Home / Copilot agents

Plain English guide

Microsoft Copilot agents for UK businesses.

Plain-English guidance for accountancy and professional services firms deciding where agents fit, how to govern them and when Copilot Studio is the right build route.

Good first agent signals

  • A repeated job with something that starts it: an email, a file, a date.
  • Trusted source content with a clear owner.
  • A defined point where a person approves.
  • A safe route to a person when the agent isn't sure.

There are two kinds of agent. Agents you talk to help people find and draft things, and Microsoft's own Copilot now does a lot of that out of the box. Agents that run start on their own when something happens: an email arrives, a meeting ends, a report is exported, a date comes round. They do the checking, drafting and filing, then stop for one of your team to approve. The second kind is where a firm gets hours back, and it's what we build.

Both kinds are built around a particular job, source of knowledge, audience or workflow, rather than asking general Copilot to help with whatever is in front of you. Microsoft keeps making Copilot better, and we'll tell you when it already does the job. What it doesn't know is your firm: your clients, your thresholds, your templates, your filing conventions and who signs off what. That's the part we connect, build and look after.

In an accountancy or law firm, an agent you talk to might be a knowledge assistant or a client intake guide. An agent that runs might file the morning post, push WIP exceptions to managers or turn a client meeting into a filed note. The important detail is not that the agent is clever. It is that the scope is narrow enough to trust.

Agents work best when they sit between broad Copilot use and hard automation. They can guide a person through a process, answer from approved material and collect useful context. They should not remove professional judgement, client responsibility or the need to check the output. When the job needs a build, it is built as part of a Firmwide install, one agent per department; when it only needs rule-based steps, a standard automation inside the same install is usually the better route, and Run keeps both working.

For the wider picture of how agent capability grows, the Agent Journey maps five stages from the Copilot licences your firm already has to a team of agents. A Firmwide install builds the skills stage and the agent stages, on a foundation of permissions checked and the firm's templates and procedures in SharePoint.

Types of agents

Not every agent needs the same build route.

The right route depends on the job, the source content, the audience and how much control the firm needs over testing, actions and handover.

In Microsoft 365

Microsoft 365 agents

Agents that sit close to everyday Microsoft 365 work are usually best for helping people find knowledge, follow internal guidance or complete a repeated task without leaving the flow of work.

  • Work near Teams, SharePoint, Outlook and Copilot
  • Useful for internal guidance and knowledge retrieval
  • Depend heavily on permissions and source content

In Copilot Studio

Copilot Studio agents

Copilot Studio agents are better when the job needs a defined conversation, a managed set of instructions, specific knowledge sources or a connection into a workflow.

  • Good for intake, triage, onboarding and guided support
  • Need a clear owner, test plan and escalation route
  • Can work alongside Power Automate when actions need to happen

Configuration choice

Declarative and custom agents

Some agents can stay close to existing Microsoft 365 content and instructions. Others need a more custom build with tools, connectors, data sources and stronger testing.

  • Declarative agents are often faster to prove
  • Custom agents suit specific processes and integrations
  • The right choice depends on risk, source material and maintainability

Microsoft 365 Copilot

Variable knowledge work where a person stays in control.

Drafting, summarising, meeting follow-up, analysis support and rewriting.

Copilot agents

A narrower assistant around a repeated task, audience or knowledge source.

Internal policy guidance, client intake, onboarding support and document review preparation.

Agent automations

Work that starts on its own when something happens, needs AI for one judgement step and must stop for a person to approve.

Starts when an email arrives, a file lands or a date comes round. Uses AI to read, match or draft, acts through Power Automate, then waits for approval in Teams.

Power Automate

Repeatable, rule-based steps that should run without a conversation.

Approvals, reminders, notifications, task creation, file movement and status updates.

Human work

Judgement, relationship handling, final review and decisions with professional responsibility.

Client advice, matter strategy, partner sign-off, exception handling and ethical judgement.

Good use cases

Where agents tend to earn their keep.

Incoming post

Starts when the day's scanned post lands in a folder. Each letter is read, matched to a client and filed with an audit line, and anything uncertain goes to a person.

WIP exceptions

Starts when the practice system's WIP report arrives. Each manager gets their own exceptions in Teams with bill now, defer and write off buttons, and every decision is logged.

Records chasing

Starts when a job is created. The client gets the request list, everything that arrives is matched to it and chasers are drafted at day 7 and day 14 for the manager to approve.

New client checks

Starts when a new client record is created. Companies House details are pulled, the firm's approved searches run and the evidence is filed, and only the hits go to the MLRO, or a partner where the firm has no MLRO.

Client onboarding

Starts when the intake form is submitted. Gaps are chased, the checks run, the client workspace is created and the letters are generated, with one partner approval before anything goes to the client.

A client meeting filed and actioned

Starts when the Teams meeting ends. The file note is drafted in the firm's template with the actions, then the note is filed, the tasks created and the follow-up drafted only once the adviser approves.

  • Unreviewed client advice, legal advice, tax advice or regulated judgement.
  • A vague AI assistant for everything with no audience, owner or quality bar.
  • Processes where nobody agrees the current steps, exceptions or handoffs.
  • Knowledge sources that are out of date, duplicated, unowned or permissioned too widely.
  • Work where the answer needs emotional judgement, negotiation or a direct client relationship.
  • A build that only exists because someone wants to show off an AI demo.
  • A chat agent that repeats what Copilot already does.

Governance and security

Useful agents have boundaries before they have users.

Agent governance is not there to slow the work down. It is what lets a firm use agents without turning source content, permissions and professional judgement into avoidable risk.

Permissions

An agent people talk to should only reach what the person asking can already see. An agent that runs on its own uses a service account limited to the mailboxes, folders and systems its job needs. Either way, permission hygiene is part of the product.

Source content

An agent built on poor content gives poor answers faster. Useful builds usually start by naming the approved sources and cleaning up the obvious gaps.

Owners

Every agent needs a business owner for scope and content, plus a technical owner for configuration, access and support.

Review process

Good agents are tested against normal questions, awkward questions, edge cases and the moments where they should say they cannot help.

Lifecycle and retirement

Agents should be reviewed, improved or retired. If the process changes and the agent does not, trust will fade quickly.

Inside an agent build

How agent work is scoped and built.

The aim is not to build the cleverest thing possible. It is to build the smallest useful agent that has a clear job, safe boundaries and a route to being maintained. Every agent FiveForward builds follows the same five steps, the FiveForward Framework: Check, Plan, Build, Handover, Run. The job is named on the Readiness Check, the price is fixed on the Plan on a Page, the agent is built and tested on real cases in the install, handed to a named owner with a one-page scope and a kill switch, and kept working by Run.

What is a Microsoft Copilot agent?
A Copilot agent is Copilot given one job. There are two kinds: agents you talk to, which find and draft things when someone asks, and agents that run, which start on their own when something happens, do the checking, drafting and filing and then stop for a person to approve.
Is a Copilot agent the same as Power Automate?
No. Power Automate is best for repeatable, rule-based steps. An agent you talk to is better for a guided conversation or answers from approved material. An agent that runs uses both: Power Automate starts it and does the actions, and the agent handles the one step that needs judgement.
When should a business build its first agent?
Start when there is a repeated task, a clear audience, trusted source content, a named owner and a realistic way to test the answers before wider rollout.
Can Copilot agents use SharePoint content?
Yes, where the content is suitable, current and permissioned correctly. SharePoint clean-up and ownership are often part of agent readiness.
Do agents replace professional judgement?
No. Agents can support preparation, guidance, intake and knowledge work. Final client advice, regulated decisions and professional judgement should stay with qualified people.
How does FiveForward scope Copilot agent work?
FiveForward starts with the Readiness Check, which shows where Copilot can and cannot see the firm's data and names the two jobs to automate first. The install then checks permissions, puts the firm's templates and procedures in SharePoint, builds the agents with defined sources and boundaries, tests them with the people who will use them and hands over ownership, department by department.

Next step

Find out where Copilot can and cannot see your data.

Free, 30 minutes, online, with James. You leave knowing where Copilot can and cannot see your data, the two jobs to automate first and a rough install price.