The Plant Changed. The Drawing Did Not.

Most brownfield facilities operate on drawings that quietly drifted away from reality. Today, AI makes it practical to bring them back in sync.

Two days lost to a line that was not on the drawing

Break a flange on a six-inch line, fit a new spool, and get back online by the end of the shift.

The crew climbed up with the P&ID in hand. The flange was there, near enough. So was another line, running straight through the space where the new spool had to go.Nobody on the permit knew what it was. Work stopped. Two days later, an engineer found the drawing in an old revamp folder. The line had been added years earlier. Nobody had ever updated the drawing.

We sat down after the unit was back up. The maintenance manager did not make excuses. “Everyone here knows the drawings are wrong,” he said. “We just don’t know which bits.”

That one sentence captures the whole problem. A drawing that is clearly wrong gets ignored, so it does no harm. A drawing that is mostly right gets believed. And it lets you down at the worst possible moment: during an isolation, on a hot work permit, or in a shutdown with a thousand contractors on-site and the clock running.

Nobody intended the problem, and that is precisely why it persists.

He walked me through how it happens, and nobody in the story is careless. A change gets approved. The pipe goes in over a weekend. The unit starts up. Updating the drawing drops to the bottom of the priority list, and one day it falls off.

Do that for 20 years, through two owners, a control system upgrade, and a few debottlenecking jobs. The plant on paper slowly drifts away from the plant in the field.

When I ask most refineries or chemical sites for as-built drawings, what I often get are as-designed drawings with years of quiet change sitting on top.

None of this shows up in the budget as bad drawings. It hides somewhere else:

• Turnarounds that overrun
• Permits that have to be reissued
• Scaffolding built where nothing needed fixing
• Engineers walking a unit that somebody else already walked the last year
• And the near miss nobody puts in the safety newsletter, where an isolation boundary was drawn from a sheet that no longer matched the steel

So, what would you actually do about it?

For years, the only real answer was a walkdown. A team, a clipboard, a camera, six weeks, and a result that was out of date the day it was handed over.

That answer has changed for three reasons.

Reality capture has become cost-effective enough to document an entire process unit in a single afternoon.

Modern computer vision models can now read legacy scanned drawings and extract symbols, tags, line numbers, and connectivity. Combined with field-capture data, these models can automatically identify discrepancies between what is documented and what physically exists.

With both in hand, AI can lay one over the other and show where they disagree. That last step is what changes the job.

No single way of capturing a plant is good enough on its own. Each method has blind spots where another one sees clearly. That is why I see the strongest teams use several methods together and let AI stitch the results into one picture.

The result is more than a better drawing. It creates the foundation for a Plant digital twin that can connect what we see in the field with engineering information already held in the enterprise.

There is no single product that solves this problem.

“So which product do I buy?” the maintenance manager  asked.

None of them, and all of them.

There is a good tool for each part of the job. The value comes from how they work together.

The workflow layer

• SAMP.AI: Reads 3D scans against P&IDs, flags equipment found in the field but missing from the drawings, and queues suggestions for an engineer to confirm or correct.
• Plant360.AI: Turns folders of PDFs and datasheets into an active engineering data graph that you can query.
• Veerum: Brings drone, LiDAR, and CAD data into a single visual workspace for frequent, accessible site checks.
• Kongsberg’s Kognitwin: Ties the digital model directly to live operating data, turning a static 3D representation into a living tool.

The AI & infrastructure layer

Underneath these applications sits the AI layer that makes this automation possible. NVIDIA is a clear example of the current capability.

• NeMo Retriever: Uses advanced optical character recognition (OCR) to read the layout and structure of faded 1980s P&IDs, not just the letters.
• Omniverse NuRec: Rebuilds ordinary phone-camera and LiDAR passes into fully walkable 3D scenes.
• OpenUSD: Does the decisive job of holding the scan, CAD model, and tag data in one unified coordinate system so they can finally agree.
• Cosmos and NIM Microservices: Build training simulations and serve vision-language models that learn a site’s specific tag conventions.

The Open-Source foundation

Reconstruction and point clouds: Free tools such as COLMAP, Open3D, PDAL, and CloudCompare do the heavy lifting for reconstruction and rendering.

Data standards: Frameworks such as CFIHOS and ISO 15926 help prevent every new contractor from inventing a different naming scheme.

None of these is the answer on its own. The answer is the loop they create together: capture, read, compare, correct, and push the correction back into the systems of record.

That way, the next crew inherits the truth.

AI does the reading and the matching, the tedious work that no person should spend a career doing. People do the judging, the part no machine should be trusted to own. Where that split works, a scan taken on Tuesday can produce a list of differences by Thursday. The drawing office can then spend its week approving changes instead of hunting for them.

Pick one unit, and keep score

He asked where to start. Not with the whole site. I would take one unit, ideally one going into a turnaround next year. I would capture it properly, run the old drawings through extraction, and get a list of what does not match.

That is where P&ID digitization becomes practical. Instead of treating legacy drawings as files to archive, I can turn their information into something that can be compared with what exists in the field.

Then I would put my engineers through it becauseAI can propose changes, but engineering judgment remains responsible for validation and approval..

Next I would measure the results using numbers the turnaround team already argues about:

• How many tags in the field match the register, before and after
• Turnaround hours lost to wrong drawings, discovered work, and reissued permits
• Walkdown and engineering hours for each modification
• Isolation plans redone because the field did not match the sheet
• Days between making a change and having that change reflected in the drawing

That last one is the measure that matters.

In closing

A plant that cannot describe itself is carrying a risk it has never priced. The old habit was to throw people at the problem every few years and live with the drift in between. I see a better way now, and it does not start with spending more. It starts with choosing to Outcreate that habit: capture cheaply, let AI do the reading and the comparing, keep engineers on the judgement, and close the loop so a change reaches the drawing in days instead of never.

I saw that manager again a year later. His team now scans a unit before every major job. He put it plainly: “We stopped arguing about what is out there.” The plant will keep changing. The drawings can finally keep up.