A bank's AI telling you you're not financeable is a decline with no paperwork

Six agents, and only one of them is interesting

Huntington Bank’s new small-business platform launches with roughly six trained artificial-intelligence agents. Five of them do admin: walking an owner through registration and licensing, drafting digital ad copy, building a social media campaign, simplifying the daily workflow of someone who, as the bank’s president put it, wears lots and lots of hats. Fine. Useful, even.

The sixth is different in kind. The bank says 1866 Studio will identify issues the business might run into when seeking financing. Read that again. A lender is offering to tell a business owner, in advance and for free, what a lender will object to.

That is the only thing on the list that no one else can credibly provide, which is precisely why it is there. It is also the only thing on the list that touches a regulated act.

The decline that never happened

American credit law is built around the application. A creditor that turns down an applicant has to notify them and give the principal reasons. The applicant ends up with a document, the bank ends up with a file, and a supervisor years later can count the declines, sort them by neighbourhood, and ask awkward questions about the pattern.

None of that machinery engages when the answer arrives before the application. An owner who is told by an agent that their receivables look thin, their licensing is incomplete, or their records will not support the ask does not become a declined applicant. They become someone who decided not to apply. There is no notice, no stated reason the owner can challenge, and nothing that necessarily lands in a dataset.

Fair-lending rules in the United States have long reached beyond formal declines: a creditor is not supposed to say things that would discourage a reasonable person from applying. That rule was written with a loan officer at a desk in mind — a human whose discouraging remarks were at least episodic and deniable. An agent trained on the bank’s own appetite is that loan officer at industrial scale, running on a product that is explicitly open to people who do not bank there, and leaving behind only whatever log the bank chooses to keep.

Nothing here says Huntington is doing anything improper. The bank has described a feature, not a decision, and the feature may well be conservative, generic and heavily guardrailed. The point is that we cannot tell, and neither can anyone else, because the obligation to show your working stops at the application door.

Free tooling for non-customers is a funnel, and that is fine

The more commented-on feature of 1866 Studio — that it is available to people who are not Huntington customers — is the least troubling part. A bank with half a million small-business relationships giving away operating software to businesses it does not yet serve is building a pipeline. That is honest competition, and it is the same instinct Fifth Third’s chief executive described this week when he said incumbents have a lot of leverage here provided they respond.

But it does change who is standing inside the tent. A non-customer who uses the studio hands over an operating picture — what they sell, who owes them money, whether they are licensed, what they spend on marketing — without ever opening an account. They get none of the relationship a customer has, and the bank gets most of the information a lender would want. If that information later informs how the same business is treated when it does apply, the applicant should know. If it does not, the bank should say so.

The strongest case against me

Here is the honest version of the other side. Small-business owners waste enormous effort on applications that were never going to succeed, and a wasted application is not free — it costs time, documentation and sometimes a credit inquiry. An agent that says plainly “fix your bookkeeping first” is doing the applicant a favour that the formal process does badly and slowly. Regulating that honesty into existence risks regulating it out of existence: the cautious legal answer is for the agent to say nothing useful at all, which helps nobody.

That case is good, and it is why the answer is not prohibition. But it rests on the assumption that the advice is accurate, consistent between one business and another, and offered without regard to the bank’s own interest. None of those three is guaranteed. Generative systems produce confident wrong answers, including about licensing requirements that vary by state. Consistency across thousands of conversations is not something a bank gets for free; it is something it has to test for. And a lender has an obvious interest in steering the marginal case away rather than through, because approval statistics improve either way.

An agent that is accurate, consistent and disinterested is a genuine public good. The way to find out whether this one is, is to look at the records — which means the records have to exist.