Eight candidate AI agents were on the table. Before a single line of code was written, we tested each one against a readiness framework — data, definition, governance, and freedom from blocked systems. Only two passed. That decision protected the investment, and it's exactly the discipline Gartner says most agentic AI projects skip.
The client is a large regional FMCG manufacturer and distributor, running production and delivery at national scale. As part of a wider specialist consulting engagement, our team was brought in to lead the agentic AI readiness evaluation — a focused diagnostic asking one question: which of the AI agents this business wanted to build could actually survive contact with its own data?
A full-year read of the company's own invoice, order and delivery records surfaced the real picture. Over 20,000 orders had been cancelled after the delivery plan was already built, worth roughly AED 2.82 million in the twelve months examined. A separate network inefficiency was quietly costing the business about AED 5.49 million a year in revenue delivered "the long way round" — customers served from the wrong depot simply because nobody owned the decision to fix it. Peak van-sales returns were running at 28.8%, nearly ten times a healthy benchmark.
Leadership's instinct was reasonable and common: build AI agents to fix all of it, fast. Eight agent concepts were already on the table — order validation, demand-signal planning, customer messaging, vendor scoring, network optimisation, and more. The temptation was to greenlight them all.
Does the data this agent needs actually exist, and is it trustworthy? Not "could exist" — exists, today, in a form the agent can use.
Is there one agreed definition for what the agent is deciding? An agent can't act consistently on a term three departments each define differently.
Is there a named human owner and an approval checkpoint for what this agent will do? No owner, no build — regardless of how ready the technology is.
Is the agent blocked by a dependency on another system or migration still in progress? If so, it waits — no matter how valuable it looks on paper.
Every one of the eight candidate agents was scored against all four gates, using evidence from the client's own systems — not opinion, and not a vendor's demo. The scoring was blunt on purpose: green means build now, amber means blocked but fixable, red means absent entirely.
Two agents cleared every gate immediately: a network and cost-to-serve agent directly addressing the AED 5.49 million inefficiency, and a cancelled-order validation agent addressing losses inside the AED 2.82 million figure. Both had clean data, an agreed definition, and no blocking dependency.
The other six were not failures — they were accurately diagnosed. Three were blocked purely by a missing governance owner, fixable in weeks at no cost. Two were waiting on a system migration already in motion. One had no agreed definition at all: three different departments used the same term to mean three different things, and no agent can act consistently on a word nobody agrees on.
With the readiness evaluation complete, the two cleared agents are now moving into build under a phased 100-day plan — sequencing governance fixes for the blocked agents alongside deployment of the two that were ready from day one.
The plan runs in four stages, each with its own committed milestone rather than a single distant go-live date. That structure is deliberate: it gives leadership a working system to check in on every 25 days, not a single high-stakes deadline at the end of a quarter.
One agreed definition and one named owner for every number the two agents will act on.
The cancelled-order agent goes live against the AED 2.82M figure it was built to address.
A single live dashboard replaces the reporting gap the original diagnostic found — the data existed; nothing showed it.
The network agent runs the weekly rhythm against the AED 5.49M inefficiency, live and unattended.
Phase 2 is in progress at the time of writing. We'll update this case study with delivered results once the 100-day plan closes.
All eight concepts were genuinely good ideas. Only two were ready today. Confusing the two is exactly what Gartner says drives most of the agentic AI projects being cancelled by 2027.
Three of the six blocked agents needed nothing more than a named owner and a defined term — zero engineering cost. That's usually the cheapest fix available, and the one teams skip first.
Building all eight at once would have spread the same budget across agents with no chance of surviving contact with real data. Sequencing by readiness put the full investment behind the two that could actually deliver.
Readiness evaluation, ROI modelling, and a sequencing plan built around your actual data — not a vendor's roadmap.
Learn More →Bespoke AI agents designed and deployed only once they've cleared a real readiness bar — data, definition, governance, and freedom to act.
Learn More →Book a free 30-minute discovery call. No pitch — just a conversation about which of your AI ideas are actually ready.