Should you build or buy an AI agent? For a business on Microsoft 365 there is a third answer as well: make it yourself in Copilot Studio. Each route can work. A poor choice tends to show up later, as an agent nobody maintains, a bill nobody expected or a platform that cannot reach the system the job depends on.
This post compares the four realistic routes for a first agent, the questions that point to each one and the costs to plan for. If you have not picked the job yet, start with our post on choosing your first AI project.
What counts as an AI agent here?
An agent is an AI assistant that follows your instructions, draws on the knowledge you connect and uses tools to take action. Microsoft describes Copilot Studio agents in much the same terms. A sound first agent does one job well: answering staff questions from approved documents, drafting replies to enquiries for a person to approve, or gathering the details of a request before passing it on.
Build or buy an AI agent: the four routes compared
| Route | Who builds it | Suits | Watch for | How it is paid for |
|---|---|---|---|---|
| Buy a ready-made agent | The vendor | A common job a product already handles, such as bookings or routine questions | Fit with your process, where data goes, and how you would leave | The vendor's subscription |
| Agent Builder in Microsoft 365 Copilot | Staff, with no code | Internal agents grounded in Microsoft 365 content | It cannot author actions that call external services | Depends on the capabilities used and your Copilot licensing |
| Copilot Studio | A confident in-house maker or a developer | Agents that need connectors, workflows, approvals or channels such as a website | Credit use, connector rules and who owns the agent | Copilot Credits, through pay-as-you-go, prepaid packs or a prepurchase plan |
| Custom build | A developer | Jobs that cross systems outside Microsoft 365 or need custom logic | Running costs and support after launch | A build price plus monthly running costs |
When does buying make sense?
Buy when the job is common and someone already sells a product that does it the way you work. Booking appointments, answering routine website questions and scheduling social posts are typical examples. You get something already in use elsewhere, and the vendor looks after it.
The questions to ask are about fit and data rather than features. Can it follow your process, or will you have to change yours? Where is your data stored and processed? What happens to your data if you leave? Our ready-made and custom AI agents page shows the kind of product that suits this route.
When is Agent Builder enough?
If the agent needs only instructions and content you already keep in Microsoft 365, start with Agent Builder. It is Microsoft's guided, no-code way to set up an agent inside Microsoft 365 Copilot. You can share the agent with colleagues or submit it to your organisation's catalogue, subject to your admins' policies and approvals. Microsoft's own example is an onboarding or project information agent grounded in approved content.
Its limit is action. Agent Builder does not support authoring actions that call external services. If you outgrow it, you can copy the agent into Copilot Studio, although the copy is a separate snapshot rather than a linked upgrade, and some knowledge sources and settings need setting up again.
When should you use Copilot Studio?
Copilot Studio is Microsoft's low-code studio for building agents and workflows. It connects to other systems through prebuilt or custom connectors and publishes agents to channels such as Microsoft Teams, Microsoft Copilot and websites. Microsoft's guide to choosing between Agent Builder and Copilot Studio points to Copilot Studio when you need an external channel such as a customer-facing website, actions in other services, workflow logic, or separate development, test and production environments.
Two things are easy to underestimate. The first is cost. Under Microsoft's Copilot Studio licensing, agents consume Copilot Credits, which you buy through pay-as-you-go billing on an Azure subscription, prepaid packs or a one-year prepurchase plan. The credits each response uses depend on how complex the task is, and unused prepaid credits do not carry over to the next month. There is a useful exception: when someone with a Microsoft 365 Copilot licence uses an agent in Copilot Chat, Teams or SharePoint for answers or Microsoft Graph grounding under their own identity, that use is not charged. Billing also varies with the engine the agent runs on, which Microsoft calls a harness, so check the current rules for the type you choose.
The second is ownership. A Copilot Studio agent is software. Someone has to test changes, watch its analytics and keep its connectors in line with your policies.
When is a custom build the right call?
Build when the job depends on systems outside Microsoft 365, such as your CRM, booking system or a line-of-business database, or when it needs logic and checks that a low-code platform handles awkwardly. A custom agent also suits work that must run on a schedule with little supervision, or where you want a say in hosting and models.
A custom build should still be narrow. Write down its goal, the tools it may use, the data it can reach, its limits and where a person approves its work, then test it against real examples before it goes live. At APPS 365 that design sits in the written quote. Our custom AI agent development starts from £999, with a fixed quote after a free call, and monthly running costs apply because models and hosting charge by use.
If the agent belongs inside Teams or SharePoint, our Copilot Studio agent development chooses between a declarative agent, a Copilot Studio agent and a custom build, and includes a review of the SharePoint content and permissions the agent will use. Microsoft licences and Copilot Studio capacity are separate from our price and set by Microsoft.
Five questions that settle it
- Does a product already do this job well enough? If so, buy it and put your effort into checking fit and data.
- Does the agent need only Microsoft 365 content and internal users? Start in Agent Builder.
- Does it need to take actions, run a workflow or face customers? Look at Copilot Studio.
- Does it depend on systems Microsoft 365 cannot reach easily, or need custom logic? That points to a custom build, so get it quoted.
- Who will look after it in six months? If nobody in-house can, favour a bought product or a supported build over a do-it-yourself agent.
Whichever route you take, check SharePoint permissions first if the agent will answer from your files. Agents grounded in Microsoft 365 work within each user's access, so anything shared too widely can surface in their answers. The free SharePoint check before Copilot is worth doing before launch.
Get a straight answer
For a structured start, our AI guides and kits include an AI Agents Kit that walks through choosing a platform and setting up a first agent with approvals, permissions and tests. To talk through your own job and which route fits it, book a free 30-minute call.
