AI & Automation
AI Agents & Assistants
Narrow jobs, done the same way.
An AI agent is software given a goal, a set of tools it may use, and permission to take several steps on its own until the job is done, as opposed to a chat window that only answers a question. happier IT builds narrow, permissioned agents for Canadian organizations of roughly 15 to 200 people, and says plainly when an ordinary workflow would be safer and cheaper.
Who it's for
The useful ones are much smaller than the demonstrations.
Four shapes work at this size. Everything grander is either a workflow in disguise or a research project you are funding.
An internal question desk. "What is our overtime policy", "who approves this", "which supplier do we use in Kelowna". An agent that searches your own documents and answers with the source attached takes a real load off whoever currently gets asked.
Triage on incoming work. Reading a request, classifying it, checking a system to add context, and routing it. Not deciding the outcome, deciding where it goes and what a person needs to know when they open it.
A first pass on a repetitive research job. Pulling the same six facts about a new customer from the systems that hold them, so the person doing the real work starts from something rather than nothing.
A drafting assistant with your material behind it. Quotes, service reports, standard letters, drafted from your templates and your data, then edited by the person who signs them.
The honest limit
An agent that takes several steps compounds its own mistakes: a wrong assumption in step two becomes a confident wrong action in step five, and nobody sees the middle.
So we build narrow, we let agents read widely and write rarely, and anything with a money, legal or safety consequence gets a person in the loop. That is a design rule, not a phase we grow out of.
What's included
How we build one so it can be trusted.
The model is the least interesting part. Everything below is about boundaries, evidence and knowing what it did.
-
A written, narrow job description
One task, one audience, an explicit boundary, and a defined response when a request falls outside it. Broad agents are less accurate and much harder to evaluate. Narrow ones are boring and they work.
-
Its own identity and its own permissions
An agent gets its own account in Microsoft Entra ID, the identity service behind your Microsoft 365 sign-ins, with only the access its job needs. Never a shared administrator login, and never a departing employee’s account, which is how these things quietly become invisible.
-
Read widely, write rarely
Reading data is low risk. Writing to a system, sending an email, or moving money is not. Write actions are enumerated one by one, and anything consequential drafts for approval instead of acting.
-
Grounding in your own content, with citations
Retrieval-augmented generation, or RAG, the agent searches your documents first and answers from what it found, quoting the source. Citations are not a nicety. They are how a person checks an answer in ten seconds rather than trusting it.
-
Tools that are real functions, not free rein
An agent can only do what it has been given a tool for: look up an order, check a schedule, create a draft ticket. Each tool is a specific function with its own validation, which is what keeps the behaviour predictable.
-
A logged trail of what it did and why
Every step, every source, every action, kept. When somebody asks why an answer was given six weeks ago, the trail has to exist. Without it there is nothing to investigate and nothing to improve.
-
Evaluation against real questions
A set of real questions with known right answers, scored before launch and after every change. Including the ones it should refuse, because a well-behaved refusal is a correct answer.
-
Built on Copilot Studio or on open tooling, deliberately
Microsoft Copilot Studio is the obvious route where the work lives in Microsoft 365 and agents can be published into Copilot. Where an agent needs to reach systems outside it, or to run inside your own environment, we build on other tooling instead. See <a href="/ai/microsoft-copilot/">Microsoft 365 Copilot</a>.
How it works
One agent, one team, one measurable job.
The failure mode is an assistant for everything, launched to everybody, evaluated by vibes.
-
Pick the narrowest useful job
Usually a question people ask each other constantly, or a triage step that delays everything behind it. We write down what a good answer looks like, what it must never do, and how we will tell the difference.
-
Build it read-only first
The first version answers and drafts but changes nothing. A fortnight with one team, with every answer reviewable, tells you what it is actually good at. Write permissions are added afterwards, one at a time, each with a reason.
-
Measure, widen slowly, keep watching
Re-run the evaluation, look at the questions it handled badly, fix the content or the boundary rather than the prompt alone. Then a slightly wider audience. An agent nobody reviews after month one is an agent nobody should be relying on.
What it costs
A build cost, then a running cost that moves with use.
The second one surprises people. Agents take multiple steps per request, so usage costs scale faster than a chat tool.
Building one A narrow question-answering agent over a defined document set is a small project. One that reaches three systems and takes actions is not.
Platform costs. Microsoft Copilot Studio bills in Copilot Credits, the unit that replaced messages on 1 September 2025, available as a prepaid monthly capacity, pay-as-you-go through an Azure subscription, or a one-year prepurchase. Prepaid capacity is enforced monthly and unused credits do not carry over. Certain agent uses by people who already hold a Microsoft 365 Copilot licence are zero-rated.
Running it, monitoring, reviewing what it did, updating content as things change, sits inside managed IT
Why usage costs climb
A chat answer is one call. An agent that searches, checks a system, reasons about the result and drafts a reply might be five or ten. Multiply that by however many people find it useful.
We estimate it before you commit and alert on it after. A budget cap and a usage alert are part of every agent we build.
Why us for this
Where an agent is the wrong tool.
If the steps can be written out in advance, do not build an agent. Build a workflow. It will be cheaper, faster, auditable, and it will do the same thing every time, which is what you actually want from a process that already works.
The other honest limit is accountability. Canada’s privacy commissioners, in their joint principles for generative AI, are explicit that accountability for decisions rests with the organization rather than with any automated system. An agent does not absorb responsibility. Whoever would have owned the outcome still owns it, which is why anything consequential drafts for a person rather than acting.
What we have built in this space: connections into software APIs that pull real-time information from several locations into one view, candidate processing for a human resources team, and automated skills coaching for professional development. Where an agent touches personal information, the questions on AI governance and policy apply before the build starts, not after.
Go deeper
- Business process automation The better answer when the steps are known.
- Custom AI development RAG, evaluation and the build underneath.
- Microsoft 365 Copilot deployment Where Copilot Studio agents are published.
Questions
What people ask before they sign anything.
What is an AI agent?
An AI agent is software given a goal, a set of tools it is allowed to use, and permission to take several steps on its own until the goal is met. The difference from a chatbot is action: a chatbot answers, an agent can look something up, check a second system, and produce or do something as a result. The difference from an automation is that the steps are not written out in advance, the agent decides the order, which is exactly why it needs narrow boundaries.
What is the difference between an AI agent and an automation?
An automation follows a sequence you defined and does the same thing every time; an agent works out its own sequence from a goal. Automation is predictable and auditable, which is what you want for a process that already works, route this invoice, provision this account. An agent suits work where the path genuinely varies, such as answering an unpredictable question from your own documents. If you can draw the process as a flowchart, build the flowchart. See business process automation.
Are AI agents safe to give access to our systems?
Only with their own identity, their own narrow permissions, and a logged record of everything they do. An agent should have an account in your identity system with the minimum access its job requires, never a shared administrator login and never an old employee’s account. Read access is low risk. Write actions should be enumerated one at a time, and anything with a money, legal or safety consequence should draft for approval rather than act.
What can an AI agent actually do for a small business?
Four things reliably at 15 to 200 people: answer internal questions from your own documents with the source cited, triage and route incoming requests, gather the same set of facts about a customer from several systems, and draft routine documents from your templates. What it will not do is run a department, make a judgement call you would want defended, or replace a role. The value is measured in interruptions removed from specific people’s days.
What does it cost to build an AI agent?
There are two numbers. The running cost moves with usage, because an agent may make several model calls per request. Copilot Studio, for example, bills in Copilot Credits, either as prepaid monthly capacity, pay-as-you-go via Azure, or a one-year prepurchase. We estimate the running cost before you commit and put a usage alert on it after.
How do we stop an agent giving a wrong answer confidently?
Ground it, cite it, narrow it, and measure it. Grounding means it answers only from your retrieved documents rather than from general knowledge. Citations let a person check in seconds. A narrow scope with an explicit refusal for anything outside it prevents the most common failures. And an evaluation set of real questions with known right answers, re-scored after every change, is what turns "it seems fine" into something you can rely on. You will not get to zero, which is why the consequential paths keep a person in them.
Do we need Copilot Studio or something else?
Copilot Studio if the work and the content live in Microsoft 365 and you want agents published where people already are, in Copilot Chat, Teams or SharePoint. Something else when the agent has to reach systems outside Microsoft 365, when it needs to run inside your own environment, or when the usage pattern makes credit-based billing awkward. We decide per agent rather than standardising on one platform.
Related
Where to go next.
Want to know what this would look like for you?
A 30-minute call. No slides, no audit fee, no obligation. We ask what is breaking and tell you honestly whether we are the right fit.