AI steel takeoff software, without the demo gloss.

Every vendor demo shows beams being detected on a plan view. That isn't the bar. The bar is: elevations, plates, base plates, braces, miscellaneous steel. A product that detects 95% of beams on a plan and zero on an elevation produces a takeoff that misses half the job. The right question for a demo is simple: "show me a moment frame elevation with stiffener plates," then watch what happens.

Counting beams is the easy part. Capturing shape, size, length, stud counts, camber, copes, holes, and moment connections is the part that makes the BOM bid-ready. Ask each vendor to itemize what attributes they extract per member. The list is shorter than the marketing implies.

Excel export is table stakes. The real test is whether the output drops into Tekla, Strumis, Fabtrol, or E.J.E. without a manual reformat. Ask to see the export landing inside the platform you actually run. If they have to send you a CSV and you have to clean it up, the "integration" was a screenshot.

Most projects ship two to five drawing revisions before bid. A takeoff tool that doesn't diff revisions cleanly turns every revision into a re-run. Watch a vendor handle a real revision on a real job, not a manicured demo set. The friction shows up immediately.

"60% throughput improvement" sounds like a benchmark; in practice it is one customer's quote, often during onboarding. Ask for at least three named customers reporting outcomes after six months in production. Anything less is a pilot, and pilot data does not predict steady-state.

A $399/mo plan with a five-takeoff cap can be more expensive than a $9,950 perpetual seat if your shop bids 40 jobs a month. Ask how pricing scales: per seat, per takeoff, per project, per user, per shop. Beta-pricing rhetoric ("lock in current rates before prices increase") almost always means the next renewal is going to land higher.

If the AI takeoff product comes from the same company that ships your detailing platform, the roadmap follows detailing priorities, not estimating ones. Independence isn't an absolute virtue, but it is a signal worth knowing about. Lift is independent. Steel Genie is not.