents running in parallel
• His rule: he does not step into anything on the critical path, because he cannot follow it through
• What he does do: prototypes he brings to the team, already brought down to something concrete
• Why it matters that the founder is technical: you know how much something should cost, and you can say that does not take a week, it takes three hours

The Real Obstacle to Adoption

• Abiel: non-technical people do not know how simple it is, and if you do not know how to do it the model itself tells you how
• There is no barrier to entry, but you only find that out by trying it once
• What they did: a week in which nobody did anything except use AI, across the whole company
• The recommendation from both of them: first give open access and maximize adoption, then optimize

Security and Permissions

• Akua put the effort into platformizing: standardizing flows, tokenizing information by default, putting the safety nets in place before opening the door
• They are part of a global consortium on Amazon’s agent platform, which already comes with good security practices
• Común built its own MCP connected to all of its internal tools
• Their philosophy: read access to everything for the whole company, except PII, which by regulation is not even in the system or is encrypted
• With write access they were more cautious, starting with the engineers
• The most complex problem, according to Juan José: the 25 customer-facing agents, where you have to guarantee that each one sees only its own data, with no prompt injection and with well-delimited scopes
• The result they report: 70% operational efficiency gained for their customers
• The anecdote: they built an internal agent so powerful, with access to all the APIs of the acquiring infrastructure, that it scared them and they decided to leave it capped

Build or Buy

• Común’s prior mindset: buy first, because when you launch a new product you do not know how it works or how your customer is going to interact with it
• The way to see it: you are not only buying the solution, you are buying the learning
• Their concrete limit: fraud products in the United States do not cover foreign identities
• What they did when they moved into merchant acquirers: ignore the fraud vendor stack people recommended to them, use data sources only, and build their own models
• The cost: one or two months of pain, and today they work spectacularly
• Their acknowledgment: they would not have made that decision three years ago
• Akua’s inverse process: the vendors they hired for speed started to slow them down, because many of them have been in the market for 10 or 20 years
• Their result: between eight and ten SaaS products eliminated in a very short period
• What they are not going to build: AWS, their own models, a CRM, a Slack, or data providers
• Their conclusion: if you are more tech than fin, today it is almost impossible for a vendor to give you more speed than building it yourself

Why Removing Vendors Improves the Product

• The case Enzo brings in: customer support depends on data from the whole company, not only from the support area, so a third party will never give the best experience
• Akua’s position: they still use the ticketing tool, but they chose it because it is open and can be connected via MCP
• The goal: that the human resource comes last, and that the ticket almost never arrives
• Abiel’s reading of the full arc: at the beginning everything is speed, and the rest of the journey is taking away, taking away, and taking away in order to gain control
• The consequence for SaaS companies: the ones that survive will be the modular, flexible ones, more infrastructure than packaged product
• The contrast they describe between vendors: the ones that publish their own MCP so you can work with their data, and the ones that ship an AI feature inside their site and charge you separately

What Happens to the Business Model

• Juan José expects costs to come down in a competitive market, but for the value-added layer to get charged by outcome
• The pre-AI example Abiel brings in: Robinhood eliminating the commission and monetizing order flow
• The version they imagine for Común: if margins grow, make remittances free at the official exchange rate, and give the margin back to the customer so it turns into virality
• The counterweight: today the cost of the models is subsidized, and it has already started to rise
• Akua’s data point: AI is 50% of their technology cost today
• The precedent they cite: Uber spent the annual budget in the first four months and could not trace the impact on its business metrics
• The obvious optimization Abiel points out: if your team writes emails with the most expensive model by default, an open source one at 5% of the cost gives the same quality

The Organizational Changes

• At Común an operations team no longer exists: those responsibilities moved over to product
• The reasoning: as the models get better, the profile that gets the most out of them looks more like a product manager than someone from operations
• What changes in support: from reactive to proactive, because the agent spots the problem and tells the user before they write in
• At Akua they never created a product area: it is called product engineering, and they hire engineers with a business and product mindset
• Their reason: putting a product person in the middle adds a layer that does not help an engineer who already thinks in systems
• The resulting paradox: their operations team ended up looking like a product team
• The ritual that holds all of it together: workshops twice a week showing what each team is doing
• The zero-ego lesson: they built their own agent platform, the next day Amazon launched its own, and they decided to destroy theirs and migrate
• The other case: they assembled a full spec driven development kit in November and threw it out in January when the coding agents arrived
• What did survive: the connectors. They launched their payments MCP before Stripe

The Shared Brain

• Abiel’s goal: that the company have a single brain, so everyone decides in the same direction without meetings to communicate
• How they do it: they record every group conversation, pass the transcripts to agents, and publish summaries on Slack according to the scope
• And an interface where you can ask for the status of the roadmap, the ETA of a feature, or its impact
• The reference he mentions: the gbrain
• Juan José’s concept: platformize, standardize so you can apply controls and safety nets a single time
• The result: 25 customer-facing agents and more than 70 internal ones, accessible from Slack or from the desktop with each person’s permissions
• Their agent that knows everything is called Homero

Autonomous Loops

• In engineering they are close, but they still do not let the agents do the merge and the deployment
• The idea that occurs to them: feature flags deploying to 1% or 5% of the user base
• Why the models are so good at programming: the languages are standardized and there is an endless supply of open source code to train on
• Why a product manager agent is a different thing: there is no data of someone doing exactly what they do
• Akua tried it and took it out: it worked technically, but nobody was responsible and nobody could walk into a meeting and explain it to the customer
• Where they do use it: an agent watching the metrics detects a fraud spike, proposes a rule to the prevention agent, runs it in shadow mode measuring without applying it, and only then asks to move it to production
• The measurable result: their chargeback win rate went from 12%, which would already be very good, to 55%
• The pending technical challenge: running agents for five, six, or eight hours, which until recently the industry had not solved

The Renaissance of the Engineer

• Juan José’s thesis: the engineer’s role went back to its foundations, which is creating systems
• What he considers dead: the developer understood as the person who writes the code
• His stated ambition: a company with few people, very senior, the best in the market

What They Imagine for Next Year

• What excites Abiel is automating the management system: how the success of each function gets measured, which metrics matter, what the inputs are
• What he discovered along the way: giving instructions to an agent sharpens his thinking about what he wants from his team
• The experiment he mentions: preparing a one-on-one with the head of product as if he were giving instructions to an agent
• His more uncomfortable conclusion: you can build high-impact businesses with quite a lot fewer people, because the most complex part of growing is the humans
• The anecdote that closes it: an agent told him it was tired and to see it tomorrow
• Juan José no longer plans two years out, but two, three, or four months out

Silicon Valley Glossary

People, Companies and References

People Mentioned

• Abiel Gutiérrez: founding engineer at Brex, cofounder and CTO of Común
• Juan José (Juanjo): from dLocal and Pomelo, cofounder and CTO of Akua
• Rodrigo Figueras: from Story, cited for the conversation about eliminating SaaS
• Enzo Cavalie: host of the podcast

Companies and Organizations

• Común: neobank for Latino immigrants in the United States, tens of millions with 35 people
• Akua: payments infrastructure, about 70 people, in production in 9 months
• Brex and Ramp: the contrast between building everything in house and leaning on vendors
• dLocal and Pomelo: the two opposite cultures Juan José lived through
• Amazon: whose agent platform they adopted, inside a global consortium
• Klarna: cited for the case of eliminating third-party SaaS
• Sierra and Decagon: the AI support vendors they discuss
• Robinhood: the precedent for reinventing the business model
• Nubank: mentioned for the maturity of the Brazilian fintech ecosystem
• Dio and the newsletter: sponsorship and promotion of the episode