Permission
In financial services, the ability to build arrives before the permission to build. Even a great product has to be let in, by an executive or a regulator and by the operators who will use it. We call that permission. Those who grant it are cautious, for good reason, and as AI has made building easier, it has only become harder to get. The industry is under pressure to adopt AI, and every vendor claims to be the answer, so the skeptical decision-maker has more people asking and less reason to trust any of them. That is what every company is really competing for. We back the teams that earn permission and compound the trust behind it.
Financial services runs on human judgment at industrial scale, reviewing and reconciling and adjudicating and advising. In the US alone, financial institutions spend over $1 trillion a year, roughly the world’s entire software budget, on the labor that turns unstructured documents into decisions. That includes a credit analyst spreading three years of a borrower’s returns for loan committee, a medical coder deciding whether an inpatient bill carries the right diagnosis, and an AML analyst deciding whether a flagged wire is laundering or a coincidence. There are about four people in finance and insurance for every software developer in the country.
Banking, insurance, wealth, healthcare, and real estate all run on the same kind of machinery, and the systems underneath predate everyone in the room, held together with mouse clicks, keystrokes, and human judgment. They are antiquated for a reason: replacing core infrastructure within financial institutions takes years, often never finishes, and the cost of failure is too high, so each generation built around it. The operators inside face it every day, and the customer pays for it, in a claim denied on a coordination-of-benefits error or a construction loan that waits three weeks for an answer.
The last twenty years of fintech mostly built new institutions beside the old machinery rather than reaching into it. Even those needed a bank to sponsor them or a network to admit them. The ones that wanted to carry the risk themselves paid for the permission to do it. Lemonade got licensed as a carrier before it wrote a policy, and SoFi bought a bank for its charter. Charters come faster today than at any point since 2008, but the license is only the first of many permissions a company needs.
The judgment is now buildable.
The software already inside could automate the edges of a process, but it could not reach the person exercising discretion in the middle. The judgment was the expensive part, and until recently there was no way to build it. Getting to it meant a multi-year core modernization first, a systems integrator to run it, and a team large enough to write rules for every exception the adjuster already handled in her head.
A model can now read the claim file, the tax return, or the medical record. A team of three can now go from prototype to a deployment inside a live workflow in months, without replacing the system underneath. The intelligence is bought as a cost that falls every quarter, not a capital project that has to be approved and staffed.
The underwriter who could only get to the submissions that fit in a week now sees every one that needs attention, so the carrier quotes faster and writes 50% more premium. Credentialing a provider, which took a health plan 120 days, now takes a single day, and the four months of appointments that used to be lost now happen. The obvious result is lower labor spend. The larger one is a superior institution, one that grows faster because it answers faster, selects risk better because it sees all of it, and serves the long tail of customers it used to decline because they were uneconomical.
Someone has to let you in.
Somewhere inside every institution is a person whose name goes on the decision to let you in. If it goes wrong, the board, the regulator, or the customer calls them. Whoever signs carries that risk personally, and every prior wave of technology underdelivered on its promise to them. The AI wave has arrived with more pressure to adopt and more vendors at the door than any before it. Intelligence is abundant, but it is only put to work where that person lets it in.
That permission exists whether a company sells into an institution or becomes one. To run an institution’s work, you need the consent of the executive above it; procurement, security review, and model governance are how that person justifies a decision already made on instinct. To become an institution, you need the consent of the state and the market: a license, a rating, capacity behind your paper, distributors willing to sell it. Each is granted by someone with every reason to say no, and the question underneath is whether they can trust you.
Permission is also granted from below. The executive signs, but the operators decide whether the product ever gets used, and they decide early, in discovery and in the pilot. With AI that scrutiny is sharper, because what is being sold reaches into their job, whether it helps them or replaces part of it.
What both are reading is whether the founder understands their business, whether they know what a loss run is and what the adjuster is afraid of, and what a covenant breach means to the credit officer who signed it. You cannot replace what you do not understand. Some of the best founders we have backed came from outside the industry with no credential and learned it cold, sitting with the people who do the work until they could think like them.
Once they trust that you are fluent, the story does the rest. It names a pain the industry has normalized but cannot yet articulate and reframes the status quo as the real risk to a buyer with no incentive to move. It recruits A+ talent into problems the sector cannot sell on its own and keeps the team, the champions, and the capital believing through the stretches with no obvious signal. That is why we read a founder’s pitch as carefully as their product.
Trust has to compound.
Incumbents are not going anywhere. The trust behind the largest of them has compounded for more than a century, and it is the same trust a newcomer has to earn. Three of the four largest banks in the US were founded before the Civil War, and the five largest life insurers between 1845 and 1875. Their position rests on four moats — regulatory, data, capital, and brand. A new company can build each of them sooner than it once could, but none of them quickly: a license runs on the regulator’s calendar, data comes with deployments, and capital arrives fast only for those that are legible to the market. The brand is slowest of all, because a brand is a promise that has been kept, over and over. There is no way to compress the time that takes. Some incumbents will adapt with AI, and the best companies will help them do it. Others will be challenged from outside.
Whether they are enabling or challenging an incumbent, the companies we back hold permission in one of four ways. Some sell institutions the judgment they now buy as labor, and hold the institution’s permission to run it. Some take a licensed position, as a carrier, lender, agency, or plan, and hold the state’s permission to carry the risk. Some own the service outright, the administrator or the brokerage, and hold the customer’s permission directly. And some build what everyone else needs to be let in, from connectivity to capacity to custody, and hold a piece of every deployment that runs through it.
Each captures something the others cannot. But permission alone does not compound. Every institution is surrounded by firms that have held its business for decades, the auditor and the consultancy among them. They have permission but not leverage, because the hundredth client cost as much to serve as the first.
What we back is trust that compounds. Each customer leaves behind a workflow the institution now runs on, data that sharpens the judgment for the next one, and a name the next executive or the next reinsurer already knows. Once earned, trust widens what you are allowed to solve, whether an adjacent line, a neighboring function, or the same judgment at the institution next door. Generalized intelligence is becoming a commodity. The layer that holds one institution’s trust and carries its risk is not.
Illegible is not the same as small.
This has never been a build-it-and-they-will-come industry, so we focus our underwriting on whether a team can earn permission and compound the trust behind it. The obvious signals do not always show that. To us, one institution that has put its name on you counts for more than ten in pilot, and one cohort underwritten at a loss ratio that attracts quality capacity counts for more than top-line growth. While product gaps can close in a quarter, permission cannot be copied, and it is rarely quick.
From a distance, that can read as little movement, and the complex, overlooked corners of financial services can look like a collection of niches. But illegible is not the same as small. The same complexity and regulation that make these corners hard to read are what keep them overlooked. Up close, these opportunities are enormous and the returns are real, just stranded behind the proximity and time it takes to see them.
The judgment we strive to bring is knowing when slow is a problem and when it is just the shape of the work. We cannot make a founder credible in a room they did not earn. What we can do is route permission: open the door to the COO who will make the call or the reinsurer weighing a new program, put the founder in front of another founder who has already been through dozens of bank RFPs, and make sure the next round of capital reads a quiet year correctly instead of pricing it as a miss.
The machinery behind financial services went untouched because the judgment inside it could not be codified, and replacing the systems underneath took years and risked too much. The judgment can be built now. Everything that remains is permission.
