EON AI Ventures has read forty years of official accident reports — 21,386 of them — and found that the safety equipment was usually working properly when things went wrong. The company is now building AI to catch what everyone else is missing: not the machine breaking, but the person forgetting.
IRVINE, CA – August 11, 2026 — Picture a routine job at a refinery. A technician has to replace a seal on a pump. It takes twenty-six steps and most of a morning.
At step twenty-three, the pump goes back into service. The job is done. The pump is running.
Except there are three steps left. Put the guard back over the coupling. Check the earthing strap is reconnected. Close out the permit.
Those last three are the ones that get forgotten. Not because the technician is careless — because the human brain files a job as finished when the goal is achieved, and everything after that point tends to evaporate. It is the most thoroughly studied mistake in fifty years of safety research, and it has a name: the post-completion error. Read more about it in detail in our The Alarm Was On white paper.
No sensor anywhere in that refinery can see it coming. And that, EON AI Ventures argues, is where the next decade of industrial safety has to go.
Click on the image below to access the EON Delta:Why, What, and How presentation.
| The point, in one paragraph
Every company selling industrial AI today predicts the equipment: the bearing wearing out, the pipe getting thinner. That is a real problem and it is well served. But a large share of serious industrial accidents do not begin with a machine failing. They begin with a person missing a step, on a day when a safety barrier happened to be switched off. Nobody predicts that — because to predict it you have to be standing next to the person while they work, holding something that understands both the job and the equipment. EON already puts that device in workers’ hands, for a completely different reason: training and guidance. The prediction is what you get almost for free once it is there. |
Why — what forty years of accident reports actually say
Every serious pipeline incident in the United States has to be reported to the federal regulator on a standard form. EON AI Ventures read all of them — 21,386 incidents going back to 1986 — and asked an unusual question.
Not what broke. Everyone studies that. Instead: what was the safety equipment doing at the time?
The answer was uncomfortable. It was mostly working.
| What was supposed to protect them | What actually happened | In plain English |
|---|---|---|
| The control room watches the pipeline around the clock | It was the thing that spotted the problem in 24.3% of cases where it was installed and switched on | Three times out of four, a human being noticed first — often a member of the public |
| The pipe gets inspected with specialist tools | 63.4% of the failures happened within two years of an inspection. A quarter failed in the same year they were inspected | The pipe had been checked. Recently. By experts. It broke anyway |
| Diggers must phone before excavating near a line | 63.3% of the times a line was hit by a digger, the digger HAD phoned first | They followed the rules. The line was marked. It got hit regardless |
| Monitoring should catch slow, predictable decay | Corrosion — the slowest and most predictable way a pipe dies — is the least detected cause of all, at 13.0% | The problem you would most expect to see coming is the one they see least |
| What this adds up to
This is not a story about missing sensors. The pipe was inspected. The ticket was raised. The alarm system was live. It is a story about four correct pieces of information sitting in four different systems, owned by four different suppliers, that nobody ever put on the same desk and asked: taken together, what does this mean, and when does it matter? |
A separate 2026 academic study, working from the same government records but looking only at carbon-dioxide pipelines, found a control-room detection rate of 19.7%. EON’s figure of 24.3% covers all pipeline types over the full forty years. The two agree.
And there is not one problem here. There are two.
Sorting the incidents by how old the equipment was when it failed splits them cleanly into two groups that behave nothing like each other.
| Kind of failure | Typical age when it failed | What that tells you |
|---|---|---|
| The pipe or the weld gave way | 51 years | Slow. Physical. You could see it coming years ahead if you were looking properly |
| Corrosion ate through it | 43 years | Same — slow and in principle predictable |
| Equipment failed | 8 years | Fast. Days or weeks of warning at best |
| Somebody operated it wrongly | 7 years | Fast. Often no warning at all in any instrument |
One is a slow puncture. The other is a front door left unlocked. Both lose you the house, but you cannot watch for them the same way — and almost every product on the market is built to watch for the first one only.
What EON AI Ventures is building
EON Delta is a set of small, specialised AI models trained on the documented record of real industrial accidents rather than on general text scraped from the internet. The V8 package published today sets out the whole design.
It treats industrial risk as three separate problems, because they are:
| The problem | An everyday example | Who solves it today |
|---|---|---|
| Metal wearing out | A bearing grinding itself down over eighteen months | A crowded, mature industry. EON uses their output rather than competing with it |
| A person missing a step | Forgetting to refit the guard after the pump goes back on | Nobody. There is no product on the market that does this |
| Safety defences quietly rotting | An alarm switched off for a shutdown and never switched back on | It gets written in a register. Nobody connects that register to the job somebody starts tomorrow morning |
The second and third are where EON is unusual, and the third one is the killer: most serious accidents are not one of these things. They are two or three of them landing on the same shift.
The part that works on day one, with no data from anybody
Here is the piece that can be shown to a customer immediately, and it surprises people.
Because we know which kinds of steps get forgotten, we can read a written work procedure and predict where it will go wrong — before a single person has performed it.
| Which kinds of steps get forgotten
The ones after the job already feels finished — refitting a guard, closing a drain, removing a blank. The ones where nothing tells you that you did it. No click, no light, no reading changes. The ones that undo something you deliberately did earlier. You remember switching it off. You do not remember switching it back. The ones straight after an interruption. People restart one step further on than where they actually stopped. The ones that are different on this machine than on the identical machine next to it. People do what they usually do. The ones you have got away with skipping forty times before. |
None of that needs a sensor, an accident history, or a single byte of a customer’s operating data. It needs the procedure document — which EON already receives from customers in order to build their training simulations.
In practice it means that any company can hand EON a work procedure and get back, within days, a marked-up version showing which steps their people are most likely to miss and what to change. That is useful on its own, before any AI model is involved at all.
Who this is for, and what each of them gets
| Who | What they get out of it |
|---|---|
| The technician doing the job | Before starting: the one step people most often miss on this procedure, and the single check that most reduces risk today. Shown visually, on the actual equipment, not buried in a document |
| The maintenance supervisor | A procedure marked up with its weak points before the crew is sent out, rather than a report explaining what went wrong afterwards |
| The plant manager | Fewer unplanned shutdowns. The failures in this analysis were not exotic — they were routine equipment on which every warning system was working |
| The safety director | A method that can be defended. Every warning traces back to published, regulator-recognised research applied to observed conditions — not to an AI that cannot explain itself |
| The board and the insurer | A written record of what was predicted, what was recommended, what the worker did, and what happened next. Today no such record exists anywhere in the industry |
| The whole company, quietly | The judgement of experienced people gets captured while they are still there. Most operators are losing that generation now and have no way to keep what they knew |
How it works
Four steps, and the fourth is the one that makes it improve over time.
| Step | What happens | |
|---|---|---|
| 1 | Read the procedure | Before anyone goes out, the written procedure is scored for the steps most likely to be skipped, and each one is tied to the actual piece of equipment it concerns |
| 2 | Add the context | How much time the job has been given, how complex it is, how experienced the assigned person is, whether they are at the end of a run of night shifts, and whether any safety barrier on that equipment is currently out of service |
| 3 | Say it at the right moment | Not a prediction of a failure date. A short statement at the start of the shift: for this job, these are the credible ways it goes wrong, this barrier is degraded, this is the step most often missed, and this is the one check worth doing today |
| 4 | Watch what happens | The same device that gave the warning sees what the worker did, what they found, and what they chose to leave until later — and why. The system therefore marks its own homework, every shift |
| Why step four matters more than the rest
Here is a problem hiding inside every predictive maintenance system on the market. In a working plant, most warning signs never turn into accidents — because somebody quietly fixed it. Nobody records that they did. So the data says: a problem appeared, then it went away, and nothing happened. Train an AI on that and it learns, very confidently, that problems sort themselves out. It learns to keep quiet. And it will pass every test, because the test data has the same hole in it. EON’s device is there at the moment somebody intervenes — not afterwards, when a job card gets closed with three words. Recording what was done and what was deliberately left undone is what turns that history from misleading into useful. EON has filed a patent on it. |
Why nobody else is doing this
It is a fair question, given the size of the industry. The answer is not that others have missed it. It is that they cannot reach it from where they stand.
| The kind of company | What they have | What they are missing |
|---|---|---|
| Sensor and monitoring companies | Excellent data about the machine | Nobody on site. They cannot see a person about to skip a step |
| Digital checklist companies | A tablet in the worker’s hand | No model of the equipment and no AI. They record that step 4 was ticked, not that the technician thought something sounded wrong |
| Inspection contractors | The best measurement of the pipe that exists | They tell you the condition today. Not the date it starts to matter |
| General AI assistants | Fluent language | No knowledge of your specific plant. Ask about your pump and you get a confident, plausible, unverifiable answer |
| EON AI Ventures | A device already in the worker’s hand that knows the job AND the equipment, plus the accident record and a physics simulator | A published performance result. See below — we do not have one yet |
EON’s claim is deliberately narrow, and it is not that its AI is cleverer than anyone else’s. It is that the missing information can only be created by something already standing next to the work, already trusted, already being used for another purpose. A sensor company has no person there. A checklist company has no model. EON has both, deployed today.
And a customer’s own information never leaves their building
This normally stops these conversations dead, so it is worth being direct.
| What it learns from | Who owns it | |
|---|---|---|
| Your model | Only your own material, inside your own systems. It never leaves your building. | You do. Deleted on request, and the deletion can be verified |
| The shared model | Only patterns that have shown up independently at several unrelated companies — and therefore belong to none of them | EON, and it is licensed back to every participant free, permanently, even if they stop taking part |
Two companies running identical equipment cannot hand each other their operating records — commercially or, in many countries, legally. But both can benefit from a lesson that has proved general and points at nobody. That is the entire reason for the split.
What EON AI Ventures is not claiming
| There is no accuracy figure in this announcement, and that is deliberate.
Nothing has been trained. Nothing has been measured. EON is not stating a detection rate, a false-alarm rate, or a comparison against any existing product, because no such number exists yet. The findings in the first half of this release describe the pipeline industry as it operates today. They come from public government records and have nothing to do with whether EON’s system works. They are not evidence that it does. |
EON has published the test it intends to be judged by, before having a result to report. It is called the Shift Test: it gives a system the information available at a shift handover and scores what it does with it. The rules are public and anyone may run it, including against EON.
EON has not run it yet.
Check the analysis yourself
Every figure in the first half of this release comes from public United States government files. EON has released the software that produces them; it takes about four minutes to run.
Operators, regulators, universities and competing suppliers are invited to reproduce the analysis and to publish anything they find that differs.
| One figure that is easy to report wrongly
The finding that 63.4% of failures happened within two years of an inspection is about pipes that failed. It shows that a recent inspection did not save them. It is NOT a measure of how well inspection works overall, and should not be written up as one. To know that, you would need to know how many inspected pipes were fine — and nobody publishes that number anywhere in the world. |
Read more about it in detail in our The Alarm Was On white paper.
Learn more by tuning to our podcast.
About EON AI Ventures
EON AI Ventures is the company behind Work Intelligence — the captured, verified, and compounding knowledge of how expert work is actually done. Its Intelligence Flywheel platform (Genesis, Field IQ, Assess IQ) enables industrial enterprises to encode expert procedures into AI-guided simulations, deliver them to any worker on any device, and verify competency in the field. EON AI Ventures builds on a 25-year foundation of immersive learning technology deployed across more than 80 countries. For more information, visit www.eonaiventures.com