When AI tools start eating the margin
7 min read
Microsoft has removed the Claude Code licences for its Experiences and Devices division. The cutoff date is 30 June 2026. Officially, the move is justified as “toolchain unification.” In reality, it’s about the token bill-Uber burned through its entire 2026 AI budget in just four months. For mid-market companies in the DACH region, this isn’t a Big-Tech drama; it’s an early warning for their own 2026 and 2027 AI budgets.
Key takeaways
- Big Tech is consolidating because tokens are expensive. Microsoft, Meta and Amazon are merging internal AI licences because agentic workflows can consume up to 1,000 times more tokens than classic chat calls.
- Tool sprawl is the costly half of the problem. Three to five parallel AI coding and chat tools inside the same company create licence duplication, onboarding overhead and no clear ROI milestone.
- Mid-market still holds the lever that corporates have lost. If you’re not running 20 AI tools in 2026, you can audit now, calculate cost-per-task and lock in a licence strategy before the next annual budget lands.
Related:Fujitsu AI Platform July 2026 / Process optimisation fails at the handover
What happened-without the PR froth
What does tool consolidation mean in an AI context? Tool consolidation is the deliberate merging of multiple parallel AI tools into a smaller, curated stack to cut licence costs, better control token burn and reduce the organisational complexity around onboarding, data flows and compliance. In the Big-Tech sphere, the current push is away from third-party tools like Claude Code and back toward in-house platforms such as GitHub Copilot CLI.
Windows Central reported in mid-May that Microsoft is shutting off internal Claude Code licences for the Experiences and Devices group-the division behind Windows, Microsoft 365, Outlook, Teams and Surface. The official line is “toolchain unification.” Yet the fiscal year-end for Microsoft falls exactly in this window, making the reading “consolidation as cost-cutting” entirely plausible.
Important: Anthropic models remain available via Copilot CLI. This isn’t a vendor boycott; it’s about licence sprawl and who ultimately foots the bill when employees access the same models through three different front-ends.
The number that should keep every board awake in 2026
This figure sounds impressive, but it isn’t a game. It’s the operational reality the moment a tool starts researching on its own, writing code, running tests, and compiling reports. Each step generates tokens. No one pays for a single question. But when a workflow has 40 intermediate steps and a team performs it daily, the monthly bill tells a different story.
Tom’s Hardware cites OpenClaw initiator Peter Steinberger with a monthly token bill of $1.3 million for a single team. Uber reportedly burned through its entire 2026 AI budget within four months, according to the same reports. Meta recorded more than 60 trillion tokens in just 30 days during an internal competition called “Claudeonomics.” These aren’t slide-deck numbers. They belong in a quarterly P&L.
What’s quietly happening right now in DACH mid-market companies
Mid-market firms see things differently, yet the mechanics are identical. Three tools running in parallel is commonplace: GitHub Copilot in development, ChatGPT Enterprise in marketing, a separate AI platform for customer service. Add the personal accounts nobody officially approved. Add the demo license that kicked off in an innovation workshop and has been running in the background ever since.
What explodes at scale in large enterprises often costs mid-market companies proportionally more because no central procurement bundles licenses. Senior management rarely knows which AI tools are active across the business-and which of them redundantly burn tokens on the same tasks.
What breaks
- Licenses for three AI-coding tools in the same team
- Token consumption without cost-per-task visibility
- Shadow accounts in departments without IT approval
- Pilot projects without end dates or success criteria
- ROI measured in licenses per employee instead of business outcome
What works
- A single, centrally managed tool inventory with a named owner
- Monthly budget caps per team, not per person
- Cost-per-completed-task as the KPI, not raw token count
- Sunset clauses baked into every pilot from day one
- Centralized model-gateway strategy instead of “everyone buys their own”
The right-hand column is what a leadership team can green-light in half a morning without rolling out a new framework. It does, however, demand someone take ownership of the tool inventory. Mid-market companies rarely have that role explicitly defined today. By 2027 at the latest, they’ll need it-or they’ll learn Microsoft’s lesson two years too late.
Four questions every executive team should ask in 2026
Concrete, no framework bingo. These four questions drive more impact than any new AI strategy slide deck. If you can’t answer them, you’ve lost track of your own AI spending.
First: Which AI tools are currently active in your organization, under what license, and what’s the monthly bill? If the answer takes longer than two days to compile, the problem is already staring you in the face.
Second: What task does each tool actually perform, measured by real output? Not “adoption rate” or “daily active users.” Instead: Which task that used to be handled differently now runs through the tool, and how much measurable time or money does it save?
Third: Where do functions overlap? Two tools generating marketing copy are consolidation candidates. Three tools drafting service-center responses are a license penalty waiting to happen.
Fourth: What happens if token prices triple? The question sounds academic, but it isn’t. It forces an answer: Which AI use would immediately stop paying for itself? That use should already be on the chopping block today.
A 60-day plan without slide theater
Nine weeks sounds like a lot. It’s less time than most vendors reserve for a single sales cycle. Completing the sequence delivers an honest cost picture by Q4 and lets you budget for 2027 with facts instead of hoping token prices will drop.
What’s still left open
What Microsoft, Meta and Amazon are learning right now, mid-market companies will learn one to two years later. The upside: the numbers are already on the table. A consolidation decision in May 2026 no longer requires the courage of a pioneer-only the discipline to face your own tool stack and token bill.
Act this summer and you walk into year-end with a tidy AI landscape. Wait until February 2027 and you’ll have to explain why the line item suddenly ballooned beyond plan.
Frequently Asked Questions
Does Microsoft’s decision mean that Claude Code is no longer recommended for enterprise use?
No. Microsoft’s move is an internal consolidation decision timed to coincide with fiscal-year-end. Anthropic models remain available via Copilot CLI. For a mid-sized company, the question isn’t “Microsoft-compliant or not,” but rather: which tool actually performs each task better-and what does that cost per month in the real workflow?
What exactly does “cost-per-task” mean?
Cost-per-task is the metric that measures token and license spend for each completed business task. Example: if the marketing team generates 40 product texts per month via an AI tool and the bill is €1,200, that’s €30 per text. This figure is the foundation for every consolidation discussion and replaces fuzzy ROI statements.
How can senior management spot shadow tools in the company?
Through departmental credit-card statements, browser auto-login suggestions from staff, and a short, openly worded team survey. Crucially, the survey must be communicated without any punitive tone-users usually had a solid operational reason for adopting a shadow tool, and that rationale is valuable input for consolidation decisions.
Is a single AI tool enough for an SME?
Rarely. In most mid-market cases, two to three clearly scoped tools with distinct use-cases is the honest answer: a generalist chat assistant for office workflows, a coding assistant for IT, and possibly a specialized solution for service centers or sales. More is seldom needed, and more often than not, it leads to tool drift.
What does a realistic tool audit cost?
For a company with 100–500 employees, expect two to four person-days of internal work plus an external spot-check if internal data is unclear. That investment guards against duplicate licenses that, without an audit, often quietly run into five-figure annual totals.
Editor’s Reading List
More from the MBF Media Network
Source of header image: Pexels / Sergei Starostin (px:6466141)
Source of header image: Pexels / Sergei Starostin
