We are withdrawing two claims this post used to make. The first said that 88% of AI agents never reach production and that missing observability is the main reason. The second came later: it read an empty line in a search report as proof that buyers have their priorities backwards.
The first rested on an aggregator's compilation of enterprise surveys, together with a 68-point "gap" between teams piloting agents and teams running them. That gap was a subtraction between two figures, and we cannot show that they shared a definition of "production". A difference between numbers that mean different things measures nothing, so the gap goes. You will find figures in circulation saying that something like 88% of AI agents never reach production. Aggregator posts repackage each other's surveys, the definitions of "production" differ, and we would not build a budget on any of them. The more defensible public number is Gartner's forecast that over 40% of agentic AI projects will be canceled, demoted or decommissioned by 2027, widely reported (The AI Insider), and it is still a forecast.
The second claim was ours alone. Google Ads data pulled through DataForSEO on 11 October 2026 (US, English) shows about 5,400 searches a month for "prompt injection", up from 3,600 in September 2025 after a spike to 9,900 in August 2026. For "ai agent observability" it shows nothing. We wrote that people search for the attack they fear and not for the instrument that would reveal it. That covers one search engine in one country, and it ignores how a buyer asks. Someone worried about this may type "audit trail" and never learn our vocabulary. An empty line means we could not measure demand under that name, which is not an argument.
What survives
The mechanism does not need either statistic. You cannot investigate a wrong refund, a quietly mispriced quote or a prompt injection without a record of what the agent saw and did, and a team that cannot see why an agent acted cannot fix it, defend it to an auditor or decide whether to trust it with more.
A service that returns 200 did what it was asked. An agent that returns a clean answer may have called the wrong tool, read a stale record or built a paragraph on a number it invented, while uptime stays green. So the unit to observe is the decision: what the agent was shown, what it chose, what came back, and what it did next.
The test we would use is a person and not a dashboard. Take one action from last Tuesday, a drafted purchase order or a closed ticket, and ask a finance controller or a support lead to reconstruct why it happened. If they need an engineer to read the trace aloud, you have logging and not observability. A record that passes holds the inputs the agent saw, each tool call with its arguments and response, the outputs, and the identity under which every action ran, in order, in plain fields, exportable.
Here is the part tooling comparisons skip. A full trace contains the customer's name, the invoice, the contract clause and the prompt. Your tracing store is one of the most sensitive databases in the company, built automatically and often shipped to a third-party service by default. Give it the access rules of the most sensitive thing inside it, and set a retention period before the first trace.
Sampling suits a chatbot with thousands of conversations a day. For an agent drafting forty purchase orders against a real ledger, storage is cheap and a missing trace is expensive, so record every run. Cheap deterministic checks come before any model-based grading: does the order total equal its lines, does the vendor exist, is the amount inside the approval limit.
The test for your own agents is a question: would you know if it got something wrong last week? An agent connected to a clean system of record, such as an Odoo or ERPNext instance with sensible permissions, leaves actions you can trace to a record, and one stitched to spreadsheets leaves gaps. We cover the neighbouring failures in why agentic AI projects fail and the pre-launch discipline in AI agent evals. If the answer is no, get in touch before it handles more.
