The advice you keep hearing is simple: artificial intelligence can do it now, so point it at your hardest problem and watch the cost fall out. Every vendor deck says it. Every keynote says it. And for a wide band of work, it happens to be true. I spent eight years building automation inside fintech shops — reconciliation, regulatory reporting, the unglamorous plumbing that moves money — and I watched teams take that advice straight into a wall.
The wall, for me, had a number on it. A reconciliation pipeline a team I knew had wired to a large language model produced a $40M discrepancy that nobody caught for two weeks, because the model had quietly "rounded" a set of intercompany entries to make them balance. It was confident. It was fluent. It was wrong in a way that looked exactly like being right.
That is the lesson nobody puts in the keynote. AI is not one thing. The advice "use AI for everything" is built on a quiet assumption that all of it works the same way, and it doesn't.
Where the advice is roughly right
Give the hyped version its due, because it earns it. The generative models everyone means when they say "AI" are extraordinary at a specific shape of problem: open-ended, subjective, and forgiving of error. Drafting a first version of a marketing email. Summarizing a 40-page deposition into something a human will check anyway. Suggesting five names for a feature. Writing the boilerplate around code you'll review.
What these tasks share is that there is no single correct output, and a wrong-ish answer costs you almost nothing. If the model's draft email is mediocre, you edit it. If its summary misses a nuance, your eyes catch it. The model is a fast, tireless first-drafter, and used that way it is genuinely a step-change in productivity. I am not here to talk you out of that. I use it daily.
The trouble starts when the same enthusiasm gets pointed at problems with the opposite shape.
Where it breaks down
Some tasks have exactly one right answer, and the cost of a wrong one is not an edit — it's a fine, an audit finding, or a $40M hole.
A bank reconciliation either balances or it doesn't. A tax calculation is correct or it's fraud. A KYC check either matches the sanctions list or it exposes the firm to a regulator. A trade settles at the agreed terms or someone is liable. These problems don't want a plausible answer. They want the answer, the same one every time, traceable back to the rule that produced it.
A generative model can't promise you that, and it isn't a flaw in the model — it's the architecture. These systems predict likely text. "Likely" is the whole game. Ask the same question twice and you can get two answers; ask it a third time and it may invent a citation to support either. For drafting, that variability is a feature. For a regulated calculation, it's a defect you can't engineer away by writing a better prompt.
I watched smart people spend months trying. They added guardrails, they fine-tuned, they bolted on a verification layer that was itself a model. The error rate dropped. It never hit zero, and in compliance, "almost always correct" is a sentence you say to an auditor right before you lose.
Is generative AI reliable for financial or compliance work?
For tasks that require the same provably correct answer every time, the honest answer is no — not because the technology is bad, but because it's the wrong category of tool for that job. Generative models are probabilistic by design; regulated financial and compliance work is deterministic by requirement. You can use a generative model to draft a compliance memo a human signs off on. You should not use one as the system of record that decides whether a transaction is legal. The reliability you need there comes from rule-based, deterministic systems — the kind of software that, given the same input, returns the same output and can show its work line by line. That category has been around for decades. It is unglamorous. It does not write you a poem. It also does not hallucinate a $40M adjustment.
What the investor and the journalist are actually looking at
If you cover this market or put money into it, the segmentation is the story, and most coverage misses it because the word "AI" flattens two different businesses into one ticker.
When you read that a company is "applying AI to compliance," the first question worth asking is which kind. A firm selling generative tooling into a domain that demands deterministic output is selling a demo that will fail an audit. A firm selling deterministic automation but marketing it with generative-AI buzzwords is doing the opposite — underselling something durable to chase a hype cycle. Both are common. Both are mispriced in different directions.
The useful frame is not "is this company AI or not." It's "does the problem this company solves tolerate a wrong answer, and does their technology match that tolerance." Companies like Wayfair tuning product descriptions live in the tolerant zone, and generative tools fit. Companies like Capital One or JPMorgan running regulatory reporting live in the zero-tolerance zone, and they know it — which is why their production compliance stacks are not, despite the press releases, a chatbot.
How to tell which problem you have
Before you reach for a tool, sort the task. One question does most of the work: if the system gives a slightly different answer next time, who pays, and how much?
- If the cost is "I edit it" — generative tools fit. Drafts, summaries, brainstorms, code scaffolding.
- If the cost is "a regulator calls" — you need deterministic, rule-traceable software. Calculations, reconciliations, eligibility checks, anything that ends up in a filing.
- If you're unsure, assume the second. The expensive mistakes all live there.
Here's the same split in a table you can keep:
| Question about the task | Generative AI | Deterministic systems |
|---|---|---|
| Is there one correct answer? | No | Yes |
| Does a wrong answer cost real money or legal exposure? | Rarely | Always |
| Does a human review every output? | Usually | Often automated end to end |
| Can it show its work to an auditor? | No | Yes |
| Same input, same output every time? | No | Yes |
Neither column is the smart one. They're answers to different questions. The mistake — the one the advice encourages — is reaching for the left column on a right-column problem because the left column is the one that got the magazine cover.
The honest version of the rule
"Use AI for everything" isn't wrong so much as unfinished. The honest rewrite: use the kind of artificial intelligence that matches the cost of being wrong. For work where a near-miss is fine, the generative tools are a genuine accelerator and you should be using them. For work where the answer has to be the answer — money, law, filings — you need software that's deterministic by construction, and no amount of prompt engineering converts one into the other.
Both will matter for a long time. They are not rivals. They're two halves of a stack that most people are still talking about as if it were one.
Here's the thing to try this week. Take one process you've been tempted to hand to a generative model — invoice matching, an eligibility check, a numbers-heavy report. Run the exact same input through it three times. Read the three outputs side by side. If they differ in any way that would matter to an auditor, you've found a deterministic problem wearing a generative costume, and you've just saved yourself the version of my $40M.
