What Is an AI Financial Agent? A Definition That Can Be Failed
The short answer. An AI financial agent is a system that is given a goal rather than a query, and that plans, calls tools, reads the results and decides what to do next until the goal is met or it gives up. The distinction that matters is not intelligence, it is autonomy over the sequence of steps. A chatbot answers what you asked. An agent decides what to look at. As of 27 July 2026, most products marketed as financial agents are the former with a longer context window.
The word has been stretched until it means almost nothing. A newsletter with a summarisation step calls itself an agent. So does a screener with a text box. Both are useful; neither plans anything. It is worth having a definition that can be failed, because a definition nothing fails is decoration.
What actually separates an agent from a chatbot
Four things, and you can check all four from the outside.
| Capability | Chatbot | Retrieval assistant | Agent |
|---|---|---|---|
| Decides which sources to open | No | No, fixed retrieval | Yes |
| Runs more than one step without you | No | Rarely | Yes, that is the point |
| Checks its own output before showing it | No | Sometimes | Yes, or it is not an agent |
| Can decline the task | Almost never | Rarely | Yes, and it should |
That last row is the one people skip and the one that predicts whether a system is usable on money. A system that cannot say "I could not establish this" will invent it, because generating a plausible sentence is the only thing it knows how to do when the evidence is thin.
Where the products calling themselves agents actually sit
Scored against the four capabilities above, out of four. Our own screening agent is on this chart at three, because it plans and it checks, and it is still poor at declining.
What has to be built on top of a model before it is an agent
Illustrative, and it is the honest shape of the work. The model is the smallest layer. Everything above it exists to stop a fluent sentence being mistaken for a fact, and it is where almost all of the engineering in this category actually goes.
The abstention problem, which is the whole problem
An agent that answers everything is wrong more often than an agent that answers most things, and the difference is not small. Raise the bar at which a system is willing to commit and two things happen at once: it says less, and what it says gets sharply more reliable. The interesting number is neither of those. It is how many confidently wrong answers still reach the reader.
Coverage against correctness
Drag the confidence bar. Notice that the third cell, confidently wrong answers per hundred questions, is the only one that matters to a person acting on the output, and it is the one nobody publishes.
Anyone can make that third number look good by refusing to answer anything. The engineering is in moving the whole curve, not in picking a point on it, which is why a vendor quoting you an accuracy percentage without a coverage percentage has told you nothing at all.
What a hedge costs against what a refusal costs
Out of a hundred questions the system cannot answer. A refusal costs one thing: the reader goes elsewhere. A hedge costs several, and the expensive one is the last row, where somebody acted on a sentence that was never an answer.
What we run today, stated plainly
SteadyShares has agents in production. They screen, they summarise ownership and filings, and they answer questions about companies against data we hold rather than against the open internet. Right now we currently hold 7,992 live companies, of which 3,065 carry a rating and 2,871 carry a discounted cash flow estimate.
They are useful and they are not finished. Three specific things are true and we would rather write them here than have you discover them:
- The abstention is weak. Our agent is more likely to produce a hedged answer than to say it does not know. Hedged is worse: it reads as an answer.
- The planning is shallow. It runs a mostly fixed sequence and adapts within it. Genuine replanning, where a dead end causes it to go somewhere it was not going to look, is in build rather than shipped.
- Coverage is uneven. A large US listed company gets a far better answer than a mid-cap listed outside the US, because the underlying filings coverage is better and the agent inherits that.
The next generation is being built now, and the section on it further down says what specifically changes rather than that it will be better. If you want the full self-assessment rather than the summary, we scored ourselves on ten capabilities and published the four we fail.
Why we are building abstention before we build anything more impressive
Set the error rate to what you would accept and then drag the review coverage. Reviewing everything is not a policy anyone runs twice. That leaves making the system decline, which is why it is first in the queue rather than fourth.
What we are actually building, and in what order
- Shipped+1
Grounded answers over our own database, with the source row shown next to every figure.
- Shipped+2
Screening agent that reads the filing rather than only the metric.
- In build+3
Real abstention: a calibrated refusal path, and a visible confidence on every claim.
- In build+4
Replanning: the agent chooses its next source based on what the last one said.
- Next+5
A public evaluation set, so our accuracy claims are checkable rather than asserted.
Written down so it can be held against us. Anything not shipped says in build, and a date it did not hit is a date that stays visible.
What this cannot tell you
It cannot tell you whether any given agent is good, because the only honest way to know that is a held-out evaluation set, published, that you can inspect. Almost nobody publishes one, including, as of today, us. Until that changes, treat every accuracy claim in this category, ours included, as marketing rather than as evidence, and judge the products on the things you can see: whether they show their sources, whether they ever refuse, and whether the numbers they quote match the filing when you open it.
The bottom line
The word agent is doing a lot of work in this industry and almost none of it is load bearing. The test that survives contact with reality is boring: can it choose where to look, does it check itself, and will it ever tell you no. Ours passes two of those three. If you want to see the mechanics rather than read about them, the screening agent question set is the twelve questions we ask of our own, and you can point them at anyone's.
Educational information, not financial advice. Where this page states a figure about our own coverage it is read from the database as the page loads, so it is today's number rather than the day this was written.
One well-researched article at a time. No spam, unsubscribe in one click.
No spam, no selling your address, unsubscribe in one click. The tools stay free either way.
Keep exploring: browse the stocks we cover or see what the super investors hold.