AI Agents for AEC Firms: What They Actually Do
Three AI agents running at real architecture, engineering and construction firms, with the counts, who checks the work, and what agents still cannot do.
Columnline7 min read
AI agents at an architecture, engineering or construction firm take one repeat task that eats senior time, such as chasing insurance certificates, building submittal packages or researching new projects, and do it end to end on a schedule. A person still checks the hard cases, and the firm judges the agent the same way it would judge a new hire: is the work done, on time, without someone redoing it?
Most of what you read about AI agents in AEC is about the future: autonomous design, predictive scheduling, robots on site. This guide is about today. It walks through three tasks we run for real firms, with the counts from their published case studies, and what each agent can and cannot do.
The short version
- What an AI agent is: software that takes several steps on its own toward one result, instead of answering a single question. It reads, checks, compares and drafts, then hands the result to a person or sends it on. Nomic defines construction AI agents the same way.
- Where it pays first: paperwork that follows rules, arrives in volume and lands on expensive people. Certificates, submittals, quotes, research.
- Where it does not: judgment calls with no pattern, relationships, and anything nobody writes down.
- What decides if it works: one named task, one owner at the firm, and a person on the hard cases.
What does an AI agent do that a chatbot does not?
A chatbot answers when you ask. An agent starts the work itself, on a schedule or when an email arrives, and finishes it. A chatbot can tell you what a certificate of insurance should contain. An agent reads the certificate that just came in, checks it against the project's requirements, notes the expiration date and sends the reminder before it lapses.
That difference is why firms talk about agents now. McKinsey describes the same shift from tools that assist to systems that act. The hard part is not the model. It is picking a task narrow enough to finish and owning it after launch.
Three tasks AI agents run at real firms today
Each of these is live work we run for a client, not a demo. Counts are from each client's case study, published on columnline.com in October 2026.
1. Subcontractor insurance certificates (a homebuilder)
The before. California gives homeowners ten years to bring a construction defect claim. A builder has to prove every subcontractor was insured for that whole window. At Rincon Homes, one purchasing assistant tracked the certificates by hand in a spreadsheet.
What the AI employee does. It reads each certificate when it arrives and checks the required fields and entity names for that project. It catches small errors a busy person misses, such as a holder number typed with a 1 where the letter I belongs. It sends reminders 14, 7 and 1 day before a certificate expires, reads the replies and drafts answers. For the first month, a person approved every email before it went out.
The counts. 3,977 insurance coverage records tracked across 118 subcontractors and 27 active projects. A live board shows each subcontractor green, orange or red by project. Rincon Homes' story has the full detail.
2. Submittal packages and O&M manuals (a waterworks distributor)
The before. After a win, the general contractor wants submittals: proof of the exact product, size and fitting for every item, one package per spec section. Then closeout manuals. By hand, a set takes weeks of compiling, highlighting and spec checking.
What the AI employee does. It builds each package from the quote and the manufacturer's literature, in the distributor's own format, one tab per item. Manuals are built in the required format and checked 11 ways before they go out.
The counts. At CapFlo, more than 21 submittal packages across six jobs. On one lift station project, 14 section packages and 419 pages, with all 184 quote lines placed. Nine O&M manuals delivered and accepted by the general contractor. CapFlo's story shows the quoting side too, where every takeoff line is matched against a catalog of 17,508 items.
3. Researching new projects and owners (a general contractor)
The before. A business development lead finds new owners one Google search at a time, between everything else.
What the AI employee does. Each week it researches a batch of owners and projects that fit the firm's market and trade, and writes who to contact and why. The business development lead scores each one, and the scores tune the next batch.
The proof. DCI Group gets researched out-of-state senior living operators every week. DCI Group's story shows how the weekly batch works, and our plain guide to finding projects early explains the method.
What do these three have in common?
Look past the trade and the pattern is the same:
| Certificates | Submittals | Project research | |
|---|---|---|---|
| Arrives in volume | Every sub, every project | Every item, every win | Every week |
| Follows written rules | Required fields, dates | Spec sections, format | Market, trade, fit |
| Lands on senior people | Purchasing, an attorney | Estimators, the owner | The BD lead |
| A person checks | Emails, first month | Every package | Every batch, by score |
| Someone notices if it stops | Yes, at claim time | Yes, the GC asks | Yes, no Monday list |
The last row matters most. If an agent quietly stopped and nobody at the firm would notice, it was not doing real work in the first place.
What AI agents cannot do at an AEC firm (yet)
- Hold the relationship. An agent can find an owner and draft the note. A person makes the call.
- Decide with no pattern. Go or no-go on a bid, a claim strategy, a design call. It can lay out the facts; a person decides.
- Work from what nobody wrote down. If the rule lives in one estimator's head, it has to be written down first. At CapFlo, rules like rounding odd spool lengths up to the nearest 6 inches went into the AI employee the day an estimator said them.
- Run itself. Someone has to own it after launch: read the exceptions, add the new rule, notice when a supplier changes a form.
How should an AEC firm pick its first AI agent?
Ask three questions about each repeat task:
- Does it follow rules you could write on one page? If yes, an agent can learn it.
- Does it land on someone whose time is worth more elsewhere? That is where the hours come back.
- Who at the firm would notice in a week if it stopped? Name that person. They are the owner.
Pick the one task with the clearest yes on all three. Run it for a month with a person approving every output. Then let the agent send on its own where it has earned it, and keep a person on the rest. Who should your first AI workflow be for? goes deeper on choosing.
Want an AI employee running one of these at your firm?
Columnline builds and runs AI employees for architecture, engineering, construction and real estate firms. We pick one task with you, build the agent, and run it, with a person on our team on the hard cases. Book a call and we will show you which task at your firm is the best first one.
