Why enterprise AI dies in staging
Every failed enterprise AI project we've been called into died the same way. Not with a crash — with a status update. "The pilot is in staging. We're waiting on security review. Q3, probably."
It's been in staging for eleven months.
After enough of these autopsies, the pattern is boringly predictable. Four steps, in order:
1. The strategy phase eats the conviction. Three months of workshops produce a 200-slide roadmap. Everyone agrees with it. Nobody is accountable to it. The budget is half gone and nothing has touched production data.
2. The pilot is built on a snapshot. To move fast, the vendor asks for "a sample export." The model works beautifully on the sample. It has never met your actual stack — the auth, the nulls, the region that formats dates differently, the ERP that times out on Mondays.
3. Nobody owns a number. Ask what metric the pilot is supposed to move and you get a mission statement. If no one committed to "reporting time from 6 hours to under 30 minutes," then no one notices — or minds — when it ships nothing.
4. The handoff to IT is where it goes to die. The people who built it leave. The people who receive it never wanted it. Staging is a polite hospice.
Notice what's missing from all four: users. Real people, using the thing, on real data. That absence is the disease; everything else is a symptom.
This single constraint quietly kills all four failure modes. You can't strategy-phase your way out of a Friday demo. You can't build on a snapshot when the deliverable is wired to the live stack. A weekly demo in front of the person who feels the pain forces a number onto the table. And there's no handoff cliff, because the thing has been in production — smally, safely — since day five.
"Small" is the load-bearing word. Week one is never the moonshot. It's the reconciliation job everyone hates, the report that takes someone's Thursday, the data that moves between two tools by hand. Unglamorous — and exactly where the trust (and usually the first crore) is hiding.
This is why our engagements run automate → analyze → application, three weeks to first login. Not because we're impatient. Because the alternative has a body count.
Next week: the three-week MVP, day by day — what we actually do between the scope call and the first login.
— Piyush