Most of Your Problems Do Not Need an AI Agent
There is a cartoon meme doing the rounds this week. Two doors on the same building. One is labeled “Automation: reliability, ROI, scalability” and nobody is standing in front of it. The other is labeled “AI Agent: hallucinations, cost, hype” and the queue wraps around the block.
It is a joke, but the queue is real. Gartner predicted last year that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and weak risk controls. The same note estimated that of the thousands of vendors selling “agents,” only around 130 are the real thing. The rest are rebranding chatbots and RPA. The widely quoted MIT figure that 95 percent of enterprise AI pilots showed no measurable return has its methodological critics, but the direction matches what CIOs keep reporting: the controls needed to run an agent safely often wipe out the projected savings.
Why the queue forms at the wrong door
Nobody gets a conference talk for a cron job. Automation is a solved, boring category: if-then rules, API calls, scheduled jobs, a script that moves rows from one system to another. It does not demo well and it does not sound like the frontier. So the board asks “what is our agent strategy” instead of “which of our processes waste the most hours,” and the queue forms.
The irony is that the boring door is where the ROI has always lived. Automation costs are mostly one-time and predictable. It fails loudly, so you know when it breaks. It does the same thing on run one thousand as on run one. Agents invert every one of those properties: per-run token costs, quiet and plausible failures, and output that varies with the phrasing of the input.
The thirty-second sorting test
Before anyone in your company builds or buys an agent for a task, ask three questions about the task itself.
Is it the same steps every time? If yes, automate it. Can you write the rules down clearly? If yes, automate it. Is the input structured, a database, a CSV, an API? If yes, automate it.
An agent earns consideration only when the honest answer to all three is no: the input is messy, the path depends on judgment, and the exceptions are the job. Reading an unfamiliar supplier document, diagnosing why an order is stuck, drafting a response to an unusual query. We walked through where that line falls inside an ERP in Agentic ERP vs RPA, and the answer is usually that reasoning belongs at one or two steps of a flow, not across it.
What the successful minority does
The deployments that survive contact with production share a pattern: a deterministic backbone of plain automation doing ninety percent of the work, with an agent placed at the specific judgment step, wrapped in permissions and an approval gate. Two or three use cases with a named owner, not thirty pilots. That is the same discipline we argued for in picking one workflow, and it is the opposite of joining the queue.
For an SME the practical move this quarter is unglamorous: list the ten tasks that eat the most hours, run the sorting test on each, and automate the ones that fail it before spending a rupee on agents. If you want a second pair of eyes on that list, our free assessment is exactly that exercise.