Fintech business account vs bank: four tiers, and the record on each one

Two people I trust gave me opposite answers to this in the same week, and both of them had better reasons than I expected.

The first runs finance at a company turning over about 4 million a year and told me that anything without a charter is a liability you have chosen to accept for the sake of a nicer dashboard. The second builds payment systems and told me the underlying bank is the same either way, so the distinction is marketing, and the real question is whose software you would rather live inside.

I went in agreeing with the first one. I came out somewhere neither would sign, which is the honest outcome and also the least useful thing to put in a headline.

The thing that changed my mind

It was not an argument. It was that the fintech business account vs bank question turns out not to be one question, because the products are regulated differently and the complaint record shows them behaving differently when something breaks.

In the year to July 2026 the federal record holds 84,177 complaints about checking and savings, 37,516 about money transfer and virtual currency, and 6,294 about prepaid cards. Different sized files, so the counts are not comparable, which is fine, because the interesting column is what happened next.

Checking closed with money going back 12.2 per cent of the time. Prepaid managed 746 out of 6,294, which is 11.9, near enough the same. Money transfer closed 2,699 out of 37,516 with money, which is 7.2 per cent, and that gap is too wide to be noise.

Deadlines run the same direction and more sharply still, and this is the column I would look at first if I were choosing today. 615 of 84,177 banking complaints were answered late, which is 0.7 per cent, and I would treat that as the floor of what a mature operation can hit rather than as anything impressive. Money transfer ran at 947 late out of 37,516. That is 2.5 per cent, more than 3 times the banking figure, on a file less than half the size. Prepaid sits at 274 out of 6,294. That is 4.4 per cent, 6 times the banking figure.

A short digression about the word late, because it is doing more work than it looks like. Late here means the company missed the deadline for responding to a regulator, which is the single easiest deadline in the whole relationship to hit, since it is known in advance and somebody is paid to watch it. Anyway, back to the tiers. I do not read that as anybody being worse people. It is not a character question. I read it as a measure of how much apparatus sits behind the answer, and apparatus is expensive, unglamorous and almost impossible to demonstrate in a product demo. A chartered bank has a compliance function built around the assumption that an examiner will eventually read its files. A money services business has a smaller one. Under load the first thing that goes is not honesty, it is speed.

What the API actually buys

This is the part the second person was right about and I had been undervaluing.

Developer access is why most technical founders move, and 4 quite different things get sold under that 1 word, which is the source of more disappointment in this market than pricing is. Read only access to balances and transactions. That is what accounting integrations need and nearly everyone offers it. Payment initiation, so that software can move money without a person sitting in a browser to approve it. Card issuing, which is a separate regulatory animal with its own paperwork, its own approvals and its own reasons to say no to you six weeks after you designed around it. And ledger level access, where you open sub accounts programmatically. You need that if you hold money for your own customers. It also turns you into the middleware in somebody else's diagram, which is the same position the company at the centre of the 2024 failure occupied, and if you are going to stand there you should at least know that you are standing there and that the record keeping obligation has followed you.

Ask which of those 4 you are buying before you look at the price, and get the answer in writing. I have watched two companies choose a provider on the strength of a documentation site and then find the endpoint they needed sits behind a partnership tier with a monthly commitment. The documentation had not lied about anything. They described a different plan, and I have stopped treating a documentation site as a description of what I can buy.

The time saved is real and I do not want to talk it down. Nobody puts a number on it, so here is one, stated as a model rather than as a measurement. Take a bookkeeper spending 4 days a month on reconciliation and an integration that takes a fortnight to build. If it lands where these integrations usually land, that 4 days becomes something under 1. At any sensible hourly rate that pays for itself several times a year, and it is a stronger argument for the software led providers than anything in their marketing.

Where the risk actually sits

The risk is not that the money vanishes. In the ordinary case the money is exactly where the disclosure says it is. The partner bank holding it is usually a small regional institution, with a real charter and real examiners.

It sits in 2 narrower places, and the regulator has written about the first of them in a way worth quoting. In October 2024 the FDIC proposed a rule requiring banks holding custodial accounts with transactional features to keep records "identifying the beneficial owners of those deposits, the balance attributable to each beneficial owner, and the ownership category in which the deposited funds are held". The stated purpose is to let the agency "promptly make deposit insurance determinations" if such a bank fails, and the word promptly is carrying a lot in a sentence about money somebody expects to reach on a Tuesday. A regulator does not propose a rule requiring somebody to know whose money is whose unless there was a case where nobody did. The record, because if a company between you and the bank keeps the ledger of who owns what, your access depends on that company continuing to operate and its books continuing to reconcile. And closure. A non bank can decide you sit outside its risk appetite and close the account on days of notice. The appeal process is whatever the agreement says, which is frequently nothing much.

The complaint file agrees on the second point in a way I did not expect. Closing an account produces 11,225 complaints against 7,412 for opening one. That is 51 per cent more noise on the way out than on the way in, and every guide written for business owners is about arriving.

A note on prepaid, which people bring into this comparison and probably should not. Prepaid against a business account compares 2 things doing different jobs: prepaid exists largely to serve people without a bank account, its complaint file is the most heavily tagged in the set, with 491 older Americans and 422 servicemembers out of only 6,294, which is 14.5 per cent of the file between those 2 tags alone, and its late response rate is the worst of any product I looked at. It is not a business banking product that happens to be cheaper. It is a different product for a different person, and I keep thinking about the 491 and the 422 in that file, because those 2 numbers describe who ends up carrying the worst response times in American payments.

Where I ended up

I went looking for a rule I could state in 1 line and this is the closest I got. Payroll and reserves at a chartered institution. Operating flow wherever the software is best. That splits the 2 risks apart and costs 1 transfer and about 1 hour a month of reconciliation, which is a real cost that never seems to reach anybody's comparison table, mine included until now.

This describes an arrangement rather than advice on yours. The sums here are usually big enough that whoever does your accounts should see the plan first.

I cannot tell you which of the 4 tiers is safest, because the record measures how companies behave when a complaint arrives and not how often money is actually lost, and those are different questions. I went looking and found nobody publishing the second one at all. I also do not know how much of the 7.2 per cent refund rate in transfers is product mix rather than conduct. A settled transfer is harder to reverse than a card charge. Some of that gap is physics rather than policy. I could not work out how much. My guess is that most of it is mix and a real part of it is not, and that is a guess with nothing behind it except the shape of the 3 files. What neither of my two people could help with, and what I still handle badly, is the middle case. A company running 400,000 dollars a month through a single provider, with an integration that took a quarter to build, has a switching cost that is not really about money and an exposure that is not really about insurance. That conversation has reached me 4 times this year and I have given 4 different answers, and I am no longer sure that is a failure of mine rather than a property of the question.

The scorecard, with the record attached to every row

Put the three public records side by side and the tiers stop being a matter of taste. Each row below is a federal aggregation bucket rather than an opinion, covering the year ending 1 July 2026.

Three things fall out of that table that I did not expect when I built it. The first is that banking is the tier most likely to give money back and the tier most likely to answer on time, which is an argument for keeping the boring account even when the exciting one is better designed. The second is that concentration rises sharply as you move away from banking: the top two names hold about a fifth of the checking record and 47.5 per cent of the transfer record, so in transfers you are picking from a much smaller field than the marketing suggests.

The third is the deadline column, which climbs from 0.7 to 2.5 to 4.4 as you move down the tiers. A late response is not a small thing when payroll is sitting behind it. It is the clearest single signal in the whole record that operational maturity and product novelty run in opposite directions, and it is why I would hold a second rail at a different tier rather than consolidate everything into whichever provider has the best dashboard this year. I rerun all 3 pulls every quarter and publish the movement, because a table like this is worth nothing if nobody keeps it current, and the whole reason for building it off a public API rather than a survey is that keeping it current costs 1 afternoon.