Is it normal for AI-generated Revit models from laser scans to need this much cleanup?

You didn't set it up badly. What you're describing is exactly what these services produce right now, and honestly your result sounds better than average. I ran a similar warehouse job through one last spring and got comparable numbers, maybe 30-40% of elements needing a cleanup pass. The wall thickness thing is the big one, they love to snap everything to a standard width and let reality be damned. And anything tucked behind or crossing another element tends to get guessed rather than measured, which explains your pipes through the column. The heatmap is the other gotcha. It only flags deviation from the point cloud, so if the model has a clean fit in a sparse area where your registration was weak, it can look confident while being wrong. It also can't judge whether the right family got used, just whether geometry lines up. So manual spot checking is never optional, treat the heatmap as a hint not a QA signoff. As for your liability worry, you're right to take it seriously. Whoever issues the model owns its errors, the software vendor's terms will make that very clear if you read the fine print. So yes, the workflow most of us have landed on is exactly what you did, machine output as a starting draft, then humans redo anything that ends up in a clash model or a deliverable. For us that means walls, structure, and anything within a metre of MEP gets checked element by element. Ceiling voids and plant rooms get the full treatment, open floor areas we mostly trust.

One thing that helps, if the service lets you export the raw points alongside the Revit file, do the checking by overlaying sections through the cloud rather than eyeballing the model. Cuts through the point cloud expose bad wall thicknesses and pipe runs in seconds. Also check the level datum they used, I've had a job come back floating 20mm off and that error rippled through every element.

So no, not a setup problem. It's where the tech is at, and the vendors quietly know it, which is why they all advertise hours saved rather than accuracy guarantees.

I'd push back on the bit about open floor areas being mostly trustworthy. My experience is the opposite of that. Big empty spans are exactly where the automated stuff drifts, because there's less clutter for it to anchor against. On a job I did, the model had a mezzanine edge sitting nearly a metre from where it actually was, dead centre of an open floor, and nothing flagged it. The busy areas came out closer to right, prolly because the software had more to grab onto. I'd flip the checking order. Spend your manual time in the plain zones where you assume nothing can be wrong, that's where the silent errors live.

The tape measure is still the best QA tool you own. Before I hand over anything from these services I pick a dozen or so random dimensions and check them against my own field notes. On a small retail fit-out last year I sampled 15 door openings and 5 were off by 30mm or more. If a third of a random sample is wrong, that tells you everything about the rest of the file. Cheap insurance when you think about what one disputed dimension costs you in a client call.