Ten Days to Six Hours: Putting AI Inside the Workflow, Not Beside It

4 MINS

Market research used to take my team ten days. It now takes about six hours. That number gets quoted back to me a lot, usually with the assumption that we found a clever tool. We didn't. We changed where the tool sat.

Beside the workflow is where AI goes to die

Most teams adopt AI as a side quest. Someone opens a chat window, pastes in a question, gets a decent answer, copies it into a doc, and the doc goes into a folder. The output is good and the process is unchanged, which means the time saved is roughly zero and everyone concludes the technology is overhyped.

The shift for us was treating AI as a step in an existing pipeline rather than a place you visit. Bringing GitHub Copilot and similar tools into the workflow where the work already happened, rather than asking people to context-switch out to them, is the entire difference between a ten-day cycle and a six-hour one. Same models, different plumbing.

Speed changes what you're willing to ask

The compounding effect isn't the time saved on any one piece of research. It's that cheap research changes your behaviour.

When a competitive scan costs ten days, you do it once a quarter and you defend the conclusion for three months, because re-running it is expensive and admitting you were wrong is more expensive. When it costs six hours, you run it before every meaningful decision and you stop being attached to last quarter's answer. Three features shipped in two months for us, and $2M of revenue growth for InstallShield followed, less because the research was better and more because we asked more often.

That's the underrated part of the AI conversation. Everyone measures output per hour. The real change is in how many questions you're willing to ask when asking is nearly free.

What still needs a person

I'd be lying if I said this scaled infinitely. Three things have not gotten cheaper:

Knowing which question matters. A fast wrong question is still a wrong question, delivered sooner and with more confidence.

Judging source quality. In enterprise licensing, half of what's publicly written about a competitor's pricing is out of date or aspirational. A model won't tell you which half.

Owning the call. Nothing signs off on a roadmap bet except a person who will be in the room when it doesn't work. So the research got fast and the decision-making got slower, in a good way. There's more time to argue about the right things now that nobody is spending a week collecting inputs.

The honest version

If you're trying to reproduce this, don't start with the tool. Start by drawing the workflow you actually have and finding the step where humans are moving information rather than judging it. That step is your candidate. Collection, formatting, summarising, cross-referencing, those are the ten days.

Then put the tool in that step, not in a browser tab next to it. That's the whole trick, and it's much less interesting than the number it produces.

Sprint Variance Was 60%. The Fix Wasn't Better Estimation.

Problem Discovery Is the Whole Job, Everything Else Is Scheduling

Ullas skipped presentations and built real AI products.

Ullas D.P was part of the June 2026 cohort at Curious PM, alongside 20 other talented participants.