The stacks
One method, three runtimes.
We take no licence margin from any vendor, so nothing pushes the recommendation in a particular direction. Five questions decide which runtime you get.
01
Where must the data stay?
Your Microsoft tenant, an EU cloud, or your own servers and nothing else.
02
What do you already pay for?
Microsoft 365 E3 or E5 with Copilot seats changes the economics completely.
03
Judgement or plumbing?
Reading and deciding is a different problem from moving data between systems.
04
What volume runs through it?
Under 2,000 runs a month, anything works. Above 10,000, unit pricing decides.
05
Who owns it after go-live?
Your IT team, or nobody. If nobody, the answer is the stack with fewest moving parts.
Below, each stack with the case for it and the case against. The second one is the part worth reading.
01 / Microsoft
Microsoft
Copilot Studio · Power Automate · Azure AI
Choose it when
You are on Microsoft 365 E3 or E5, you have an IT function, and the agent needs to sit where the work already happens — SharePoint, Teams, Outlook. Also when governance, DLP and tenant-boundary data handling are formal requirements rather than preferences.
What gets built
- Copilot Studio agents published to Teams and M365 Copilot
- Power Automate flows with proper error handling and retry
- AI Builder and Azure OpenAI prompts
- Azure Document Intelligence for OCR past AI Builder’s size limits
- Dataverse solutions packaged with environment variables, so dev to production is a configuration change rather than a rebuild
What it costs you
- Copilot Studio, prepaid
- $200/month for 25,000 credits
- Copilot Studio, pay as you go
- $0.01 per credit via Azure
- M365 Copilot licence
- $30/user/month enterprise
Credits per action run 1 for a scripted answer, 2 generative, 5 an agent action, 10 tenant-graph grounding, and 100 per ten premium-model responses. That hundredfold spread is why Copilot Studio estimates are unreliable until the agent design is fixed. The part most buyers miss: internal agent interactions by users who already hold a Copilot licence consume no credits at all.
Do not choose it when
You are on Google Workspace. The work is high-volume background processing where credits compound. You need to choose your own model.
02 / Claude
Claude
Anthropic API · MCP · Agent SDK · Claude Code
Choose it when
The work needs judgement rather than plumbing — reading long documents, comparing and evaluating, drafting, extracting structure where the rules are decisions rather than fixed fields. Also when you want your own team using these tools properly.
What gets built
- Agents on the Agent SDK, running where you choose
- MCP servers exposing your systems to the model under controlled permissions
- Your own procedures packaged as reusable skills
- Evaluation harnesses, so output quality is measured rather than asserted
- Claude Code adoption for your developers
What it costs you
- API
- Token-based, no per-seat floor
- Scaling
- With document volume and output length
- Quoting
- After a measured sample, never before
Cost here is predictable once a real sample has been run, and unpredictable before that. We measure first and quote second, which is slower to a number and faster to a correct one.
Do not choose it when
Your compliance position rules out processing outside your own infrastructure and no data processing agreement will satisfy it. Or the task is pure plumbing with no judgement in it, where n8n is cheaper and simpler.
03 / n8n
n8n
Self-hosted, in your own infrastructure
Choose it when
You are not on Microsoft, the budget is tight, volume is high enough that per-task billing dominates the cost, or the data genuinely cannot leave your servers.
What gets built
- Workflows and agent nodes, deployed into your own n8n instance
- Hosting on your infrastructure or an EU VPS you control
- Zapier only where an integration exists nowhere else, or where a first result in days matters more than unit cost
What it costs you
- n8n Cloud Starter
- €20/month annual, 2,500 executions
- n8n Cloud Pro
- €50/month annual, 10,000 executions
- Self-hosted
- Free software, €5–20/month server, unlimited executions
- Zapier Professional
- $19.99/month annual, 750 tasks
n8n bills one execution per workflow run regardless of how many steps it contains. Zapier bills per step, with overage at 1.25× and workflows paused at 3× the plan limit. On a ten-step workflow at volume that difference is an order of magnitude. Note that SSO, role-based access, git-backed environments and external secret stores are n8n Enterprise features — if your security team will ask for them, that belongs in the budget from the start.
Do not choose it when
You need an agent inside Teams or M365 Copilot for people who will not open another tool. Nobody on your side can own a server.
Work it out
Answer the five questions.
Same five questions we would ask on a call, scored the same way. It runs entirely in your browser — nothing is sent, stored or tracked.
01 / 05
Where does the data have to stay?
The single constraint that rules stacks out fastest.
What do you already pay for?
Existing licences change the economics more than anything else here.
What kind of work is it?
Reading and deciding is a different engineering problem from moving data.
Roughly how many times a month would it run?
Below a few thousand, unit pricing barely matters. Above it, it decides.
Who owns it after go-live?
If the answer is nobody, the right stack is the one with fewest moving parts.
Result
Best fit
What it costs to run
What to watch
What drove this
Nothing you selected was sent anywhere. The link above opens your own mail client with the answers written into it, and you decide whether to send it.
A recommendation is not a quote.
Five questions narrow it to one stack. Working out whether the task is worth automating at all takes a proper look at how it is done today.
Book a 30-minute call