Should You Wait for Microsoft to Sort Out Agent Pricing? An Honest Answer

Should you wait for Microsoft to sort out agent pricing before building anything? It depends, but mostly no.

That is the short answer. The longer one gives the case for waiting a fair hearing first, because it is a better case than the people selling agents tend to admit, and I am one of the people selling agents. I design and build them in Copilot Studio for a living, so I have an interest in your answer. I also spend a fair part of every week reading Microsoft’s licensing documents, and I have quoted figures on this site that Microsoft moved a month later. If anyone has felt the ground shift, it is me. So take what follows as one practitioner’s view, stated plainly enough that you can disagree with it.

The case for waiting

Concede the facts, because they are facts.

Pricing has changed three times in twelve months. The currency changed in September 2025, when Copilot Studio’s messages became Copilot Credits. The billing model changed in August 2026, when Copilot Studio gained three harnesses and the new default one went onto usage-based billing. Then on 1 September 2026 every agent and workflow built before 3 August moved onto credits whether its owner had noticed or not. A fourth change lands on 1 November, when the AI Builder credits seeded in Power Automate Premium and Process licences disappear. Anyone who says the pricing has been stable is not reading the documents.

The new experience is priced in ranges, not rates. The classic harnesses have a per-feature rate card: so many credits per answer, per tool call, per hundred flow actions. The GitHub Copilot harness and Copilot Cowork do not. Microsoft’s September 2026 Credits Guide gives 100 to 300 credits for a light run, 300 to 500 for medium and more than 500 for heavy, and labels those rough planning estimates. A firm can budget a Cowork-style agent to within a factor of two or three. It cannot budget it to the pound.

Building now costs credits too. On the new harness and in Cowork the meter runs while you build: natural language authoring, preview and evaluation all draw credits. A firm experimenting on the new experience pays to learn. We set out what the new engine changes and what it costs when it landed, and that point has not softened.

Early buyers of the capacity pack paid for credits they never used. The 25,000 credit pack at $200 a month resets monthly with no carry-over. Firms that bought packs before they had usage data spent real money on credits that expired. That is not a hypothetical. It is a pattern I have seen on invoices.

None of that is unreasonable grounds for caution. If you have read that list and decided to wait, you are not being foolish. You are being careful about a platform that has given you reasons to be.

The case for building now

Now the other side, which I find stronger for firms of the size I work with.

Most first agents live on the covered harnesses. A knowledge agent that answers procedure questions, a process agent that drafts a standard letter from a form, an assistant in Teams that helps a new starter. That is what a 30-person firm builds first, and it belongs on the standard or Copilot Chat harness. There, internal users with a Microsoft Copilot licence draw no credits within fair use. The metered risk barely touches the work most firms actually want done.

The classic rates are published and small. Twelve credits for an answer grounded on Microsoft 365 content, 2 for an answer from files or a website, 5 per tool call. A firm can model its first agent to within a few pounds a month before building it. I have done the sum for an imaginary firm in what a first agent costs a 30-person firm in year one. The credits line is the smallest of the seven.

The worst case is a number you chose. Pay as you go at $0.01 a credit, with a spend threshold set in the Microsoft 365 admin centre before publish, means the bill cannot exceed the cap without someone deciding to raise it. No commitment, no pack, no annual pre-purchase. The full set of prices and the controls around them are on our Copilot Credits reference.

Waiting has a cost that never appears on a Microsoft invoice. The work keeps being done by hand. In builds I have delivered for firms of this size, a work in progress report that took managers up to half an hour became a five-minute conversation with an agent. Letters of engagement that took up to an hour became a form and a review. Every month of waiting is another month of that, and no pricing change Microsoft could make would cost a firm as much as the hours it spends not building.

The lock-in fear is overstated for this kind of build. Agents built inside Power Platform solutions are portable between environments, the logic is documented at Design and the knowledge sources are the firm’s own SharePoint. If it ever came to it, the process map and the content leave with you. Readers who want the alternative costed properly should read our Copilot Studio vs n8n comparison, which is not a foregone conclusion.

My actual position

Three things, and the second is the one I hold most strongly.

Microsoft has largely done the sorting. The moves in the list above were not random. They were one transition, from a flat message price to a metered currency that can charge for agentic work in proportion to what it costs to run. Currency first, then the billing model, then the legacy estate moved across. What is live today is the model, not a preview of it. There will be edits, the ranges will tighten as Microsoft gets usage data and I will not predict specific price moves in either direction. But the shape is settled.

The real pricing risk is underneath Microsoft, not inside it. This is my view rather than a documented fact, so weigh it accordingly. The consumer subscriptions from the frontier model providers are, I believe, priced below the cost of the work they deliver, because the companies behind them are spending investor money to win market share. At some point those prices rise. Copilot, by contrast, already charges roughly what the underlying model calls cost. My own illustration from running both: the work I get for an £18 a month consumer subscription to a frontier model would cost me £20-odd in Copilot Credits in Cowork. A handful of medium tasks a week gets there. That is not Copilot being expensive. That is Copilot being priced at cost while the subscription is subsidised. A firm that builds around true prices now has no shock coming later. A firm that builds its habits around a subsidised subscription does.

So build on the parts that are priced and covered, keep the metered parts small until you know your usage, and do not put off a capability because the meter might move a penny. That is the whole position. It is not “build everything now”. It is “build the covered things now and measure the metered things before you scale them”. The ten habits for keeping credits under control and the effort selector exist because the measuring is the easy part once someone decides to do it.

What building carefully looks like

Five lines, in order.

1. Start on the classic harness. Save the GitHub Copilot harness for work that needs the loop.
2. Licence the people who will use the agent most. Let the rest run on the meter.
3. Stay on pay as you go with a deliberately low spend cap for the first month.
4. Check every connector’s tier before the build starts, not during it.
5. Re-run Microsoft’s estimator every quarter, because the inputs move.

The three numbers to check before your first agent goes live is the working version of that list, written for the person who ended up as the firm’s admin by default.

The last word

My honest position is that prices for AI everywhere are more likely to go up than down over the next two years, and that Microsoft’s are the ones already closest to the real cost of the work. If a firm is going to build agents on any platform, the one that already charges the true price is the safer bet, not the riskier one. That is not a guarantee. Nobody who tells you they know where AI pricing will be in 2028 is being straight with you, and I am not going to start.

What I can offer is the number for your firm rather than mine. A free consultation is enough to work out what your first agent would actually cost to build and run, on the harness it belongs on, with the cap set where you would set it. Then decide whether to wait, with the figure in front of you rather than the fear.

Sources checked

Last checked: 12 September 2026. Microsoft prices are US dollar list at September 2026 and subject to change.

• Microsoft, Copilot Credits Guide, September 2026
• Microsoft Learn, standard harness licensing in Copilot Studio
• Microsoft Learn, usage-based billing on the GitHub Copilot harness
• Anders Jensen, Copilot Studio pricing 2026
• PlugJunction, the November 2026 AI Builder change