Internal questions
Policies, procedures and how-tos answered from the source documents, with a link to where the answer came from.
An agent with a job. And a point where it stops.
Good agents do not pretend to know everything. They gather context, perform bounded steps and hand decisions to people when the stakes or uncertainty rise.
Agent: HR questions
- Its job
- Answer leave and policy questions
- It knows
- The handbook and leave policy on SharePoint
- It may
- Raise a leave request for approval
- It must not
- Approve leave or discuss pay
- It hands over
- Complaints, and anything it isn’t sure of
Useful, not chatty. An agent should finish a job.
The agents worth building do one thing well: answer the question staff ask the HR inbox every week, raise a request properly, or check a policy before something is sent. We start from that job, not from the technology.
Useful agents are specific.
- A vague brief. “Build an AI assistant” gives nobody a clear success measure.
- Too much authority. A bot can create consequences faster than a person can correct them.
- No escape route. Uncertain answers and failed integrations need a named owner.
Where it helps
Three kinds of agent worth building.
Requests that need doing
An agent that gathers the details, checks them and raises the request through a flow, instead of starting an email chain.
A first line for customers
Common questions answered on your website, with anything unusual passed to a person with the conversation attached.
Give the agent four responsibilities, not unlimited freedom.
Boundary-led agent design
- Receive intentA defined request and user identity
- Gather evidencePermitted sources and tools
- Draft or prepareStructured output with uncertainty
- Hand overA personHuman reviews or approves
See it run: Tina sketches this workflow, step by step, and shows which parts a person keeps.
What we build
- Agent design: its job, its knowledge, its actions and its limits
- Knowledge from SharePoint, documents and websites you choose
- Actions through Power Automate and connectors, with approvals where they matter
- Hand-over to a person, with the context, when the agent shouldn’t decide
- Testing against real questions, including the awkward ones
- Publishing to Teams, your intranet or your site, with usage monitored
- Evaluations and logs, so you can see what the agent did and why
What stays with your team
- Anything that commits the business
- Exceptions and complaints
- What the agent may know and do
AI prepares. People decide. Routine actions run only within limits you agree.
An agent with a job and a boundary.
Illustrative architecture — the shape, not a fixed design
What you already run
- Your knowledge sources
- Service desk or CRM
- Teams or your website
What carries the work
Copilot Studio the agent, its topics and its limits
Power Automate the actions it is allowed to take
Dataverse what it may read, and what it may not- Hand-off to a person, with the conversation attached
What stays with people
- Anything outside the agent's remit
- Judgement calls
- The final word with a customer
Microsoft product icons are used under Microsoft’s icon terms for architectural diagrams. Product names belong to their owners.
Usually built with
- Copilot Studio
- Power Automate
- Dataverse
- APIs & webhooks
Technologies we work with. Product names and logos belong to their owners, who do not endorse Boring Future.
The evidence. Each with where it came from.
Founder experience · employed role
700 work orders
Same underlying case; approximate preparation saving, human review retained.
A preparation agent built around a real review queue.
- The problem
- A team needed to prepare a large work-order backlog for expert review without removing the expert from the process.
- What was built
- Combined automation and agent-style summarisation to prepare structured information from roughly 700 work orders.
- What changed
- Approximate preparation fell from 35 hours to 30 minutes.
The same project is shown on Power Platform & Power Automate and Facilities & maintenance: one piece of work, not 3 separate cases.
When not to build an agent
- If the answer is a link to one page, you probably don’t need an agent.
- Agents shouldn’t approve, refund or promise. They prepare; a person decides.
- If the knowledge is out of date, the agent will be too. Fixing the source comes first.
How it works
- Free conversationFree
- Discovery & Design£2,500 + VAT, fixed
- CompleteFixed scope and fee
- OperateOptional, monthly
Running a business of 5 to 50 people? Essentials starts with a £395 + VAT half-day session
Questions. Anything else, just ask.
Ask us directlyIs Copilot Studio included in our Microsoft 365 licence?
Some Copilot Studio use comes with certain Microsoft 365 and Copilot licences, and heavier use is billed separately. Microsoft changes the details often, so we confirm what your agreement covers before we design anything.
Can it act in our business systems?
Potentially, through approved integrations such as Power Automate flows, connectors or an API, and within explicit permissions and confirmation rules.
How do we prevent made-up answers?
We constrain tasks, use reliable sources, test against real questions, and stop or hand over to a person where confidence is not sufficient.
Do you build agents outside Microsoft?
Yes. When Copilot Studio isn’t the right fit, we build custom agents and decision steps.
Is an agent the right solution for every task?
No. If a simple rule or conventional automation is more dependable, we would recommend that.
Where it’s used. Where to go next.
The cleverest thing an agent can do is know when to ask.
Start a free 30 to 45 minute conversation. No deck, no obligation.