Gates before spend

Logistics

Gates before spend

Reading a vendor's AI showcase like an analyst

  • 1 of 7demos that ran on real production data
  • 4recommendations to leadership
  • 90 daysgated plan before any spend
Gates before spendscreenshots and video coming

The problem

A global cloud vendor and its integration partner ran an AI session for the company. Leadership asked me for an independent view. Every demo but one ran on dummy data; the one production case was from another industry. The most dangerous number in the room was “two to three days to deploy”.

What I did

  1. Classified the session. A capability showcase, not discovery. Separated signal from theatre.
  2. Ranked use cases by fit, not by demo quality. Keep: grounded question-answering on the Workspace licences already paid for; a multilingual WhatsApp shipment-status bot; a revenue-leakage agent on detention and surcharges where the data was already being captured. Kill: generic growth tooling, a price-variance operator as shown, image and video generation.
  3. Wrote the gates. A document Q&A pilot must reach 60 percent weekly active use and 90 percent citation-correct answers before any spend. A status bot must hit 85 percent intent accuracy, 50 percent containment and zero wrong confirmations.
  4. Four recommendations. Fix data before buying AI. Use the licences already paid for. Run gated pilots, not platform deals. Put clear AI ownership inside the company.

Decisions I would defend

  • Bring your own requirements. A vendor discovery workshop writes them for you otherwise.
  • Offer your own tools as evidence. Doori (no AI in the product) and FleetSetu (built with AI) showed that in-house capability exists and that data readiness is the constraint.
  • Independent voice. An early draft signed as an “AI transformation lead”. I corrected it. I hold no such title and the note was stronger for saying so.