“Should we just use Copilot, or do we need something built?” is still the question I am asked more than any other. When I first answered it in January the reply was simple: Copilot for general desk work, a custom tool for anything specific to your business. That answer is no longer good enough, and part of it is now wrong.
Two things moved the line in September 2026. Copilot Cowork gained an App skill that builds small interactive apps from a description. And Copilot Studio gained frontier models, GPT-6 Astra and Claude Fable 5.1, which means the model you would reach through a custom build is now also reachable inside your Microsoft tenant. Both of those eat into territory I used to describe as custom-only, and it would be dishonest to pretend otherwise.
Use Copilot for tasks that live inside the Microsoft Graph and are scoped to one person or one team. Build custom for workflows that cross systems, run at high volume, face people outside your company, or where the cost of each transaction matters. Most SMEs have both kinds of work, so most end up with both.
The rest of this post explains why the line sits there and how to place a given process on one side of it.
If the task lives in the Graph (email, calendar, Teams, SharePoint, OneDrive, the Office apps) and one person or one team is both asking for it and checking the result, Copilot is usually the right tool and a custom build is usually a waste of money. That covers more than it did a year ago:
The last one is worth dwelling on. For users who hold a Microsoft 365 Copilot licence, classic answers, generative answers and Graph grounding inside Copilot Chat or Teams do not draw credits. A conversational agent for licensed staff is close to free at the margin, quick to build, and governed by the permissions model you already have. I would not build that custom today.
Copilot sees what the Graph sees. If the process starts in a shared mailbox, needs a lookup in your CRM, and ends with a record in your finance system, most of the work is outside its natural reach. Connectors and plugins narrow the gap, and Anthropic’s Claudeforce now puts Claude inside Salesforce, so “off-the-shelf AI cannot reach your CRM” is no longer true as a general rule. But stitching three systems together with a defined hand-off at each stage is integration work whichever tool you start from, and a purpose-built service usually does it more reliably than an agent improvising the route each time.
Anything that runs thousands of times a month without a person starting each run is a pipeline, not a conversation. Pipelines want fixed behaviour, structured output and a cost that does not grow with every step the tool takes.
Copilot is licensed and designed for your own staff, working with their own permissions. A customer-facing assistant on your website, or anything that sends output to clients without a human reading it first, needs behaviour you can specify, test and pin to a model version. That is a build.
Copilot Credits charge per interaction: about 2 credits for a generative answer, 5 for an action, 10 for Graph grounding and 25 for an autonomous action, at US$0.01 a credit pay-as-you-go. A custom tool on the Claude API charges per token. Using the illustration from our Power Automate vs Copilot Studio vs Claude API comparison: classifying, checking and routing 3,000 inbound emails a month comes to roughly 130,000 credits as an autonomous Copilot Studio agent, which is five or six prepaid packs at US$200 each. The same job on Sonnet 5 is about US$21 of model spend plus modest hosting. The build has to be paid for, but at that volume the running cost differs by an order of magnitude.
Reverse the shape, 300 licensed staff asking an HR agent questions in Teams, and Copilot Studio wins just as clearly. Per-interaction pricing suits conversations. Per-token pricing suits pipelines.
A year ago, a team that wanted a small tracker, calculator or intake form with some logic behind it had two options: a spreadsheet nobody trusted, or a quote from someone like me. If the App skill does what Microsoft describes, a good share of those requests no longer need a developer, and I would rather tell you that than quote for the work.
Before you rely on an app built this way, I would ask four questions. Who owns it when the person who described it leaves? Where does its data live, and who else can see it? Does it need to talk to anything outside Microsoft 365? And if it produced a wrong answer, would anyone notice? If the answers are “the team”, “in our tenant”, “no” and “yes, quickly”, use it. If the app is going to become part of how you bill, quote or report to a regulator, treat it as software and build it properly.
With Fable 5.1 available in Copilot Studio, “we need a custom build to get the better model” is no longer an argument. Model quality has stopped being what separates the two routes. What separates them now:
I go through the model side of that decision in how to choose the right AI model.
| Scenario | Copilot | Custom build |
|---|---|---|
| Summarise a meeting or thread | ✓ Copilot — native, nothing to build | Overkill |
| Draft a client proposal | ✓ Copilot — in Word, with your files as context | Unnecessary |
| Prepare a weekly status pack from email, Teams and a tracker | ✓ Copilot Cowork — one person, inside the Graph | Only if it must pull from external systems |
| Small internal tracker or calculator for one team | ✓ App skill, where available in your tenant | Only if it becomes business-critical |
| HR or IT questions answered in Teams for licensed staff | ✓ Copilot Studio — near-free at the margin | Hard to justify |
| Triage 3,000 inbound emails a month, unattended | Possible, but metered on every run | ✓ Custom — per-token cost, fixed behaviour |
| Extract supplier invoices into your finance system | Partial; the finance system is outside the Graph | ✓ Custom — cross-system, structured output |
| Customer-facing assistant on your website | Not what Copilot is licensed or designed for | ✓ Custom — tested, pinned, auditable |
A small Copilot footprint for the knowledge workers who will use it, one or two Copilot Studio agents for staff-facing questions once licences are in place, and one or two custom tools for the heaviest, most repetitive workflows. Not Copilot for everyone, and not a custom build for everything.
Whichever side a process lands on, the groundwork is the same. Copilot acts with the user’s permissions and a custom tool acts with whatever its service principal has been given, so both depend on the permissions hygiene in our oversharing guide, and both need an owner for the budget. The question I ask every client has not changed: what does your team do every week that takes significant time, follows a consistent pattern and produces a predictable output? Those are the candidates for a build. You can see what that looks like in practice in our case studies.
If you are not sure which side of the line a particular process falls on, that is a half-hour conversation, and I would rather have it before you buy licences or commission a build than after.
Bring one process to a free 30-minute discovery call. I will tell you whether it belongs in Copilot, Copilot Studio or a custom build, and why. It is the starting point for our Microsoft 365 AI consultancy and our custom AI work.
Book a Free Discovery Call