What It Costs to Run an AI Agent Across a Whole Stock Market
The short answer. Running a genuine multi-step agent over an entire equity market is a five to six figure annual cost, and that number is the reason almost every AI screening product quietly limits its universe to a few hundred large companies. The arithmetic is simple: companies covered, times checks per company, times cost per check, times how often you refresh. Every product decision in this category is downstream of that multiplication.
Nobody publishes this, so here is ours, with the numbers you would need to disagree with it.
The multiplication
An agent pass over one company is not one model call. It is a sequence: fetch the filings, extract the statements, compute the derived figures, check them against the previous period, look for the one-off distortions, and write the result. Call it six checks, which is on the low side for anything you would trust.
The arithmetic every vendor does privately
Where a full pass actually spends its time. The model calls are the bill; the fetching is the reason it is slow.
Set the universe to the number of companies you would want covered and the checks to what a trustworthy pass takes. The monthly figure is why the universe you are actually offered is smaller than the one you wanted.
Two numbers fall out and both are uncomfortable. Wall-clock time, which is why "real time" in this category usually means a cache, and cost per month, which is why coverage is the first thing to get cut.
Where the money in a full pass actually goes
The model calls dominate and they are the part that does not amortise. Everything else gets cheaper as the universe grows; that slice does not, and it is why the cost line is close to straight.
What gets cut, and in what order
The pressure is always the same and the concessions are always made in this sequence.
| Concession | What it looks like to you | Honest? |
|---|---|---|
| Shrink the universe | Only large caps are covered | Yes, if stated |
| Cache aggressively | Data is a day or a week old | Yes, if dated |
| Drop checks per name | Faster, shallower answers | Rarely disclosed |
| Run on demand only | First request is slow | Yes, and sensible |
| Pre-generate for popular names | Great on Apple, thin on everything else | Almost never disclosed |
The first two are fine and every serious product makes them. The last two are where the category gets slippery, because both produce a demo that is dramatically better than the product. If a tool is brilliant on the ten companies you first tried and vague on the eleventh, you have found a pre-generation strategy rather than a bad day. The twelve question checklist turns that instinct into something you can score.
Cost per company covered, as coverage grows
Unit cost falls with scale, because the fetching and parsing amortise. It never falls to nothing, because the model calls are per company and do not amortise at all. That floor is the whole economics of the category.
The same budget, spent two ways
Breadth and depth are the same money and every product in this category has silently picked a point on this line. Asking which one a tool chose is more informative than any feature list it will show you.
You can also run the strictness side of the same trade, which is where the other half of the cost hides: a screen narrow enough to be cheap is usually narrow enough to have thrown away most of what you were looking for.
The cheaper the screen, the more it throws away
Every filter you stack makes the reading step cheaper by shrinking what has to be read. Watch the good names cut column while you do it. That is the saving, itemised.
What we do, and the trade we made
We cover a wide universe: we currently hold 7,992 live companies, of which 3,065 carry a rating and 2,871 carry a discounted cash flow estimate. We got there by making a specific and arguable trade, which is that the expensive agent work does not run over everything.
The cheap deterministic layer runs across the whole universe every day: prices, statements, derived ratios, ownership, filings. That is a data pipeline, not an agent, and it costs what pipelines cost.
The agent runs on a shortlist and on request. When you open a company page or ask a question, the reading happens then, against data already in the database. That keeps the universe wide and the depth on demand, and the honest cost of it is that a company nobody has looked at recently has not been read recently either.
We think that is the right trade and we would rather state it than let you infer it from an unexpectedly thin answer on a small cap.
What this cannot tell you
The cost per call is the least stable number in this article. It has fallen by an order of magnitude in two years and may well do so again, which would change every conclusion here. The structure survives even so: it is a multiplication, so whichever term is largest is the one that determines the product, and coverage is usually it.
The bottom line
When a tool tells you its universe, it has told you its budget. That is not cynicism, it is the only inference available, and it is more informative than any feature list. Ask what is covered, ask how fresh it is, and ask what runs on demand rather than in advance. Then work out what you actually want covered with the screening cost calculator, which is the same arithmetic with your own numbers in it.
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.