Microsoft is trying to sell business buyers on a simple idea: Copilot can deliver similar AI help for less money than competitors. At its Ignite conference, Judson Althoff, Microsoft's commercial business chief, said companies pick Copilot over Anthropic's Claude and OpenAI's ChatGPT because it costs less to run. He pointed to a study Microsoft commissioned from research firm GigaOm.
That matters because AI spending has moved from "interesting experiment" to "real budget line." Companies are starting to cap employee usage, and finance teams want to know whether licenses actually pay off. Still, a comparison funded by a vendor isn't the same as an independent estimate of what your company will spend.
So the better question isn't "Which vendor has the lowest headline number?" It's "Which tool ends up costing less for your day-to-day work?" That includes licenses, usage, integration, oversight, and the time employees save (or lose) correcting outputs. This article breaks down what Microsoft's claim does - and doesn't - show, then outlines a practical way to compare enterprise AI costs before you commit.
What does Microsoft's Copilot cost claim actually show?
Microsoft's core argument is that Copilot can provide comparable AI assistance for less than what companies would spend using alternatives. At Ignite, Althoff said, "Anything you can do in Claude or ChatGPT, you can do better with Copilot, and for less." It's both a product comparison and, plainly, a sales message. Microsoft has reportedly told its salespeople to frame Copilot against Claude and ChatGPT since July.
The cost figure comes from a GigaOm study commissioned by Microsoft. The reported summary says the study estimated $205 per month on "basic work" with the competing option(s), while Copilot was lower for the same work. But the underlying report isn't fully provided here - no full assumptions, methodology, or line-by-line breakdown. So that $205 should be treated as the outcome of a specific comparison, not a universal "monthly cost for Claude/ChatGPT users."
And that distinction matters. A study can be careful and still reflect choices about which tasks were tested, how often those tasks occur, which model or plan was used, and what it counts as "cost." Since Microsoft paid for the research, buyers should ask for the details before using the headline number in a budget or procurement decision.
There's also the adoption angle, which Microsoft cites in passing. Microsoft said Copilot had more than 30 million paid users in June, compared with 450 million business users of its office software. The company also said daily users grew fivefold over a year. That suggests adoption is rising - but it doesn't prove Copilot is cheaper per completed task, or that adoption automatically translates into measurable savings.
For buyers, the clean way to treat the claim is as a hypothesis: Copilot may be less expensive for certain workloads, especially where a company already standardizes on Microsoft productivity tools and prefers predictable subscription costs. It doesn't establish that Copilot is the cheapest option everywhere, for every department, task, or usage pattern.
What to ask before relying on a vendor study
If you want to compare the study's result to your own budget, ask for the assumptions behind it:
- Which products, plans, and model versions were compared?
- What tasks were tested, and how many times were they performed?
- Did the cost include only software access, or also setup, administration, and support?
- How did the study compare quality and completion time?
- Were users already familiar with one of the products?
If those details are missing, the study can still hint at a possible cost advantage. It just can't replace a pilot using your employees, your tools, and your actual work.
Why do the $205 and under-$75 figures differ?
The most obvious pushback comes from Anthropic. Anthropic says non-coding Claude Enterprise users spend under $75 per month on average. That's far below the $205 monthly figure Microsoft cites. These numbers aren't guaranteed to be contradictory, because they may refer to different groups, plans, tasks, or definitions of "spend."
For instance, "spend" could mean consumption-based charges in one analysis, while another might model a broader assumed cost of using a product. A study could focus on heavier users doing a defined set of tasks, while a vendor average could include lots of light users. Coding workloads can also drive higher resource use than everyday writing or summarizing. And Anthropic's statement explicitly excludes coding, so you shouldn't assume it applies to engineering teams without more data.
What's also missing is enough context to normalize the comparison. Without the details, you can't confidently say which figure is the "real" number for your business. A low average can hide heavy users, while a higher benchmark might assume usage levels your staff never reach.
The deeper issue is that enterprise AI spend isn't one simple bill. A flat subscription can be predictable for finance teams, but it doesn't necessarily reflect how much value each user actually gets. Usage-based pricing might be efficient for occasional work, but it can make total spend more sensitive to volume and model choice. Either way, a small group of high-use employees can drive a disproportionate share of the bill.
Microsoft's own experience illustrates why usage economics can surprise people. Microsoft reportedly reduced Claude Code licenses for its engineers in May because of token costs. That doesn't tell you what other companies should pay. It does reinforce why teams should track real usage rather than assume a subscription price - or a model's public sticker price - settles the question.
Separate the questions hidden inside "cheaper"
When a vendor says "cheaper," it can mean different things. Clarify which claim they're making:
- Lower list price: the plan or license fee is lower.
- Lower usage cost: employees consume fewer paid resources to complete the same work.
- Lower cost per successful task: the tool produces an acceptable result with less time, rework, or model spend.
- Lower total cost of ownership: licensing, implementation, governance, and support add up to less over time.
A product can look good on list price but still lose on total cost if it creates extra manual fixes or integration work. The reverse can also be true: a higher-priced service may be worth it if it completes valuable work faster or reliably handles tasks the cheaper tool struggles with.
How should you compare enterprise AI costs?
Start with a defined set of workloads, not a vendor demo. Pick a handful of tasks employees already do - drafting internal documents, summarizing information, preparing spreadsheets, or answering questions about approved company materials. Make the tasks specific enough that different tools can be tested against the same expected outcomes.
Then cost the work needed to deliver that outcome, not just the monthly plan price. For each tool, include the relevant license or usage charges, the number of employees who need access, and the time required to review and correct outputs. If implementation and administration are material in your environment, include those too. A model that answers quickly but needs extensive checking can shift the work rather than reduce it.
A practical pilot should run for several weeks with a representative mix of people. Include occasional users and heavier users, and record which tasks they actually complete. Ask employees whether the tool saved time, improved the result, or increased review work. Don't rely only on usage counts: frequent use can mean value - or repeated retries on weak output.
For each task, track a few consistent measures:
- Cost: license or usage expense attributable to the task or user group.
- Time: time to produce an acceptable result, including review and correction.
- Quality: whether the output meets the team's practical standard.
- Adoption: whether the tool fits existing work without workarounds.
- Risk and control: whether access, data handling, and review requirements are acceptable.
A simple calculation usually reveals the trade-off: compare the pilot's total cost to the value of time saved on completed work. Use conservative assumptions. If someone "saves" ten minutes but spends that time later fixing errors elsewhere, the net saving is close to zero. If freed-up time goes to work that has real business value, capture that separately instead of treating every saved minute as cash.
Build a cost model that reflects real usage
A monthly model can look like:
Total cost = licenses + usage charges + implementation and administration + review and rework
Track the value side separately:
Net operational value = time saved on successful work + quality improvements − additional review and correction time
Avoid rolling everything into a vague "ROI" number unless you can explain how you measured each input. In practice, the key distinction is often between a tool that's cheap to provision and a tool that actually reduces the cost of the workflow.
Also test more than one usage scenario. Model light, expected, and heavy usage. Decide whether your pricing keeps costs predictable as adoption grows, and who approves expanding usage. This matters when you want broader access but need guardrails against unexpected consumption.
Finally, compare products under the same conditions. Use the same task descriptions, comparable user groups, and clear quality criteria. If one tool integrates directly into an existing workflow and another forces people to move data between applications, document that difference. Integration isn't automatically a win - but the time and friction it removes (or adds) are part of the real economics.
When does Copilot make financial sense?
Copilot is more likely to make financial sense when it fits the work employees already do inside Microsoft's productivity environment - and when predictable costs matter for routine tasks. A fixed monthly fee can simplify budgeting and reduce the need to micromanage small usage charges per interaction. For many organizations, that predictability is worth paying for, especially when broad access is more useful than tightly metered access.
But software adoption shouldn't be mistaken for proven value. Microsoft's large base of business software users creates a potential distribution advantage, not proof that every employee will use Copilot effectively. The reported 30 million paid users are also only a slice of the company's 450 million business users of its office software. That gap makes rollout design and sustained adoption part of the business case: paying for access isn't the same thing as getting useful work out of it.
The competitive landscape is also moving beyond chat assistants. Anthropic has introduced Claude Sheets and other office-focused tools aimed at Excel users. At Ignite, Microsoft executive Jacob Andreou criticized rivals' "simplified viewers and editors," arguing they compromise on power and collaboration. That's a positioning argument, not a cost comparison - but it raises a real procurement point: the economics depend on whether a tool supports the full workflow (including collaboration and editing) or only a single step.
There may also be a case for using more than one tool. AT&T reportedly pays for tens of thousands of employees to use Copilot while also using OpenAI's Codex. That example doesn't prove multi-vendor setups are right for everyone. It does show that companies can choose tools by job rather than forcing one assistant to handle every use case.
A decision framework for buyers
If you're considering a broad commitment, use this sequence:
- Define the work. Name the tasks and teams that should benefit. Don't buy licenses based only on general interest in AI.
- Check the workflow fit. Identify where the tool fits into employees' existing applications and what extra steps it adds.
- Pilot against alternatives. Test Copilot, Claude, or ChatGPT on comparable tasks where each is a realistic candidate.
- Measure outcomes, not enthusiasm. Track acceptable completions, review time, and adoption alongside spend.
- Set usage and governance rules. Decide what data can be used, who reviews consequential outputs, and how high usage is monitored.
- Expand only where the evidence holds. Keep the tool for the teams and tasks where it performs well. Don't treat a company-wide rollout as proof.
The trade-off isn't only fixed pricing versus usage pricing. It's predictability versus flexibility - balanced against how much control you need over adoption and how well each product matches your real workflows. A fixed fee can make budgeting easier, but it can also mean paying for inactive users. Usage-based billing can align cost with demand, but it only works if you can monitor consumption and prevent avoidable spikes.
Key takeaways
- Microsoft's cost claim is based on research it commissioned - review the assumptions before treating it as a benchmark.
- Anthropic's under-$75 average for non-coding Claude Enterprise users is a separate claim, not a direct rebuttal on identical terms.
- Compare cost per successful task and total operating cost, not just monthly prices.
- Pilot with real users, measure review effort, and set controls before scaling.
- Choose tools based on workflow fit and evidence. One platform doesn't have to serve every team.
Conclusion
Microsoft's case for Copilot is plausible, especially for companies that want predictable costs and a close fit with Microsoft's productivity environment. But a vendor-funded study and competing vendor averages can't tell you what your organization will actually spend - or save. Treat the numbers as starting points for a comparison, not as a procurement shortcut.
Measure the cost of completing real tasks, including employee review and operational overhead. Pilot with representative users, monitor usage, and expand only where the tool produces reliable value. If you do that, you'll end up with a better answer than picking the assistant with the most confident price claim.
