I sat in on a meeting last month with the operations lead at a mid-sized civil engineering firm. She had just finished rolling out Copilot to 200 staff. The adoption dashboards looked decent. People were opening the tool. When I asked whether her leadership team would put an AI-generated number in a stamped drawing without a manual check, she went quiet for a second. Then she said the quiet part out loud: nobody on her team would.
That gap, between using a tool and trusting what it produces, is the actual story of AI in AEC right now. Unanet’s 2026 AEC Inspire Report found 75% of AEC firms now use AI, up roughly 20 percentage points year over year. The same report found only 29% of firms report high confidence in the data underlying their AI outputs. Roughly seven in ten firms are running a tool whose outputs they do not fully trust. That is not a rollout problem. It is a data confidence problem.
The adoption number is counting shadow use
Here is the part most people miss. The 75% adoption figure almost certainly includes individual and shadow use: engineers pasting site data into ChatGPT, project managers dragging cost models into Copilot on their own laptop. The formal, firm-level workflows, the ones with an owner and an audit trail, are a much smaller set. Most of the 75% is people opening a chat window, not a sanctioned process.
The 29% high-confidence number is not a coincidence. It sits close to the share of firms that have actually built a documented AI workflow. Firms with a sanctioned, documented AI workflow can name where the inputs came from, what model and version touched them, and which engineer reviewed the output. Firms where individuals “use AI” without that workflow cannot. Reviewers and regulators cannot see what the AI touched versus what an engineer verified. Confidence collapses at exactly that seam.
This is why the gap is a workflow and verification gap, not a tooling gap. Buying a better model will not close it. The model is already good enough. What is missing is the wrapper around the model: tagged inputs, logged prompts, a review step, an audit trail.
The Copilot numbers tell the same story
Microsoft’s own numbers tell a parallel story. The same source puts weekly active use at 20 to 30 percent of licensed seats and daily active use below 40 percent even in committed organizations, with most seats generating zero return.
Firms I talk to usually explain this as a training problem or a change-management problem. Sometimes it is. More often the reason is the one the operations lead gave me. Staff open the tool once or twice, get an answer that looks plausible, then quietly go back to doing the work the old way because the old way is the way they trust. They cannot verify the AI output quickly, and they do not have time to verify it slowly. So they stop using it.
That is a rational response to a tool with no verification layer. It is not a failure of the team. It is a failure of how the tool was rolled out.
What closes the gap
Closing the data confidence gap is not a modeling problem. It gets solved in a specific order.
First, get the source data honest. Know where your project data lives, what format it is in, and what is wrong with it. Most firms have never done this audit. They bought the tool before they audited the inputs.
Second, put a verification step between the model and anything that leaves the building. A second model checking the first model’s citations against the source document. A rules-based check that flags any number outside an expected range. A mandatory human review on anything below a confidence threshold. The check is what turns a plausible answer into a verifiable one.
Third, pick one workflow and ship it. One recurring task that eats hours every week, with a defined input, a defined check, and a named owner. Watch what breaks. Iterate.
Firms that do these three things stop saying “we use AI but we don’t trust it.” They start saying “this is how we do this task now, and here’s how we know the output is right.”
The gap is fixable
The data confidence gap is not a reason to slow down on AI. It is the reason most firms have not gotten return on the licenses they already bought. Firms that close it first will move faster on every bid, every submittal, every internal study. Firms that do not will keep paying for seats that sit idle.
If this sounds like your shop, the AI Workflow Diligence Sprint is the fixed-fee way to find out where your gap actually sits and which one workflow to ship first.