For a decade, the operations conversation was about automation. Scripts, runbooks, pipelines, orchestration — take the repetitive work off people and hand it to the machine. That conversation has moved on. In 2026, the question in engineering and operations leadership meetings isn’t whether to automate a task; it’s whether an agent can own the workflow end to end. From automation to autonomy, as the vendors like to put it.
The technology genuinely has moved. The thing that decides whether it pays off has not.
The gap isn’t where you think it is
There’s an anxiety running through most engineering and operations teams right now: everyone feels a step behind on the AI itself. Which model, which framework, which agent platform — it moves weekly, and almost nobody feels on top of it.
Here’s the reframe, and it’s the whole point of this piece. The part that decides whether an AI investment delivers isn’t the part most teams are behind on. It’s the operational groundwork underneath it — the processes, the discipline, the data. And that part hasn’t changed. It’s the same work that separated the automation programmes that delivered from the ones that quietly stalled. An organisation that runs its operations well already has most of what makes AI work. The gap is smaller than it looks, and it sits on familiar ground.
The part that decides whether your AI investment delivers isn’t the part you’re behind on. It’s the groundwork underneath it — and that hasn’t changed.
Automation amplifies whatever it touches
Start with the oldest lesson in automation, because it’s about to be relearned the expensive way. Automation amplifies whatever it touches. Point it at a clean, well-understood process and it makes that process faster and cheaper. Point it at a broken one and it makes the mess faster and cheaper too. Most automation programmes that disappointed didn’t fail on the technology — they automated a process nobody had fixed first, and got more of the same, sooner.
AI raises the stakes, and it does it in a specific way worth being precise about.
Automation makes a good process faster and cheaper — and a broken one faster and cheaper too. Point it at a mess and you get more mess, sooner.
Loud failure is safe failure
Deterministic automation fails loudly. A script hits a case it wasn’t written for and it stops, throws an error, pages someone. The failure is obvious — and obvious failure is safe failure, because you know to go and look.
AI fails plausibly. Give an agent a task it isn’t equipped for, or data that’s wrong, and it doesn’t stop. It produces a confident, reasonable-looking answer that happens to be wrong — and then, if it’s an agent rather than a chatbot, it acts on it. It updates the record, closes the ticket, reroutes the change, all at machine speed, all looking entirely correct until someone downstream discovers it wasn’t. The failure isn’t visible at the moment it happens.
That’s the shift from the automation era that operations teams haven’t fully priced in.
Deterministic automation fails loudly — it stops and pages someone. AI fails plausibly — it acts on a confident wrong answer at machine speed, and looks right until it isn’t.
The market has caught up
This isn’t a contrarian take any more — it’s the consensus catching up to what the analysts flagged early. Generative AI has slid into what Gartner calls the trough of disillusionment, and agentic AI — still riding the peak of inflated expectations, with only around one in six organisations having actually deployed agents — is heading the same way. The abandonment that was predicted is now being counted: reported figures have organisations walking away from the bulk of their AI initiatives at roughly triple the rate of a year earlier. And the specific failure this piece is about has started showing up in the wild — agents approving things they shouldn’t in procurement workflows, acting against live systems with nobody watching. Plausible failure, in production.
None of this is new information. Back in early 2025, Gartner warned that most AI projects unsupported by AI-ready data would be abandoned, and that more than 40% of agentic AI projects would be cancelled by the end of 2027 — pointing not at weak models but at escalating cost, unclear business value and inadequate controls. Management and governance problems, not intelligence problems. Those warnings have simply aged well.
Read that list of causes again, because it’s a list of operations problems. Undefined value. Weak governance. Data the system can’t trust. None of it is the AI. All of it is the groundwork.
What “ready” actually means
Three things, none of them glamorous, all of them familiar.
Pick candidates on value and readiness, not value alone. The instinct is to rank opportunities by the size of the prize. But the prize is irrelevant if the process underneath isn’t stable enough to hand over, or the data isn’t good enough to act on. A high-value candidate sitting on a broken foundation isn’t a quick win — it’s the project that gets cancelled in eighteen months. Score both, and start where value and readiness meet.
Give the agent a foundation it can trust. This is the one that gets underestimated: an agent acts on what it retrieves. If it reasons about your estate from a configuration picture that’s incomplete or wrong — the CMDB half the organisation already doesn’t trust — it will make confident decisions on bad information, and it will make them faster than any human ever made the same mistake. The data foundation isn’t a prerequisite you can defer until after the pilot. It is the pilot’s chance of survival. We’ve made the case for fixing that foundation on its own merits — long before AI entered the room — in Your Platform Doesn’t Know Itself.
Prove it under supervision before you let it run alone. Every worthwhile automation earned its autonomy — it ran alongside a human, on real workloads, until it had demonstrated it could be trusted. AI is no different; only the consequences of skipping the step are larger. Keep the blast radius small and reversible. Let the first real workloads run watched. Grant autonomy as it’s earned, not as it’s hoped for.
Where to start
Two questions worth sitting with honestly.
If you automated your highest-value candidate tomorrow, would you be automating a process you’ve fixed — or one you’ve simply learned to live with? The answer tells you whether you’re about to buy a genuine improvement or a faster version of today’s problem.
If an agent made a wrong call on your estate, how long before anyone noticed? If the honest answer is “when something downstream broke,” you don’t yet have the visibility to let AI act autonomously — and that’s a foundation gap, not an AI gap.
Neither question is about how good the AI is. Both are about whether the ground underneath it will hold. That’s the part you can control, and it’s the part that decides the outcome.
The AI isn’t the hard part — and it isn’t where the value or the risk lives. The foundation underneath it is both.
If you’re weighing where to point automation or AI first, that’s exactly what our Automation Value & Readiness Assessment is built to answer — a focused 3–5 day engagement that scores your automation and AI candidates on value and readiness in equal measure, with a quadrant-mapped backlog, a recommended first wave, and the readiness gaps to close before you build.
If this conversation resonates, it’s worth a chat.
Found this useful?
Share it on LinkedIn, or get in touch if you’d like to talk through whether your foundations are ready for what you’re about to put on top of them — that’s what we’re here for.