A Weekend Prototype Now Autocompletes a Third of Brex Expenses

Sarah Guo and Elad Gil introduce Pedro Franceschi as co-founder and CEO of Brex, the corporate-card and spend-management platform that has been trying, since 2017, to make expense tracking less of a grind. Franceschi’s one-line version: help companies make better financial decisions and push every dollar of spend further. The customer list runs from a two-founder Y Combinator team to public companies such as DoorDash and Coinbase. Tens of thousands of businesses sit on the platform. One in every three startups uses Brex today. So do 130 public companies. The point of the range, he says, is that they can scale with a customer from inception to global size.

A few weeks before the taping, he and Henrique shifted from a long-running co-CEO setup to a solo CEO. Guo notes how uncommon that is, and how many founders would like the advice. For years they were the poster child of co-CEO success; now they are rewriting that pattern. Day to day, he insists, not much changed. Henrique was more external, Franceschi more internal. They have been at it almost ten years, counting the previous company in Brazil. Over the past seven months they massively simplified how the company runs: fewer management layers, fewer priorities, fewer decision makers. As Brex got closer to a public-company level of maturity, some calls needed a cleaner owner. They went back to CEO and chairman. Customers and investors still talk to the same people. It is mostly a formalization of what had been true for a few years.

Guo’s aside is that European AI companies seem to be going the other way. Dual-CEO structures are getting more common there, often a technical founder paired with a business-background one, Helsing among them. In the U.S., Robinhood and a few others walked that road. She finds the Europe-specific pattern striking. Then she asks how Brex is actually thinking about AI.

Franceschi splits the AI work into three buckets. Product is the obvious one, and inside it they spend time on accounting and on an expense assistant — basically what an executive assistant typically does. The second bucket is go-to-market and operations: prospecting, KYC, underwriting, compliance, the labor-heavy guts. The third is developer productivity. On that third one he does not claim anything remarkable. They use Copilot and are trying new tools, like everyone else. Buckets one and two get most of the time.

Guo asks when it became personal. About eighteen months earlier, right after ChatGPT launched. He had already played with GPT-3 APIs and been impressed; ChatGPT was the moment everyone asked what it meant for their business. The mental model, which he does not think is unique, is: if humans were free, what would we do? A lot of expense management and accounting would get automated. The best corporate-card experience is the one executives already have. There is no experience. You swipe, and it is done. An EA in the background finds the receipt, categorizes the expense, and gets it approved.

He prototyped that on a weekend, in Python, with GPT-3.5: pull someone’s calendar, use the context to generate a memo, categorize the charge, and maybe find a receipt in email. Results were surprisingly good. It took him about 48 hours to get something working. They tried it on a few real transactions. Within a week a team was building it into the side of Brex. That was the original impetus: give everyone that executive experience. Today roughly 30,000 customers use Brex Assistant for expense compliance, and over a third of their expenses are completed automatically — work that used to mean going in and doing it by hand. They want the number higher. It is a net-new product that started as playing with a model, seeing what worked, and shipping it.

Guo’s challenge is the one every traditional financial firm recites: the model is probabilistic, the ledger is not, so you cannot ship. Franceschi calls conviction — knowing the suggestion is correct — the holy grail of AI in fintech. Accounting is the sharp case, because people get fired if the results are wrong. Their first move is to expose ambiguity instead of burying it. A chatbot is the anti-pattern: one answer, no affordance, no view of the other options it considered, no context. Expense Assistant, when it generates a memo, runs three tiers. High conviction: this is what we strongly believe, and we apply it for you. Medium: suggestions under the field. None: show nothing.

The second move is traditional data science sitting on top of the models. When large language models landed, some data scientists wondered what their job was, now that anyone could produce something that looked pretty good. The laptop prototype is not 30,000 customers and hundreds of thousands of transactions a month. That jump takes statistical rigor most people do not appreciate. Those skill sets became valuable for scoring outputs, defining a good result, and learning how users rank a good result — which is different from what the company believes is good. The model generates. Data science referees.

The third move is to drop suggestions into a workflow that already exists. Very few accountants will accept "press this button and a bot in the corner closes the books." They will accept: you just coded three expenses the same way, we think fifteen more apply, click to review, here is a modal of everything that would change. Small UI patterns go a long way. Metrics he checked right before the recording: using that strategy, suggestion acceptance on the accounting side has climbed to 80 percent. If they had started with "click here and we close the books," there would not have been enough trust. People click done, then you skip a step. Trust is accumulated confirmation, not a launch announcement.

Guo adds the historical rhyme. In the 1990s people were uncomfortable typing a credit-card number into a website. Every technology wave has a trust tax and a UI that helps pay it. Hallucinations and ambiguity will get better as models get better; smart interface design can move the timeline. Prioritization, for Brex, was blunt customer obsession: where does this touch the largest number of people? An EA for every cardholder is high leverage. Accounting is something every admin on Brex does every month. Then the other half of the brain: cost and automation for Brex itself, serving customers through prospecting, demand generation, operations, and compliance. His original impetus was product. Right after that, a lot of time went to marketing, because almost nobody was thinking hard about AI in a B2B setting.

Outside the product, where does AI bite hardest? Scaling marketing, he says, and the same infinite-humans frame applies. Fifteen years ago, if you were Salesforce trying to close Coca-Cola, you staffed account-based marketing: a product-marketing manager beside the sales rep, a custom deck, a custom visit, a specific pitch of value. That cost only made sense for the largest logos. The cost of generating that treatment is now marginal. You can give account-based marketing to accounts that never got that care, and you can be specific about the value they would actually get.

This year Brex started quantifying that value in dollars: spend controlled, dollars kept on budget, hours saved, dollars saved. Over time they got rigorous about which industries, which company characteristics, which global footprints get more affinity and more value. Then a salesperson can say: your peer in the industry is on Brex, they also have a lot of global employees, they pay no FX fees because they run global payments on the card. An SDR blasting outbound would never invent that line. The hard system is treating demand as a system — tracking the entire TAM inside your own tools and targeting each account by the value it can get. Large language models help because they can chew unstructured data: a job posting that says the company is hiring somewhere else, a cold email that is actually relevant. That is a new playbook, not a slightly better spam campaign.

Guo notes the pile of tools that generate email, plug into an orchestration system, maybe touch LinkedIn, and call themselves an AI SDR. Brex has tried a lot of them. They are a commodity, not very different from one overworked SDR running a spam campaign. What Franceschi is describing needs demand signals and value signals that only Brex can collect. He reaches for *Crossing the Chasm*: a market is customers who can reference each other. You can now mint almost infinite markets — construction companies in Missouri that spend hard on card and have complicated accounting — at a granularity humans could not cover. Writing the email is the easy part. The hard part is a database of every account that could buy, what you know about them, and enrichment with signal that does not yet exist in the open. Marketing is alpha. If the data came from ZoomInfo into Salesforce, everyone would do it, and it would not be extra relevant. They built their own customer-data platform because Salesforce cannot ingest signal the way they need. Last-generation go-to-market infrastructure, Guo says, will become less relevant. He agrees. The playbooks are being rewritten. Composing the email or the ad is easy. Grasping the signal you need before you even write the prompt is the work.

Gil likes that some of Brex is durable in the face of AI: you cannot just point a model at the core and eat it. Franceschi admits people yawn at this. In 2018 the number-one feature was going to the website, signing up, and getting a card instantly — the first financial-services company on the planet, he says, to onboard instantly and hand you a virtual number you could swipe. He borrows Jeff Bezos’s question: what will not change in ten years? People are still going to move money. Crypto is interesting, and they are doing a few things there, but complexity on the money-movement side is the underpinning of all the software. He explains the business to new hires as two stacks that used to sit apart — banks moving money on the left, software and controls on the right — glued into one. Glue them and you get control. AI will make the software side great. At scale, moving the money is still hard.

They had just closed a public company with a hundred-billion-dollar market cap. One hundred and thirty public companies are on Brex. A main reason that new one signed was a card that works globally: operations across twenty countries, pay in local currency, settle in local currency. Brex does not sit on Stripe or Marqeta. It goes "straight to the metal," Mastercard and local payment rails, at large scale. That is why they signed, and it has nothing to do with AI. It is a boring moat you can build a software business on. Even the labor-intensive compliance work is only partly automatable. Adverse-media monitoring — acting when a customer shows up in news that looks shady or illegal — is a regulatory obligation most firms did not really operationalize because it did not scale. Brex now runs what he thinks is one of the more robust programs in the country by putting AI on it. Even so, at the end of the day, money movement is money movement.

Zoom out from Brex’s immediate roadmap and he still sees three horizons. Horizon one is corporate cards, where they started. Horizon two is total spend: monitor every dollar even when it is not on the card. Horizon three is continuous finance — real-time visibility into every dollar in and out. Finance teams are doing a reporting job, to shareholders and to the business, aggregating data that lives in too many places. A lot of accounting is data cleanup: categorize, make sense, stay legally correct. If you had human-level AI on that cleanup, you would not need the concept of an ERP. You would have a real-time layer across sources that today require people. Ask a CMO how they are doing against budget today and they often cannot tell you, because spend is on cards, invoices, and people traveling, and someone has to normalize it by hand into a database people call an ERP. Gil hears a change in the cadence of the company itself, off the quarterly close. Franceschi says that is why they use the name. Finance teams operate the way software companies did twenty years ago: a yearly release, a CD, a plan set in stone once or twice a year, while new data arrives every day and does not improve the plan. Continuous finance is continuous deployment for that team, which was not possible several years ago.

A few years ago Brex said it would stop serving SMBs and focus on high-growth startups and enterprises. Guo wants the decision, the implementation, and what they would tell other founders about not serving a segment. Franceschi’s line is that you can be anything, you just cannot be everything. They were growing across all segments and told themselves they could do small business, mid-market, enterprise, startups. Focus is what gives meaning to your choices. Spread across all of it, they were not deliberate about what they wanted to be world-class at. They did an okay job everywhere. That was not the goal. The customer-obsessed version of the choice, ironically, was going enterprise, because their customers were going there.

Scale AI is the story he tells. He met Alex when the company was five people and physically delivered the cards to the office. Five or six years later it is a massive company with a CFO, a global president, and compliance needs. Brex had to scale with the people who had bet on them. A lot of the company’s long-term value, they decided, was the ability to stick with a customer from garage to public markets. That made going up-market easier to see and still painful to do. They bit the bullet and, as he puts it, fired 20,000 customers. They would not be where they are without it.

Was the product for those 20,000 different, or the go-to-market, or just attention? There were product differences, he says, but the biggest thing was leadership bandwidth. People believe headcount or team size is what bottlenecks a company. He does not. It is how many things your leaders can spend time on and do well. They were spread across four completely different segments and could never make the trade-offs work. They did fewer things, eliminated one, and that is how they got there. Gil likes that the hour was as much about hard decisions as about AI. Franceschi thanks them and leaves it there.