Understanding the AI
What the AI is good at — and what it isn't
CIM uses AI to read your drawings and specs and surface findings. It's powerful, but it's not magic. Reading this once will save you from the most common trust issues. What it's good at - Reading text on drawings and specs. Title blocks, notes, schedules, callouts. High accuracy. - Comparing two documents for consistency. Architectural vs structural column grids, drawings vs spec, revision A vs revision B. - Finding things that aren't there. Missing notes, absent details, undocumented assumptions. - Volume work. Reading 400 sheets in an hour and surfacing the 30 that need a human. What it's not (yet) good at - Counting. Car spaces, GPOs, doors, fixtures. Don't trust quantities. Use it as a starting point and recount. - Measurements. Corridor widths, room areas, setback distances. Same rule — verify on the drawing. - Implicit cross-document inference. If the answer requires combining info from three documents the AI wasn't told to look at, it'll often miss it. - Knowing what should be on a drawing. It can compare what's there to what you told it to check for, but it doesn't have construction intuition. Expect false positives. Roughly 1 in 4 findings will be wrong. This is normal and we tell every customer upfront. The right mental model: the AI is doing the first pass, you're doing the judgement call. Always click the citation to verify before you action a finding. The single most common failure mode: the AI looked at the wrong drawing.
Why counting and measurements are unreliable
This is the most common question we get. Short version: if you need a number, count it yourself. Why it's hard. Counting on a drawing requires (a) recognising every instance of the thing, (b) not double-counting where details repeat across sheets, and (c) knowing which sheets to count from in the first place. AI vision models are improving fast but they currently fall down on at least one of these on most jobs. What goes wrong in practice - Counts that include duplicates. A GPO shown on both an RCP and a unit plan gets counted twice. - Counts that miss elements behind callouts, hatches, or off the visible sheet. - Measurements that read the wrong dimension string. The AI picks up a nominal dimension instead of a clear one. - Confidence that doesn't match accuracy. The AI can sound certain about a wrong number. What you can use the AI for instead - Estimating order of magnitude. "Roughly how many apartments are in this set?" is a fair question. - Sanity-checking your own count. Ask the AI after you've counted, not before. - Finding the relevant sheets to count from. It's reliable for navigation. Roadmap. We're working on a dedicated quantity-extraction pipeline that separates counting from generative AI. When it ships, this article will be updated.