Finance, not engineering, is switching off the agents: 59% report a kill or a delay

The unattributable bill has a mechanism, and the mechanism explains why the invoice turns up without a breakdown. One agent task is not one request. It emits a top-level trace, then several model calls, retrieval operations, tool calls, retries and loops, and when the agent delegates to another agent, the trace grows a fresh branch ^[13]. Every model call carries token, latency, cost and provider data, and every tool call writes its own records for arguments, results, status and downstream activity ^[14]. The identifiers that would let you split the bill by program, things like tool_name, agent_id and trace_id, are the high-cardinality fields that are hardest to aggregate and most expensive to index ^[15].

Now the part nobody in the survey write-up computed. Compound the reported spend growth for two years and the average observability bill lands near $5.19 million, about 1.64 times today's ^[18]. Over the same two years, respondents expect telemetry volume to grow 9.5 times on average ^[11], with 44% putting their own increase somewhere between 6X and 100X ^[12]. Those two curves only meet if the cost of carrying a unit of telemetry falls to roughly a sixth of its current level, an improvement of about 5.8X in unit economics ^[19]. That is not a number you get out of a retention setting or a renewal negotiation.

Two caveats belong on the 9.5X. It is a forecast that respondents made about themselves, and the survey was commissioned by Apica ^[2], whose chief product and technology officer prescribes intervening earlier in the pipeline as the answer ^[5][17]. The measured figures are more useful anyway. Among the 54% who saw telemetry triple in a year, AI and ML workloads accounted for 43% of that growth ^[8], which works out to roughly 0.86 of the original baseline volume added by AI alone in twelve months ^[20]. Those workloads came close to doubling the estimate on their own, without any of the forecast multiples.

Mann's bank could not say what its AI programs cost, so it cancelled some of them, and he describes AI projects as cannibalising ordinary budgets ^[6]. That is the operative failure. Only 35% of enterprises claim widespread agentic deployment, and close to two-thirds describe themselves as only somewhat prepared to run those systems ^[16]. Meanwhile 83% rank AI observability as a top priority for the year ahead ^[10], which in the bank's case meant the observability survived and the agents did not.