The Assay · 5 August 2026

The Test We’re Proud to Fail

There is a test going round for anyone selling AI agents. Buyers are being told to put one question to the vendor and watch the reaction: “Show me your agent making a decision my system didn’t pre-approve.” It is designed as a trap. A real agent, the argument runs, reasons for itself when it meets something new — anything that only follows approved paths is a workflow wearing an agent’s name badge, at several times the price.

Ask me that question and the answer is no. The agents we build do not make decisions the system has not approved. They never will. And I would rather lose the sale than change the answer.

The sceptics are right about the market

Before I defend that answer, the concession — and it is a full one. Gartner has a name for what the test is built to catch: agent washing, the rebranding of chatbots, RPA scripts and assistants as autonomous agents without the capability to match. Of the thousands of vendors now marketing agentic AI, Gartner estimates about 130 are real. The same analysts predict more than 40% of agentic AI projects will be cancelled by the end of 2027 — escalating costs, unclear business value, inadequate risk controls. S&P Global found the share of companies abandoning most of their AI initiatives before they ever reach production rose from 17% to 42% in a single year. An MIT study put the share of enterprise generative-AI pilots showing measurable return at roughly one in twenty.

So the test is doing useful work. Buyers who are being sold decision trees as autonomy deserve a question that exposes it in the room — and I have sat on the buying side of enough overpromised technology, years before anyone called it AI, to have no quarrel with any of it.

Where the test goes wrong

The test rests on one assumption: that autonomy is the product. Sometimes it is. First drafts, research, options for a campaign — work where an error costs a redo, variance is cheap, and a system that improvises can earn its keep.

Now take a board pack. A quarterly review. Anything a CFO signs. Variance there is not the product — it is the defect. An agent that makes a decision nobody approved is not impressive in that setting. It is unauditable. The two words describe the same behaviour; the work decides which one applies.

The systems we build get their accuracy from constrained execution, not from cleverness. Every number in the output can be traced back through a gate a person approved — this source, this calculation, this check, signed off before the system ran and inspectable after it. When a gate stops a run because a number will not trace, nothing about the moment looks impressive. That unimpressive stop is the system working — the failure surfacing at the gate, visibly, instead of in the pack, quietly. What an executive team is actually buying is a system that can be read and repeated: one that someone who was not in the room when it was built — a new finance director, an auditor — can open, follow and trust. Improvisation is precisely the thing that cannot be handed over.

Two honest products — and one dishonest price

Which means there are two honest products on this market. One is genuine autonomy, sold where variance is affordable and improvisation adds value. The other is the constrained, gated system with reasoning inside the gates, sold where variance is expensive and every output carries a signature.

The dishonesty is not in building the second product. It is in labelling it as the first and charging the autonomy premium. The washing is in the claim, not the workflow. Gartner’s own analysts make the quieter point that gets lost in the noise: many of the use cases positioned as agentic today do not require agentic implementations. The market’s problem is not too little autonomy — it is mislabelled constraint, sold at the wrong price.

We build the second product, and we call it what it is. We price against the outcome it delivers, not against an autonomy premium we could not demonstrate and would not want to.

The build order says the same thing

The advice circulating alongside the test is a build order: data first, then content, then the system of record, then workflows — and the agent last. Skip a step, its authors warn, and you join the 40%. The sequencing is sound. Ours is older and shorter: Simplify First, Automate Last. The agent is the last brick, never the foundation — and a vendor who wants to sell you the last brick first is answering a question about their pipeline, not your business.

The question worth asking instead

So keep the circulating test for what it is good at: catching a claim of autonomy that is not there. But for the work that carries your name, the harder question is a different one: “What is its failure rate in production — and show me the gates.”

That question splits the market into vendors with numbers and vendors with adjectives. Every run of every system we build is logged against a written standard — a ledger entry per run, self-graded, patches recorded, including the runs that stopped at a gate because a number would not trace. The ledger is inspectable: ask, and you can read it. A vendor who will show you where their system stopped is the vendor whose other claims are worth something.

We would fail the autonomy test and pass the audit. Given what is stapled to a board pack, you get to decide which test matters.

Which test has your current vendor passed — and which one did you ask? If the second question is the one you would rather be asking, you know where to find me.

The Assay is written by Rudi Prenzlin at Scalable Growth, a UK practice that gives owner-managed businesses a fractional CFO and the operational structure to run without every decision going through the owner.

← All notes Begin a conversation