Copilot Studio or custom AI? Five questions that decide the route

Push every idea into Copilot Studio because it's licensed, and you end up with agents that fail on the hard cases. Commission a custom build for every idea, and you pay to solve problems a licence already covers. Copilot Studio is very good at a specific kind of problem, custom AI is right for another, and a growing share of what gets called an AI use case needs neither.

First, be clear which Copilot you mean

Copilot follows a familiar Microsoft tradition: one name, several products, and a support article to help you work out which one you have. Anyone who has had to ask “Which Outlook?” or explain that using Skype for business didn’t necessarily mean you were using Skype for Business will recognise the problem. Three of the products Microsoft calls Copilot matter here, and the differences decide what's possible.

• Copilot Chat comes with eligible Microsoft 365 licences. It's grounded in the web and in content the user opens or shares with it, and it doesn't search across your organisation's content the way the paid licence does.

• Microsoft 365 Copilot is the paid licence, which Microsoft labels Premium. It works across a user's email, meetings, chats and documents, inside Outlook, Teams, Word and the rest of the suite.

• Copilot Studio is the low-code platform for building agents. An agent can be grounded in specific sources such as a SharePoint site or a website, take actions in other systems through connectors, and be published to Teams, Microsoft 365 Copilot or a public website.

Bear with me, I'm going to get Claude to explain the Microsoft pricing dynamics: Copilot Studio usage is measured in Copilot Credits, bought pay-as-you-go or in prepaid packs, and what you pay depends on who's using the agent. Staff with a Microsoft 365 Copilot licence are largely covered by that licence when they use agents inside Microsoft 365. Staff with only Copilot Chat can still use agents grounded in specific SharePoint sites or other organisational sources, with that usage metered in credits (Microsoft's Copilot Chat FAQ has the details). You don't need a Premium licence for every user to put a useful agent in front of them. Clear?

Custom, in this post, means AI you build and run yourself, typically on Azure using all the fancy tools in Microsoft Foundry (there's another naming slight to be made here) or on private infrastructure where the data demands it.

Five questions that decide the route

For any single use case, five questions do most of the work.

1. Who uses it? Staff with Microsoft 365 accounts are Copilot and Copilot Studio territory. Members of the public rule out Microsoft 365 Copilot straight away, which leaves a Copilot Studio agent on a website or AI built into the digital service itself.

2. Where does the knowledge or data live? Content in SharePoint, OneDrive, Outlook and Teams is where Copilot is strongest. Data in a line-of-business system, such as a case management or grants platform, needs a connector or a clean API: a Copilot connector can make it searchable in Microsoft 365 Copilot, and a Copilot Studio agent can act on it. Without either, it's custom work. If another organisation owns the system, the first conversation is usually about integration, not AI.

3. What does it have to do? Drafting, summarising and finding things for one person are good starting points for Copilot. A repeatable team process, especially one that ends by updating a record, raising a ticket or routing a request, is where Copilot Studio can fit well. When the logic is the product, such as validating evidence against several systems or preparing an eligibility recommendation, test whether the platform provides the control you need. Custom components may be necessary, while Copilot Studio still provides the interface or workflow. Test sets, evaluation and an appropriate audit trail matter whichever route you choose.

4. What's the scale and cost model? Microsoft 365 Copilot is licensed per user, Copilot Studio is billed on consumption beyond what that licence covers, and a custom build carries build and running costs. An agent for 20 caseworkers and a service for 20,000 applicants land in different places.

5. What happens when it's wrong? A misrouted email is recoverable. A wrong eligibility decision, or personal data sent where it shouldn't go, is not. The higher the stakes, the more you need custom evaluation and human oversight, or a design that doesn't rely on generative AI at all. Sensitivity also decides where AI can run: in our analysis of 50 high-volume Scottish public services, we estimated that 24 would need private AI infrastructure rather than hyperscale cloud, mostly in health, policing and justice. No SaaS route works for those, Copilot Studio included.

Five routes, not two

Framing the choice as Copilot versus custom leaves out routes that matter. We work with five, and Microsoft's own decision tree for agents works the same way, asking whether you need an agent at all and whether a ready-made one does the job before it considers building anything. Our AI roadmap for Scottish Forestry ended up with use cases on all five.

• Copilot as it comes. Microsoft 365 Copilot licences targeted at the teams with the heaviest meeting and drafting load, rather than a blanket rollout. That's a licensing and governance decision, not a build.

• Copilot Studio agent. Triage for a busy shared inbox: classify incoming email, suggest a route and draft a reply for an officer to check. Keep a custom classifier in reserve for a second phase, if accuracy demands it.

• AI in something you already have. Redaction built around a tool that's already available, or imagery analysis tried in your GIS platform's own AI before any custom build.

• Custom build. Validating application documents against records held in several systems.

• Largely not an AI problem. Checking an application for the right form version, the required attachments and completed mandatory fields, which is mostly rules, with AI mainly there to read scanned forms.

Routes also combine, and a grant guidance assistant for applicants shows how the questions get you there. Applicants are the public, which rules out Microsoft 365 Copilot. The knowledge is thousands of pages of scheme guidance, which needs a retrieval pipeline built for it. The job is answering and signposting, which suits a Copilot Studio agent on the website as the front end. And a wrong answer misleads an applicant, so the first step isn't technical: it's an audit of the guidance, because an assistant grounded in duplicated or contradictory guidance gives duplicated and contradictory answers, confidently.

The landscape decides more than the use case

The constraints that shape an AI roadmap are usually about the environment the AI will live in, not the models. Check it before anyone scores a use case.

• Licensing. Copilot Chat only, Microsoft 365 Copilot for some or all staff, Copilot Studio capacity, Power Platform. What's actually licensed moves the line between routes more than anything else.

• Tenancy. Whether you run your own Microsoft tenant or share one you don't administer. On a shared tenancy, platform and automation decisions sit with the tenancy owner, so Power Platform use cases depend on their roadmap as well as yours. If you're on the Scottish Government's tenancy, Serving Scotland could change the picture, so confirm what reaches you and when.

• Systems. Which systems hold what, who owns them, and whether anything can talk to them.

• Content and permissions. Whether documents really live in SharePoint or are still spread across file shares, records systems and personal drives, and whether the permissions on them are right. Copilot can only work with what it can see, and it respects the access people already have, so an overshared site becomes an overshared answer.

• Governance. An AI acceptable use policy and an approval route before anything reaches staff.

You don't need access to the data for any of this, or for the five questions, only knowledge of it: where it lives, who uses it and how sensitive it is. That comes from people, through workshops and an honest conversation with whoever runs your IT.

Where the public sector opportunity lands

Our research paper, Automate tasks, not jobs, found that those 50 services cluster into five repeatable patterns. The hours below come from that analysis; the routes are our judgement from applying the five questions.

Copilot licences can make a real difference to routine documentation and correspondence, but the evidence is mixed. The Government Digital Service's cross-government Microsoft 365 Copilot experiment reported 26 minutes saved a day on average. The Department for Business and Trade's own evaluation found high satisfaction but no robust evidence that time saved became productivity gained, though it wasn't set up to measure that, and some tasks slowed where output quality fell short. Either way, three quarters of the hours in the table sit inside systems of record, where Microsoft 365 Copilot reaches least on its own. The biggest prizes need more than a licence.

Test it, then scale it

Whichever route a use case lands on, prove it before you scale it. Microsoft's Cloud Adoption Framework recommends time-boxed prototypes of one to two weeks per option, comparing a low-code Copilot Studio agent with a pro-code build. The five questions tell you which route to try; building it tells you what that route runs into. As Jack Strachan argues in Alpha is discovery, whatever you build becomes "a probe into the institution around it", and what it hits is often the real finding: an integration that doesn't exist, data that was never captured, a decision nobody on the team is allowed to make. On a shared tenancy, that last one is usually the tenancy question from earlier, turning up just as the pilot starts to work. It's also the moment to baseline the service, so you can show the AI moved something that matters.

We build Copilot Studio agents, custom AI components and private AI for sensitive workloads, and we'll tell you when a use case needs none of them. If you're sorting a long list of AI ideas into routes, or have a Copilot Studio agent outgrowing the platform, get in touch. Our data and AI sprints are one way to run that first test in a week, and we're running a limited number for free in 2026.