Walk any 2026 trade floor and you will be told that artificial intelligence is transforming engineering. Walk back to your desk on Monday morning and the picture is quieter. The model still has to be built. The line list still has to be reconciled. The stress isometric still has to be checked against the code. So which is it — is AI actually changing how we engineer plant, or is most of what we are sold a coat of paint over the same software we already had?
The honest answer is: both, at the same time. AI is genuinely reshaping a handful of workflows in oil and gas. It is also being sprayed onto a great many products where it changes nothing except the marketing. The skill worth having is not enthusiasm or scepticism — it is the ability to tell which one you are looking at.
01What the numbers actually say
Start with the uncomfortable data, because it cuts through the noise. SimScale’s State of Engineering AI 2026 survey of engineering leaders found that adoption is accelerating fast — but only a minority of teams are turning that adoption into measurable impact. The difference, they report, is not ambition. It is whether AI is wired into the actual workflow or merely installed alongside it.
More bluntly still: MIT’s widely-cited NANDA report concluded that the large majority of enterprise AI pilots have not yet produced a return on investment. That figure gets quoted as evidence that “AI is failing.” It is not. It is evidence that most organisations bought the tool and skipped the hard part — redesigning the work around it. A licence is not an outcome.
A licence is not an outcome. The value was never in installing AI — it was in changing the work it touches.
02What cosmetic AI looks like
You already know the smell of it. A familiar tool gains a chat box that summarises a manual you could have searched yourself. A “smart” feature that produces a confident paragraph you then have to verify line by line — which takes longer than doing it from scratch. An “AI-powered” estimator whose answer you would never put your stamp on without rebuilding it manually anyway.
None of that is fraud, exactly. The technology underneath is real. But it sits beside the work instead of inside it, and it carries no accountability. The tell is always the same: when the AI is removed, your workflow is unchanged. That is cosmetic. It demos beautifully and survives contact with a deadline poorly.
| Question to ask | Cosmetic | Real |
|---|---|---|
| What happens if I remove the AI? | Nothing changes | The workflow breaks |
| Can I trace the answer? | “Trust the model” | Back to a clause, a formula, a datum |
| Who is accountable for the output? | Unclear | The engineer, with an audit trail |
| What problem does it solve? | “AI” in general | One specific, measurable task |
03Where it is genuinely happening
Now the other side of the ledger, because in our industry the real cases are not hypothetical. The clearest wins are not the glamorous “design it for me” promises — they are the unglamorous problems that always ate our time.
Pipeline integrity and asset twins
Integrity digital twins now fuse GIS data, smart-pigging and ultrasonic corrosion runs, soil data and CFD into a model that predicts corrosion rates and stress hotspots. That moves inspection from calendar-based to condition-based — fewer wasted campaigns, earlier warning before a failure becomes catastrophic. Hexagon reports operators using this kind of asset-risk analysis to surface seven-figure savings and cut maintenance costs by double digits. That is a changed workflow, not a badge.
Process plant twins
Connecting a first-principles simulation — your Aspen HYSYS model — to live sensor data produces a model that actually mirrors the columns and exchangers as they run, letting operators optimise for changing feedstock and schedule turnarounds around real catalyst degradation rather than the calendar.
Finding what already exists
The least glamorous and most valuable category. Engineers spend a startling share of their time searching — for the right spec, the right existing part, the decision someone made on the last project before they left the company. Plain-language and geometry-based search over drawings, P&IDs and PDM vaults attacks exactly that bottleneck. It does not invent; it retrieves. That is where a lot of quiet, real productivity is landing in 2026.
04The line a piping engineer must hold
Here is the principle that separates a tool I would build from one I would walk past — and it is the principle I built ComplianceIQ around. AI is excellent at extraction, classification and pattern. It is dangerous as the authority on a code clause.
A large language model will read a plot plan and tell you, plausibly and fluently, that a spacing looks compliant. It will be right most of the time and confidently wrong some of the time, and it cannot show you the clause it relied on. That is unacceptable when the verdict has to survive a regulator, a client OE register, and your own professional liability. SAES-B-055 Figure 1 is a fixed matrix. The minimum distance between a fired heater and a storage tank is not a matter of probability — it is a number, and either you meet it or you do not.
So the right architecture is hybrid. Let the AI do what it is good at: read the drawing, extract the equipment, classify the hazard, surface the context. Then hand the verdict to deterministic code that enforces the standard exactly, every time, and cites the clause it applied. The AI accelerates the engineer; the code carries the accountability; and the trace back to the standard is never broken. Get that division of labour right and the tool is genuinely useful. Get it wrong — let the model rule on the code — and you have built something that demos well and should never be trusted.
05So — happening, or cosmetic?
It is happening. Not everywhere, not the way the keynote slide claims, and not by simply buying a subscription. It is happening in the specific, measurable places where someone did the unglamorous work of rebuilding a workflow around a tool that earns its keep — pipeline integrity, predictive maintenance, retrieval, drawing review done properly. Everywhere else, the cosmetic version is also happening, and it is louder.
For the working piping engineer, the takeaway is not to fear AI and not to worship it. It is to stay the person who understands both the code and the limits of the tool. The engineer who knows exactly where a model helps and exactly where it must never be allowed to decide is not being replaced by this technology. They are the one who makes it safe to use.