If you run IT, you have probably been asked to “add an agent” to something this year, and more often than not the agent people mean is Copilot Cowork. Before you point it at a process, the sharper question is whether the task needs an agent at all, or whether a business app would do the job better and for a fraction of the running cost. Here is how we help clients decide, starting with the outcome and the price of the work rather than the technology.
We build both business apps and agents for clients on the Microsoft 365 stack, so we have no stake in talking you into either. What we care about is that you do the right thing for the outcome you want, at a cost you can predict. And in the last year, that second part has changed the whole conversation.
The build decision is now a cost decision
For years, deciding what to build was mostly about capability and one-off cost. You scoped an app, you built it, you paid for it once, and it ran on licensing you already held. Agents changed that math. A modern agent does not just cost money to build. It costs money every time it runs, because it bills on consumption.
That single shift is why “let’s put an agent on it” can be an expensive reflex. We made the broader case for why structured apps still matter in our blog: Why Apps Still Matter in an Agent-First World. This piece is the practical follow-up: how to choose between an app and an agent, and what each will actually cost you once it is live.
What agents, and Cowork in particular, are good at
Agents earn their place where the work is conversational and hard to script in advance. Answering questions, guiding someone to the right place, routing a request, pulling together an answer from information scattered across email, chat and SharePoint. Anything where the value is in reasoning over content rather than capturing it in a fixed structure.
It helps to picture agents on a spectrum of control. At the light end sits a declarative agent, which is closer to configuration than development. You describe what you want in plain language, ground it in your content, and it works, with no integration project behind it.
At the far end sits Copilot Cowork, the agent most teams are actually weighing up today. Cowork is built for complex, long-running, multi-step work, the kind that used to take a person hours.
Where Copilot Chat drafts or recommends, Cowork executes. You define the work, and it goes and does it, using whatever tools it needs to return a finished result.
That capability is real, and for the right task it is worth paying for. The trick is knowing which tasks those are, because Cowork is also where the meter runs hardest.
What a business app still does better
The balance tips toward an app as the work gets more structured. Several forms under different conditions, validation rules, approval flows, conditional logic on the back end, anything hands-on and repeatable. These are the jobs where you want people going to the same reliable place every time and capturing clean, consistent data when they get there.
A simple example makes the point. Nobody wants a fresh conversation with an agent every time they book time off. They want to open one screen, see their previous requests, check their remaining balance, and submit. That is an app’s job, and it does it the same way every time, which is exactly what an audit trail and a system of record need.
There is a cost angle here that is easy to miss. Well-structured data does not only keep you compliant. It makes any agent you decide to build later cheaper and more reliable, because the agent thrives when it can read dependable, structured information instead of reasoning and guessing its way to an answer on every request. The app behind the agent is often what keeps the agent affordable.
The question we ask before we build anything: does the task warrant an agent?
The cost question is where we spend real time with clients, because it is the part the Microsoft marketing hype skips, and Cowork is the clearest illustration of why it matters.
Cowork is not included in your base Microsoft 365 Copilot license. It sits on top of it. You pay the per-user Copilot license, and then Cowork bills separately on usage, in Copilot Credits, for the actual work it does.
A seat license includes zero credits. They are a separate budget line. Cowork reached general availability on June 16, 2026, and it does not run at all until an admin has configured usage-based billing and a spending policy. Tenants that had not set that up by July 1, 2026 have had their Cowork access suspended until they do.
One more thing worth knowing: if your tenant is managed by a Cloud Solution Provider, the billing has to be set up through your provider. Your own Azure subscription will not be accepted, so loop your partner in early.
How much a task costs depends on how much work it involves. Microsoft prices credits at roughly a cent each and groups tasks into rough tiers: a light task might run a dollar or two, a heavier multi-step task several dollars or more.
The cost driver most people miss is context. Every time Cowork reads your emails, calendar, SharePoint files or Teams messages to ground its answer, it consumes tokens, and in agentic work the volume it reads to produce a short result can be very large. A task that feels like a ten-second request can quickly cost a lot.
Contrast that with a well-built Power App, which usually runs on licensing you already hold, at a fixed and predictable cost.
So, the design question is simple. Does this task genuinely need an agent reasoning through it every time, or is it a structured job an app can do once, cleanly, for a known cost?
Pointing Cowork at a high-volume, repetitive task can rack up charges for work an app would have handled for a flat fee. And where an agent genuinely is the right call, the discipline is to match the effort to the job, keep an eye on how much context it pulls, and set spending limits before you roll it out, not after the first invoice.
Model choice is also part of the same discipline. Copilot now lets you pick between models and modes, from quick responses to deep reasoning, and in Copilot Studio you can even mix models within one agent, using a lighter model for routine steps and a heavier one only where accuracy matters.
A simple rule of thumb. The deeper the reasoning, the more it costs to run, so the rule is the same as everywhere else in this piece: use the lightest model that does the job well, and save the heavyweight reasoning for tasks that genuinely earn it.
A note on the numbers. Credit rates and license inclusions here are Microsoft’s published figures at the time of writing, and this area is changing quickly, so check the current documentation before acting on any of them.
Microsoft’s Copilot Studio pricing page has the credit rates and worked examples are in its Copilot Studio billing rates guidance, and its usage-based billing and cost management overview covers how Cowork billing and spending controls work. The direction of travel, that agentic work meters and structured apps do not, is stable even as the exact rates move.
Sometimes the answer is neither an app or agent
Some of the best outcomes we deliver are the ones where we talk a client out of building anything at all. Every so often, what looked like an app is really a SharePoint list with a built-in form, solved in an afternoon. Our rule is to find the lowest-cost, lowest-maintenance option that genuinely solves the problem, whether that turns out to be an agent, an app, or neither.
Usually, the answer is both, used deliberately
Apps and agents are not competitors. The strongest setups use each for what it does best.
The app becomes the engine: the structured forms, the validation, the system of record.
The agent becomes the front door, so instead of hunting for the right link, someone can simply ask Copilot for the tool they need and have the app open for them.
The pattern we recommend most often is to build the app for the structured, repeatable work, and reserve the agent, whether that is Cowork or a purpose-built one, for the genuinely open-ended tasks that justify the reasoning and the cost.
The point is to choose, rather than reaching for whatever is currently in fashion.
How we help you decide
We start with discovery, because the right build almost never reveals itself in the first meeting. We map how the work happens now, find where it slows people down, and weigh an app against an agent against the simplest option of all. Then we factor in what each will cost to run, not just to build, and we build the right one. If the right one is neither, we will tell you.
If you are weighing up an app or an agent for a process that is not working as it should, talk to us. We will help you decide before you spend.
FAQs
What is a business app in Microsoft 365?
A business app is a purpose-built tool, typically created with Power Apps, that gives people a structured place to do a repeatable task: forms to fill in, rules that validate the data, workflows that route approvals, and a reliable record of everything submitted. It behaves the same way every time, which makes it ideal for processes that need consistency, auditability, and clean data.
What is an agent, and what types are there?
An agent is an AI assistant that understands plain-language requests and works things out rather than following fixed screens. They sit on a spectrum. Declarative agents are the simplest: you configure one with instructions and knowledge sources, and it answers questions grounded in your content. Custom-engine agents go further, with their own logic and integrations for more specialized work. At the far end are agentic tools like Copilot Cowork, which take on multi-step tasks and carry them through to a finished result. As a rule, the further along that spectrum you go, the more capable the agent, and the more it costs to run.
Do agents cost more to run than apps?
Often, yes. Agentic experiences like Copilot Cowork, and agents built in Copilot Studio, bill on consumption, so they charge credits each time they run, and a complex task can cost far more than a simple one. A business app usually runs on licensing you already hold, at a fixed and predictable cost. The right choice depends on whether the task needs an agent’s reasoning or simply a reliable place to capture and process data.
We already pay for Microsoft 365 Copilot. Isn’t an agent free?
It depends on the type of agent. Interactive agents that a licensed Microsoft 365 Copilot user runs inside Teams, SharePoint, or Copilot Chat are generally covered by the license, within fair-use limits, so a simple internal Q&A agent usually doesn’t add to your bill.
Cowork is different. It sits on top of your Copilot license and bills separately on usage, so every task it runs consumes credits, even inside your own tenant and even for fully licensed users. As a rule, the more autonomous and multi-step the agent, the more likely it is to meter. “We already pay for Copilot” rarely means the agentic work is free.
Can an app and an agent work together?
Yes, and this is often the strongest design. The app handles the structured work and holds the data. The agent acts as a natural-language front door, so people reach the app just by asking for it in Copilot. You get the reliability of an app with the ease of conversation on top.
What if we don’t need either an app or agent?
That happens more often than you might expect. Sometimes a SharePoint list with a built-in form solves the problem for a fraction of the cost and effort of a custom build. We would rather point you there than sell you something you do not need.
Weighing up an app or an agent for a process that isn’t working? We build both, so we’ll give you a straight answer on which fits your outcome and your budget. Book a discovery call with our team and we’ll help you decide before you commit to spend.