How AEC firms should choose between Copilot, ChatGPT, and a custom AI build
You've probably got three options on the table. Microsoft is pitching Copilot seats, your project managers already use ChatGPT for something, and someone on the leadership team keeps asking whether you should build it yourselves.
Any of the three can be the right call, but they solve different problems. Buying one gives people a tool without changing how a submittal or a bid package actually moves through your firm.
If you already know which workflow is costing you, book a 30-minute call and we can talk it through. If your firm bought seats and they're sitting idle, start with why ChatGPT and Copilot licenses don't stick at GCs. This post is about the next decision, which is where to spend from here.
Each of these options is built to do a different job
Microsoft 365 Copilot is built for work that already lives in Microsoft 365
Copilot works inside Word, Excel, PowerPoint, Outlook, and Teams. With the full Microsoft 365 Copilot license, it grounds its answers in your email, files, meetings, and chats through Microsoft Graph, and each person only gets answers from content they already have permission to see. For a firm that runs on SharePoint and Teams, that's a big reason Copilot is often the safer broad rollout.
It's worth knowing that eligible Microsoft 365 subscriptions already include Copilot Chat at no added cost. That version works mostly from web data unless someone uploads a file or uses it with content they have open in Teams or Outlook, so check what you already have before you buy more.
Copilot can also reach past Microsoft through connectors. Microsoft lists more than 100 prebuilt connectors to outside systems. That's useful, but finding a document is a different job from changing who reviews it, in what order, and where the result gets logged.
ChatGPT is built for flexible thinking and drafting
ChatGPT is good at open-ended work, like drafting a tough client letter or working through a scope question. On ChatGPT Business and Enterprise, OpenAI says it doesn't train its models on your business data by default, and admins decide which connected apps are turned on. Its SharePoint app can search files a person can already open, once a Microsoft admin approves it if your organization requires that.
Where it stops for most AEC teams is the handoff. Unless you build around it, your firm's standards aren't baked in and the output still has to be copied back into Procore, Excel, or your template.
One note for your AI policy: OpenAI says it uses data from its services for individuals to train its models unless the person turns off the "Improve the model for everyone" setting, while business plans are excluded by default. Personal accounts aren't where bid numbers or client drawings should go.
A custom build is designed around one workflow you already run
A custom build starts from the work instead of the tool. You pick one recurring process, like the Friday owner report, and design a better path for it using the data and permissions already in Procore, SharePoint, Teams, Excel, or wherever that work lives, as far as access allows. Someone on your side owns it, there's a baseline from before anything changed, and a person reviews the output wherever the work calls for judgment.
A simple decision rule for architecture and GC firms
Here's the rule I use with firm leaders:
- Keep or buy Copilot when the work is everyday Microsoft productivity and you want AI that follows the permissions you've already set.
- Use ChatGPT, under a written firm policy, when the work is one-off drafting, research, and thinking.
- Build custom when the work repeats every week or every project, costs real hours, crosses more than one system, and needs an owner and a baseline.
Many firms end up with a mix, keeping Copilot for Office work and adding one purpose-fit build for the workflow that eats the most time.
When Copilot or ChatGPT is enough for an AEC firm
Seats are the right spend for a lot of everyday work, and that's true even when it means you don't need me yet. Good fits include:
- Polishing client emails and meeting follow-ups
- Teams meeting summaries and action items
- Cleaning up a Word narrative or an Excel table
- Light research before a go or no-go conversation
- First drafts of internal memos and policies
If most of your wish list looks like that, buy the seats and train people well.
The line comes when the bottleneck is a handoff that lives in Procore, your bid templates, or a Friday report built from three systems. Copilot and ChatGPT aren't a substitute for AEC-specific workflow design there, and more seats won't redesign the handoff.
When Copilot isn't enough for AEC work
The ceiling tends to show up in a few familiar places:
- Copilot can summarize a spec section, but checking a submittal against that section and your firm's review standards, then flagging what a reviewer has to look at, is a workflow someone has to design.
- Bids come in with different formats, missing line items, and exclusions, so the useful version of AI lands in the leveling sheet your estimators already trust, and a person still confirms scope and price before any award.
- A submittal or RFI first pass only helps when it shows up where your team works the log, not in a chat window someone has to remember to check.
- A Friday status report pulled from Procore, the schedule, and a cost spreadsheet needs a defined source for each number and a person who signs off.
- Proposals and SOQs built from past performance in Deltek or SharePoint need the right project sheets, resumes, and approved language every time. If marketing is where your pain sits, here's how I approach AI for AEC marketing teams.
Every one of those only helps if the output lands where the team already works, because on a deadline people skip anything else.
How a custom build compares with another AEC AI product or Copilot Studio
Another AEC AI product can be the right home when a specialized tool truly owns the workflow and your team is willing to adopt it. If that's your situation, I'd rather tell you than build a weaker copy. The trade is another login and another vendor's way of doing the work.
Copilot Studio can be right when your data already sits in Microsoft 365 and you have IT capacity to build and maintain agents. Microsoft positions it as the place to create custom agents and connect them to your systems, and agents published from it can run inside Teams, SharePoint, and Copilot Chat. Usage is billed through Copilot Credits, though Microsoft says use of agents published to Microsoft 365 Copilot is included for people who already have that license, so check how your seats are licensed and plan for the running cost along with the build.
A custom build inside the tools you already open is the middle path I default to. It starts from the workflow, including the parts that live in Procore and other field and ops systems, and it ends with someone on your team who owns it.
On the 90-Day AI Sprint page, I lay out the order I work through these choices:
- Remove steps that don't need to exist.
- Use what you already own, including your existing software and licensed AI tools.
- Consider a suitable off-the-shelf product.
- Connect the systems already in the workflow when access and permissions allow.
- Build custom where the earlier options can't meet the need.
Before we commit to an integration, I confirm permissions, data export or API access, operating costs, support needs, and an internal owner, including how your Procore, Sage, Viewpoint, Deltek, or Bluebeam setup actually works.
What a custom path looks like over 90 days
The 90-Day AI Sprint is how I package that middle path. By default it covers one workflow we agree on, and it isn't sold by how many agents we build, since agents only come in when the work needs them.
The Sprint runs in three overlapping stages. Prioritize is about picking the workflow and writing down how it runs today, so there's a baseline to measure against. Build is where we rebuild how the work gets done while your team practices on live examples, with review kept wherever a person needs to make the call. Prove is the results review, where we look at quality, adoption, and operating cost alongside speed.
At day 90 you should have:
- One agreed workflow and the baseline we used to review it
- A working path for that workflow, fitted to the systems we can access
- A prepared internal owner who can run daily use and routine exceptions
- A results review against that baseline
- A short ranked list of next opportunities from the scope we explored
I won't promise a set financial return, and integrations depend on what we can access and what's feasible. For the full detail, see the 90-Day AI Sprint. If you'd rather talk it through first, pick a time.
If your seats are already sitting idle, start with why licenses don't stick
If you already bought seats and usage faded to email drafts, the earlier post on why ChatGPT and Copilot licenses don't stick at GCs covers what went wrong and why the fix usually isn't more seats.
Here are the questions AEC leaders ask most about Copilot, ChatGPT, and custom AI
Should our architecture or GC firm buy Microsoft Copilot or build custom AI into Procore and SharePoint?
Buy Microsoft Copilot when the work is everyday productivity inside Word, Excel, Outlook, Teams, and SharePoint and you want AI that follows your Microsoft permissions. Build custom AI into Procore, SharePoint, and the rest of your stack, where access allows, when the bottleneck is a recurring, expensive workflow that crosses systems and still ends in copy and paste. A license gives people access to a tool, while a build gives one workflow an owner and a baseline.
What is Microsoft Copilot good for in an AEC firm, and when isn't it enough?
Copilot is strong for drafting, meeting summaries, spreadsheet help, and finding content already inside your Microsoft tenant, and it only answers from what each person already has permission to see. It isn't enough when the real work is a Procore package, a bid template, a review standard for drawings and specs, or a status report assembled from several systems. Copilot can search and summarize some of what's in your Microsoft tenant, but more seats won't redesign the handoff.
What is ChatGPT good for in an AEC firm, and where does it stop?
ChatGPT is good for flexible drafting, research, and working through a problem. On Business and Enterprise plans, OpenAI says it doesn't train its models on your business data by default. It stops at the handoff, because unless you build around it, your firm's standards aren't built in and the output still has to be moved into Procore, Excel, or your templates by hand.
When is a custom AI build the right answer instead of more seats?
A custom build makes sense when the work repeats every week or every project, costs real hours, crosses more than one system, and someone on your team can own it. If you can name the handoff and measure how it runs today, it's a good candidate.
How is a custom AI build different from ChatGPT seats, another AEC AI product, or Copilot Studio?
ChatGPT seats help people think and draft. Another AEC AI product means a new login and that vendor's workflow, which is the right trade when it truly owns the work. Copilot Studio lets your IT team build agents inside Microsoft's ecosystem. A custom build starts from one painful process you already run, fits it to your data and permissions in tools like Procore and SharePoint where access allows, prepares an internal owner, and expands only after you've reviewed the result against a baseline.