Copilot, agents or custom AI: a decision guide
Microsoft 365 Copilot, agents on your own data, or a custom solution. How to pick the right level for each problem, and keep governance and cost under control.
Written with AI and reviewed by the Logol team.
Three levels, one question
Most AI discussions in mid-sized companies arrive at the same decision. Roll out Microsoft 365 Copilot and see what happens? Build agents on our own data? Or invest in something custom, or in software made for our industry?
The useful question is not which option is the most powerful. It is where the work actually happens, and what the AI needs to see in order to help. Answer that honestly and the choice usually becomes clear.
Think in three levels. Level one is personal productivity: Copilot inside Outlook, Teams, Word and Excel. Level two is agents grounded in your own data and processes, built with Copilot Studio or Azure AI Foundry. Level three is a custom solution or vertical software, where AI is part of the system that runs the business. Most organisations end up using more than one level. The mistake is using the wrong one for a given problem.
When Microsoft 365 Copilot is enough
Copilot works with what a user can already access in Microsoft 365: emails, meetings, chats, files in OneDrive and Teams. It is strong at summarising, drafting, finding and rewriting. If the problem is that people spend too much time reading, writing and searching, Copilot is often the right first step.
- The work is individual and varied, not one repeatable process.
- The relevant information is already in Microsoft 365.
- The output is a draft that a person reviews and owns.
- Success means time saved per person, not a redesigned workflow.
Two preconditions are easy to underestimate. Copilot respects existing permissions, so if your file permissions are too broad, Copilot will surface content people should not see. Clean up access before the rollout. And adoption needs training: people who get concrete examples for their own role use Copilot very differently from people who get a licence and an email.
When you need agents on your own data
Copilot reaches its limits when the work is a process rather than a task, or when the data sits outside Microsoft 365: in your ERP, CRM, accounting system, DMS or a line-of-business database. This is where agents come in.
With Copilot Studio you build agents that answer from defined knowledge sources, call connectors and Power Automate flows, and act inside a business process: handling supplier enquiries, pre-qualifying requests, preparing records for approval. Azure AI Foundry is the step up when you need more control: a choice of models, including those from OpenAI and Anthropic, custom retrieval over large document collections, systematic evaluation, and integration into your own applications.
- The same process runs often, with known rules and exceptions.
- The AI needs data from systems beyond Microsoft 365.
- The agent should not only answer, but prepare or trigger actions.
- You need to measure quality systematically, not by anecdote.
A practical rule: start in Copilot Studio when the process fits standard connectors and low-code flows. Move to Azure AI Foundry when you need model choice, deeper retrieval, or when the agent becomes part of a product or a core system.
When custom or vertical software is the right choice
Some problems are not solved by adding AI on top. If the underlying data is spread across ten systems, an agent will inherit the fragmentation. The right step is then to fix the system of record first and let AI work inside it.
Custom development makes sense when the process is a genuine differentiator, no standard product covers it, and you are ready to own and maintain the result. Vertical software makes sense when your industry has its own data model, regulation and workflows that a generic tool would take years to learn: a law firm's matters, conflict checks and time records, or a family office's positions across banks and structures. There, AI is most useful when it already has the full context of the domain, not just access to a collection of documents.
The trade-off is straightforward. Custom gives maximum fit and maximum responsibility. Vertical software gives domain depth and a vendor roadmap, in exchange for adapting to a standard. Both are better than a patchwork of generic agents on top of fragmented data.
Governance and cost control at every level
Whatever the level, the same disciplines keep AI safe and affordable.
- Access first. AI sees what its identity can see. Review permissions and data classification before rollout, not after an incident.
- Human approval for actions. Agents propose. People decide on anything that changes records, sends messages or commits money.
- Logging and audit. Keep a trace of requests, sources, proposals and decisions, so that an outcome can be explained later.
- Data protection by design. Know where data is processed and stored. Assess use cases involving personal data against the Swiss FADP and, where relevant, the GDPR. If you offer or use AI systems in the EU, check how the risk-based obligations of the EU AI Act apply to you.
- Cost visibility. Per-user licences are predictable. Consumption-based agents and model usage are not, unless you set budgets, limits and alerts. Assign licences by role and actual use, and review consumption regularly.
- A use-case register. One list of what is live, who owns it, which data it uses and how it is measured. It is the simplest governance tool, and the one most often missing.
Our AI integration team helps companies make exactly these choices on their Microsoft stack, from a Copilot rollout to agents in Copilot Studio and Azure AI Foundry, and, where it fits, our own platforms for law and finance. This guide can, however, be used on its own. Match the level to the problem, and govern it from day one.