There's a straightforward way to build browser automation with AI in 2026: hand a model the goal and the screen, let it decide what to click, repeat on every run. It demos beautifully. We built one early on, and it did the thing.
Then we ran it two hundred times, and the problems showed up in the order you'd expect.
The cost is linear in your success
Every run costs tokens. That's fine for a task you run twice a week. It's not fine for a task you run for every row of an eight-hundred-row spreadsheet, or nightly, or per customer. The economics work backwards: the more useful your automation becomes, the more it costs you. Businesses end up rationing the thing that's supposed to save them work.
With TaskMate, the model is involved while you're defining the tool. After that, a run is deterministic playback — 0.02 credits per action, no tokens. Eight hundred rows costs about $2.90 instead of a real conversation with your finance team.
The latency is per step, not per task
A model deciding what to click has to look, reason and answer for every single step. Twenty steps is twenty round trips. That's the difference between a task finishing in twelve seconds and finishing in four minutes — and if it runs a hundred times, that's the difference between overnight and not-overnight.
The part people underrate: it isn't the same twice
This is the one that decided it for us. A model working the screen fresh each time may take a slightly different path, interpret an ambiguous label differently, or handle an unexpected banner in a way it didn't last week. For open-ended research work that flexibility is the whole point. For filing a return, it's disqualifying.
Nobody wants a compliance filing to be creative. They want it to be the same filing as last month, with this month's numbers.
When someone asks whether an automation is trustworthy, what they're really asking is whether it will do the same thing tomorrow. Deterministic playback can answer that question. Per-run inference can't, quite.
So where does AI still earn its place?
At definition time, where it does something genuinely hard: turning a sentence like “find the filing status for this registration number” into a working sequence of steps, with the right elements identified and the right inputs named. That's a task where variation between attempts costs nothing, because a human is right there looking at the result.
And at the specific steps where judgement is the point. Drafting the commentary paragraph on a weekly report is a fine use of a model. Reading the numbers off five systems isn't — those should be identical every week, and a model is the wrong tool for making sure they are.
What this costs us
It's not free, architecturally. A recorded tool can break when a site changes, where a model might have adapted. We think that's the right trade: a task that stops and says which step broke is more useful than one that quietly does something else. But it's a real trade-off, and we'd rather say so than pretend the approach has no downside.
Written by the Deccan Logic team. More posts · Try TaskMate free