AGENTIC AI
When an AI agent acts on its own, a wrong call turns into a wrong action, and industrial buyers stall the purchase until someone answers for it. Thing Company scores the four questions behind those decisions.
Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. In S&P Global's 2025 survey, 42% of companies were abandoning most of their AI initiatives, up from 17% in the prior year's survey.
Neither finding says the technology stopped working. Both describe commercial hypotheses about value, budget, and governance that nobody validated. Earlier technology waves failed the same way. This time the system acts, so the cost of a wrong bet is higher.
Each dimension is scored on primary evidence from the economic buyer, the operations team, and the risk owner, instead of the innovation group that ran the demo.
Will this buyer let the system act at this level, in this facility? Trust varies by decision and does not carry over from one site to another. We score it 1-5 across discovery interviews, and a strong signal needs an average of 4.0 or above. Below that, the viable product keeps a human in the loop, and the pricing changes with it.
When the agent acts, who approved it, who can stop it, and what record survives? The risk committee wants approval gates, rollback paths, audit trails, and a named owner, and without them it will not approve, whatever the pilot showed. Score the buyer's controls as well as the vendor's.
Most agentic pilots start without a pass/fail bar, which leaves the capital request with nothing to cite. "The team found it useful" does not qualify. A criterion names the metric, how it is measured, the period, the threshold, and who verifies it, and it is declared before the pilot starts.
Agentic spend rarely has a budget line of its own, and a cloud commitment only absorbs it when the purchase runs through that cloud's marketplace. Someone has to build the approval pathway and own the return on investment (ROI) number. If the vendor's forecast is the only number, the request dies in the finance review. The Sprint confirms the budget, the owner, and the number.
The operations leader and the risk owner have each confirmed the autonomy level for this use case. The approval controls are named and available, and the budget owner has confirmed the pathway. All of it comes from scored interviews.
A successful pilot in assistive mode, vendor benchmarks, and analyst projections are all real evidence. None of them confirms that this buyer will accept this system acting alone in this facility, and an assistive pilot is often misread as acceptance of autonomy.
Business cases still rest on claims like these: the plant will accept full autonomy once it sees the accuracy numbers; governance can be worked out during deployment; the productivity gain is obvious enough that finance will approve it. Each is an Assumption, and each has a validation path.
Thing Company uses agents in its own research and outreach, with a person approving anything that leaves the building.
That working knowledge is why the scoring asks what it asks: where governance slows a team, where the chain of accountability breaks, and the gap between comfort with an assistant and acceptance of autonomy when an interviewee describes one. The four dimensions are set out in the paper.
It also makes us useful after a Proceed verdict, and we keep that work separate from the verdict: what happens after a Proceed.
The reasoning is published: Governing agents in production: what to ask before an agent acts, and the Field Notes take single questions, such as where to put the human in an agent's work and why one autonomy switch is not enough. Related papers cover agentic AI at the industrial edge and the turnkey autonomous edge platform. Robots, machine vision and autonomous machines are scored on the physical AI page. The specimen brief shows what the scoring produces.
It is testing, with the people who would buy and approve it, whether an agentic AI system will be accepted and paid for. Thing Company scores four questions on primary evidence from the economic buyer, the operations team and the risk owner: autonomy tolerance, governance readiness, a falsifiable success criterion, and budget architecture and ROI ownership. Model accuracy is not one of them.
Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Those are commercial hypotheses about value, budget and governance that nobody validated, not evidence that the technology stopped working.
Interviews score it 1 to 5, and a strong signal needs an average of 4.0 or above. Below that, the viable product keeps a human in the loop, and the pricing changes with it.
One that names the metric, how it is measured, the period, the threshold and who verifies it, declared before the pilot starts. "The team found it useful" does not qualify.
Yes. It runs its own research, outreach, content and software delivery on agentic systems built in-house, with tiered autonomy gates, human approval on outward-facing actions and audit logging. The dated field notes describe what those systems do in practice.
One conversation to find out which of the four dimensions is still an assumption.
Or reach us directly at hello@thing.company