Custom Copilot Agents: What it actually takes to build agents that can handle pursuits

If you lead marketing or business development at an AEC firm, you’ve likely had some version of this conversation. You asked IT if Microsoft Copilot could streamline your proposal development, and IT answered with confident optimism: "We can just build that in Copilot Studio."

It is a fair instinct. Copilot is genuinely effective at horizontal enterprise work—summarizing meeting notes, drafting initial emails, and indexing project files scattered across last year’s drive. But once you examine what high-stakes pursuit work actually requires, building a custom Copilot agent for proposal workflows proves to be a vastly different challenge.

Most in-house AI builds hit the exact same structural roadblocks. This guide covers what every custom build must account for before your pursuit teams can actually trust an agent’s output on an active RFP. But before we get to that, some basics.

What do Microsoft Copilot agents do?

Microsoft Copilot agents connect to your enterprise data (SharePoint, Dataverse, Dynamics, or other connected systems) and carry out a defined task without a person prompting every step. Pre-built agents ship with Microsoft 365 Copilot for functions like meeting facilitation, research, and IT helpdesk triage. Custom agents, built in Copilot Studio, extend that to whatever workflow your team configures. The question this piece answers is whether that configuration is enough to safely run something like RFP shredding or compliance-matrix generation, or whether it just relocates the manual work.

Does Copilot have an agent mode?

Yes. Copilot Studio is Microsoft's dedicated platform for building custom agents on top of Copilot, and Microsoft 365 Copilot also ships with several pre-built agents out of the box. Whether that agent mode performs like a true autonomous agent in a pursuit workflow, or just a well-dressed chatbot, comes down to the four gaps below: taxonomy, fact verification, production hand-off, and governance.

Four Critical Gaps in General Enterprise Agents or Assistants (even if you customize them)

#1 Does pointing Copilot at SharePoint give it your taxonomy? (The Taxonomy Gap)

Pointers to a SharePoint directory allow Copilot to read your documents. It can index last quarter's submittals, staff resume archives, and boilerplate project descriptions.

What it lacks, out of the box, is your business development taxonomy, your pursuit history, and your underlying project data model. Grounding on a folder is not the same as grounding on what your firm has won, lost, and learned. A general assistant treats every uploaded document as equally true. Pursuit work, by contrast, requires understanding which specific project references apply to a target opportunity—and precisely why.

Every AEC firm categorizes its experience differently. One firm’s "Healthcare" portfolio might span acute care facilities, ambulatory surgery centers, and lab spaces—each governed by distinct regulatory language, codes, and client expectations. No general agent arrives pre-loaded with that sector breakdown; you have to construct and maintain that relational schema manually.

When pointing at directories containing conflicting drafts, outdated templates, or unverified spreadsheets, the DIY Copilot agent simply averages across all available text. The output is generic and flat because it has no way to distinguish your best-fit, winning narrative from a simple keyword match.

#2 Can Copilot verify a fee number against your CRM? (The Fact-Verification Gap)

Copilot grounds its responses in SharePoint content to mitigate hallucinations. However, grounding on a document folder and verifying facts against a central system of record (like your CRM) are fundamentally different operations.

Query a general assistant for a project’s final square footage or contract value, and it will return an answer based on whichever text file it indexes first—whether that file was updated yesterday or abandoned three years ago. Pursuit submittals cannot rely on "plausible" data. Submitting incorrect project metrics creates legal liability and risks immediate disqualification during technical evaluations. Pursuits require specialized tools that source provenance, audit trails, and verification against your ERP or CRM (like Deltek or Salesforce).

#3 Where does a Copilot draft land? (The Last-Mile Production Gap)

Copilot workflows generally terminate at a Word document or a chat window. Proposal production ends at an Adobe InDesign file, strictly configured to page limits, visual grid systems, and complex typography rules.

If your proposal team receives an unformatted text draft in Word, someone still spends hours manually copying, reformatting, and fitting that text into InDesign frames under deadline pressure. The operational bottleneck hasn't been eliminated—it has simply shifted downstream to layout assembly.

#4 Who maintains a custom Copilot agent after the builder leaves? (The governance & maintenance gap)

A DIY agent built by a single motivated person in innovation or IT becomes an orphaned project the day that person changes roles.

Without ongoing ownership, a custom agent rapidly deteriorates:

• Model Drift: When Microsoft updates the base model, systems can break unless someone retests them.
• Taxonomy Stagnation: Entering a new market or adding new assets requires manual schema updates that nobody is assigned to make.

Proper Copilot Studio governance and support from a Power Platform CoE (Center of Excellence) are critical. Gartner projects that over 40% of agentic AI deployments will be canceled by the end of 2027, largely due to escalating costs, unclear business value, or inadequate risk controls.

What does a custom Copilot agent cost?

Basic Copilot usage is included in your Microsoft 365 Copilot license, up to an unpublished fair-use threshold. What isn't included is anything resembling a real pursuit agent. The moment you build in Copilot Studio, or an agent starts grounding itself in tenant data and taking autonomous actions, it draws down separately metered Copilot Credits — on top of the seat license, priced at $0.01 per credit pay-as-you-go or $200 for a 25,000-credit monthly pack.

That meter starts at build time with testing and iteration burning credits before the agent ever touches a live pursuit. And the pricing isn't flat across actions: a simple grounded answer might cost a couple of credits, while an autonomous action or tenant-data grounding step can run 25–30 credits each, according to Microsoft's own published consumption rates. An agent stuck reasoning in a loop over a malformed 200-page RFP, or one rerunning a check because nobody defined a clean stopping condition, keeps drawing credits until someone notices the bill.

By default, there's no hard cap on that spend. Microsoft's admin center lets you set one, but it's opt-in, not automatic.

Before signing off on a build, get straight answers to three questions:

• Which actions are zero-rated under our license, and which draw Copilot Credits?
• What's the estimated credit burn per agent run, and per month, at our team's actual pursuit volume?
• Have we set a hard spend cap, and what happens when a run hits it?

Should you build Copilot Agents?

Keep Copilot for horizontal workflows across the enterprise. Internal meeting recaps, first-pass email drafts, searching static guidelines, and general market research? Use Copilot agents there.

For pursuit workflows that directly impact win rates and proposal accuracy, equip your pursuit teams with a system purpose-built to understand opportunity history, project fit, source provenance, and how your firm wins work.

See how specialized agents bridge these gaps across intake, compliance, and experience matching in our complete inventory of 24 AEC proposal use cases.

If you’re looking to have a candid conversation about your agentic AI strategy or want a sneak peek into what we’re building around agents, book a chat with our team.

FAQs

Is Microsoft Copilot agent free?

Partly. Basic Copilot Chat and agent interactions are included in your Microsoft 365 Copilot license, up to an unpublished fair-use threshold. But building a custom agent in Copilot Studio, or running one that grounds itself in tenant data and takes autonomous actions, draws down separately metered Copilot Credits on top of that license. 

What are some examples of Copilot agents?

Out of the box, Microsoft ships agents for meeting facilitation, IT helpdesk triage, and researcher/analyst-style summarization. Custom Copilot Studio builds at AEC firms typically start narrower still, indexing SharePoint folders, drafting first-pass emails, summarizing past submittals. That's the horizontal work Copilot already handles well, as opposed to full pursuit workflows like compliance matrix generation, which is where the gaps in this piece show up.

‍