October 2026 update — Copilot Credits are now spent while you build, and you can cap them per agent. First, credits are now consumed while makers build agents in Copilot Studio, not only when agents run. A pilot has a cost before it has a single user, so budget for the build and test phase as well as production. Second, the controls have caught up: limits can now be set per agent, and there is a tenant-wide view of credit spend. That turns “do the credit maths before the pilot” from a spreadsheet exercise into something you can enforce. Set a limit on each agent before it goes live, and give one named person the job of reading the tenant spend view every month. None of this changes which engine wins which process. Per-interaction pricing still suits conversations and per-token pricing still suits pipelines. It does remove the main excuse for an agent bill nobody saw coming. The Copilot Business introductory rate also ended on 30 September, and the prices below reflect that.
Every UK SME on Microsoft 365 that decides to automate a business process with AI now faces the same fork in the road. There are three credible ways to do it, all of them live in your tenant today, and they are priced so differently that the wrong choice can cost ten times the right one for exactly the same outcome.
The three options are Power Automate (the workflow engine you probably already own), Copilot Studio (Microsoft’s agent builder, billed in Copilot Credits), and a custom tool built on the Claude API (or another frontier model), hosted in Azure and wired into Microsoft 365 through Graph. Vendors will tell you their option is always the answer. It isn’t. Each one wins a specific class of process and loses the others, and the September 2026 pricing picture has shifted enough — Anthropic locked in Sonnet 5’s introductory rate, Microsoft is about to strip AI Builder credits out of Power Automate licences — that it’s worth redoing the maths.
This is the practical decision guide we walk clients through. It builds on the broader Copilot vs custom AI question, but zooms in on the one thing that matters once you’ve picked a process: which engine should run it?
Power Automate runs rule-based flows. Something happens (a form is submitted, an email arrives, a row changes in a SharePoint list, a file lands in a folder), the flow executes a fixed sequence of steps across any of several hundred connectors, and it stops. It does not reason, and that is the point — it does the same thing every time, which is exactly what you want for approvals, notifications, data movement and anything with an audit trail. Licensing is per user: Power Automate Premium is around £12 per user per month at UK list price, and many Microsoft 365 plans include a “seeded” version that covers standard connectors.
Copilot Studio builds AI agents — conversational ones that answer questions grounded in your SharePoint, Teams and Dataverse data, and increasingly autonomous ones that are triggered by events and decide for themselves what to do. Agents can call Power Automate flows as tools, so the two products are complementary rather than rivals. The critical difference is billing: there is no per-user licence. You pay in Copilot Credits, pooled across the tenant, at US$0.01 per credit pay-as-you-go or US$200 per month for a prepaid pack of 25,000. Each interaction burns a variable number of credits — roughly 1 for a classic scripted answer, 2 for a generative answer, 5 for an agent action, 10 for grounding on Microsoft Graph tenant data, and 25 for an autonomous action triggered with no user in the loop. Those stack. A single well-grounded generative answer is 12 credits; an autonomous agent that runs three actions and a grounded summary is pushing 90.
One wrinkle that changes the economics entirely: if a user already holds a Microsoft 365 Copilot licence (£16.10 (US$21) per user per month at UK list for the Business SKU since 1 October 2026; the introductory £13.80 (US$18) rate ended on 30 September), their use of agents inside Copilot Chat or Teams for classic answers, generative answers and Graph grounding does not draw down credits. For a licensed tenant, conversational agents are close to free at the margin. For an unlicensed one, they are metered on every turn. Since September 2026 Copilot Cowork also shows the credit cost of a task before you run it, and carries a reasoning-effort slider that trades a cheaper, faster pass against a slower and more thorough one — which makes the budget conversation considerably easier to have.
The third option is to write the process yourself — typically a small Azure Function or App Service in a UK region that receives a trigger, calls the Claude API to do the reasoning (classify, extract, summarise, draft), and writes the result back to Microsoft 365 through the Graph API. Nothing is metered per user or per interaction; you pay per token. As of September 2026, Claude Sonnet 5 costs US$2 per million input tokens and US$10 per million output tokens — Anthropic cancelled the increase to US$3/US$15 that had been due to take effect in September, confirming in August 2026 that the launch rate is permanent. Haiku 4.5 is US$1/US$5 for lighter work, Opus 5 is US$5/US$25 for hard reasoning, prompt caching cuts repeated context to a tenth of the base price, and the Batch API halves everything for jobs that can wait an hour. For the kind of volumes an SME actually runs, model spend is tens of pounds a month. The cost is the build and the maintenance, not the tokens. We covered the model-and-vendor choice in the Claude API buyer’s guide; here we’re only concerned with when a custom build beats the two Microsoft options.
Microsoft has confirmed that the AI Builder credits seeded into Power Automate Premium and Power Apps licences will be removed on 1 November 2026, for all new and existing customers, Enterprise Agreements included. The AI Builder capacity add-on was already withdrawn from sale to new customers in November 2025. From November, any AI step inside a flow — invoice extraction, text classification, a GPT prompt action — draws on Copilot Credits instead, and when no credits are available the step is blocked.
The practical consequence: a lot of SMEs have flows that quietly use a bit of AI Builder for document processing and have never paid for it separately. Those flows will start failing or start costing money on 1 November. Audit them now. For each one, decide whether the AI step is worth buying Copilot Credits for, or whether it is cleaner to move that single step to a direct model call and leave the rest of the flow as it is. That hybrid — Power Automate for the plumbing, a custom endpoint for the thinking — is often the most economical shape of all.
Take an illustrative (not client-specific) process: a services firm receives around 3,000 inbound emails a month that need classifying by type, checking against an internal knowledge base, and routing to the right team with a two-line summary. No human conversation, just unattended processing.
Power Automate alone: the routing and summary need language understanding, so this needs AI Builder — which from 1 November draws Copilot Credits. Rough order: a few credits per email for the AI steps, so a few hundred pounds a year in credits on top of Premium licences for the makers. Workable, but the classification quality of a generic prompt action is the weak point.
Copilot Studio autonomous agent: each email is an autonomous trigger (25 credits) plus a grounded generative step (12) plus an action to route (5) — call it 40–45 credits per email. At 3,000 emails that is roughly 130,000 credits a month: five to six prepaid packs, so around US$1,000–1,200 per month, or more on pay-as-you-go. Excellent governance and zero code, but you are paying a conversational-agent price for a batch job.
Custom Claude API tool: 3,000 emails at perhaps 2,000 input and 300 output tokens each on Sonnet 5 is about US$21 in model spend, half that with batching, plus perhaps £20–40 a month of Azure hosting. The build is the cost — typically four to eight weeks of consultancy — but the run cost is an order of magnitude below the agent and the classification quality is whatever you engineer it to be.
Reverse the shape of the process — 300 staff asking an HR agent questions in Teams, all with Copilot licences — and Copilot Studio wins just as decisively. The lesson is not “custom is cheaper”. It is that per-interaction pricing suits conversations and per-token pricing suits pipelines, and most SMEs have both kinds of process.
Whatever you choose, the governance layer sits above the tool: the same AI usage policy, the same data classification, the same permissions hygiene from our oversharing guide. A Claude-based tool calling Graph inherits exactly the permissions you give its service principal; a Copilot Studio agent inherits the user’s. Both are governable. Neither is governed by accident.
Most SMEs end up with a deliberate mix: Power Automate for the deterministic plumbing, a Copilot Studio agent or two for staff-facing Q&A once Copilot licences are in place, and one or two custom tools on the Claude API for the high-volume, judgement-heavy processes that justify a proper build. The mistake is picking one tool and forcing every process through it — usually because it is the one somebody already knows.
Do the credit maths before the pilot, not after the first invoice. And diarise 1 November for the AI Builder change — it is the sort of licensing shift that surfaces as a broken flow at 9am on a Monday.
We map UK SME processes to the right automation tool — Power Automate, Copilot Studio or a custom Claude build — with the credit and token maths done before anything is bought. It’s the core of our Microsoft 365 AI consultancy. Get in touch or book a 30-minute call — no sales theatre.
Book a Free Discovery Call