Copilot or custom agent: how to decide

It is the first question any organization with existing licences asks itself: “we already have Copilot, why build anything?” It is the right question, and the honest answer is: often, you should not. We do not resell licences and take no commission, which lets us give that answer without ulterior motive.

At a glance

Criterion	Copilot / off-the-shelf	Custom agent
Strength	Assisted individual work	Multi-step processes, fine-grained rules
Systems	Microsoft ecosystem	Your line-of-business systems, via API
Deployment	Fast, via licences	Project, investment per process
Rule control	Generative, may improvise	Constrained, human in the loop
Best when	Save 30 to 60 min/day per employee	Quantified gain on a critical process

What Copilot does well

Copilot and its equivalents (Claude, Gemini, in their enterprise versions) excel at assisted individual work: drafting and summarizing documents, preparing a meeting from email threads, finding information across the organization’s files, producing a first draft of almost anything.

If your goal is for every employee to save thirty to sixty minutes a day on those tasks, the off-the-shelf tool is the right choice: fast deployment, predictable cost, security handled by the vendor. The real challenge becomes adoption, and that is a training problem, not a technology problem.

Where the off-the-shelf tool plateaus

Three situations come up constantly:

Multi-step processes. Copilot assists one person with one task. It does not orchestrate a workflow that crosses three systems, waits for an approval, then triggers invoicing. As soon as the word “then” appears three times in the process description, you have left its territory.

Systems outside the ecosystem. Your niche ERP, your Québec-specific line-of-business software, your Access database inherited from 2009: the general-purpose assistant does not see them. Custom work exists precisely to connect AI to the systems that actually run your business.

Strict business rules. When an error is expensive (compliance, invoicing, contractual commitments), you do not want an assistant that improvises. You want an agent constrained by your rules, escalating to a human as soon as it leaves the frame, with every decision traceable.

The three conditions for custom work

Before recommending a custom agent, we check three conditions:

1. The gain is quantified. The targeted process costs enough hours to pay back the development. A custom agent to save two hours a month is a bad deal.
2. The process is stable. You automate what has proven itself, not what changes every quarter.
3. The data is ready. The systems involved are accessible and their data is usable. Otherwise the project starts with a data cleanup, and you should know that before signing.

One more condition applies in Québec: if the agent processes personal information, Law 25 compliance is designed into the architecture, not bolted on afterwards. Hosting, data residency and the privacy impact assessment are part of the design.

The middle path, our most frequent recommendation

Start with the off-the-shelf tool, with real training so usage actually takes hold. Measure what it changes and note where it blocks. Those blocking points, documented over two or three months, form the business case for custom work: you know precisely which process to automate, and you already have the baseline numbers.

Buy first, measure, then build where buying plateaus: it is less spectacular than a grand project, and that is exactly why it works.

To situate your organization before deciding, the AI readiness quiz takes two minutes. And if you are already at the decision point, let’s talk: that is exactly what an audit settles.