A few weeks ago I wrote about Gartner's forecast that over 40% of agentic AI projects will be cancelled by 2027 — and why the three named causes (escalating costs, unclear business value, inadequate risk controls) are governance failures, not technology failures. This week I want to show what the fix actually looks like in practice, because we just ran it on a real engagement.
A regional FMCG manufacturer had eight candidate AI agents on the table — order validation, demand-signal planning, customer messaging, vendor scoring, network optimisation, and more. Leadership's instinct was reasonable: build them, fast, and start capturing value. Our job was to test that instinct before a single line of code was written.
Key Takeaways
- The framework: every candidate agent was scored against four gates — Data (does the information actually exist?), Defined (is there one agreed meaning for what it's deciding?), Governance (is there a named human owner?), and Free (is it blocked by another system's migration?).
- The result: of 8 candidates, only 2 cleared every gate. Both were tied to named, quantified value: one addressing roughly AED 5.49M in network inefficiency, the other addressing losses inside a roughly AED 2.82M cancelled-order figure.
- The other six weren't failures — they were accurate diagnoses. Three were blocked purely by a missing governance owner, fixable in weeks at zero engineering cost. Two were waiting on a system migration already in motion. One had no agreed definition at all across the departments that would use it.
- The cheapest fix is usually skipped first. A named owner and an agreed definition cost nothing to put in place, yet they were the single biggest blocker across the eight candidates — not data, not budget, not technology.
Why eight good ideas isn't the same as eight ready ideas
Every one of the eight agent concepts on that list was a legitimate business idea. None of them were hype, and none were "agent washing" in Gartner's sense — rebranded chatbots dressed up as autonomous systems. The problem wasn't the ideas. It was the gap between an idea being good and an idea being ready, and nobody had built a way to tell the difference before committing budget.
That gap is exactly where Gartner's 40% cancellation forecast lives. Escalating costs happen when a team builds against data that turns out to be wrong. Unclear business value happens when nobody agreed what the agent was actually deciding. Inadequate risk controls happen when there's no named human accountable for what the agent does. All three are visible before you write any code — if you go looking for them.
The four gates
We score every candidate agent against four binary-ish questions, using evidence from the client's own systems rather than a vendor's demo or a stakeholder's optimism.
- Data. Does the specific information this agent needs exist today, in a form it can actually use — not "could exist with a project," but exists now.
- Defined. Is there one agreed definition for what the agent is deciding? An agent cannot act consistently on a term that three departments each define differently — and in our case, one candidate failed here outright: three functions used the same word for a cancelled order, and none of them meant the same thing by it.
- Governance. Is there a named human who owns this agent's outcome and a checkpoint where they approve its actions? No owner means no build, regardless of how ready the technology is.
- Free. Is the agent blocked by a dependency on another system or migration still in progress? If so, it waits. Building on top of a moving foundation just guarantees a rebuild later.
Four green lights means build now. Anything else means the agent isn't cancelled — it's sequenced, with a clear, usually inexpensive fix identified before it goes anywhere near a development budget.
"Saying no to six agents wasn't caution. It was how we protected the budget behind the two that could actually survive contact with real data."
— Lisa Warren, Founder & CEO, Neural Horizons AI
What this means for your own agent backlog
If your organisation already has a list of agentic AI ideas waiting for budget, the highest-leverage thing you can do this week isn't to build the most exciting one first. It's to run all of them through these four gates and see what the list actually looks like once you take data quality, definitions, governance and system dependencies seriously.
Most lists come back looking like this client's: a couple of ideas genuinely ready to go, and several more that are one cheap governance fix away from being ready. That's not a disappointing result. It's the result that keeps you off the wrong side of Gartner's 2027 number.
The three questions worth asking before you build anything
- Can you name, in one sentence, the human who is accountable if this agent does something unexpected?
- Would every department that touches this agent's decision define the key term the same way, right now, without checking with each other first?
- Is the data this agent needs sitting in a system that isn't mid-migration, mid-cleanup, or mid-anything?
If the honest answer to any of the three is no, you haven't found a reason to cancel the idea. You've found the cheapest possible moment to fix it — before the budget, not after it.