What actually happened at Microsoft
Microsoft spent years pushing employees to use AI aggressively, then started asking what all of it cost. The numbers surfaced through an internal spreadsheet where staff share pay and working-condition data, which added a column labeled "AI $ Usage Per Month" during 2026 (Ynetnews).
The eye-catching figure: one worker in Customer and Partner Solutions reported roughly $28,000 in AI usage over 28 days (Ynetnews). That is one person, one month.
Treat the spreadsheet as anecdote, not census. Only about 350 US-based employees filled in the field, out of more than 223,000 Microsoft workers worldwide, so it is a voluntary sample rather than a representative one (Ynetnews).
The per-division spread
Among those roughly 350 self-reporters, the medians varied a lot by team. The table below is drawn from Ynetnews's reporting on the internal spreadsheet:
| Group | Median AI spend (per 28 days) |
|---|---|
| All reporting employees | ~$300 |
| CoreAI | ~$975 |
| Security | ~$526 |
| Microsoft AI | ~$490 |
| Cloud AI | ~$325 |
Source for the figures above: Ynetnews. Several departments also had individual employees clearing $10,000 in a single 28-day window.
"Tokenmaxxing" and the response
The behavior driving the extremes had a nickname: "tokenmaxxing" — staff inflating token consumption with low-value prompts to climb a visible usage leaderboard. Jay Parikh, the executive VP running CoreAI, sent a memo saying "Tokenmaxxing is not what we are trying to maximize" and pushed the team toward "maximizing outcomes that create real change" (Ynetnews).
Microsoft also shifted some internal work to OpenAI's less token-hungry GPT-5.6 Sol for "higher value for the token investment" (Ynetnews). The pattern is not Microsoft-specific: Uber reportedly exhausted its entire 2026 AI coding budget within four months, and Microsoft itself canceled most of its Claude Code licenses after six months over unexpectedly high usage (Fortune).
Why this matters to an operating store
You are not Microsoft, but you face the same trap in miniature. The moment you adopt a usage-metered AI tool, your bill stops being a fixed line item and starts tracking behavior you can't fully see.
The uncomfortable part is that AI is not automatically cheaper than a person. As one Nvidia VP put it, "the cost of compute is far beyond the costs of the employees" for his team (Fortune). Per-token prices keep falling — Gartner expects inference costs on advanced models to drop nearly 90% by 2030 versus 2025 — yet total spend rises because agentic tools burn far more tokens per task (Fortune).
So "the tokens got cheaper" does not mean "your bill got smaller." For a store owner, the defensible question is not how much AI am I using but how much profit did that AI usage protect or create. That is the profit angle every news piece on this story skips.
A worked example on your own P&L
Say you run 340 orders a month at a $31 average order value, with about $2,800/month in Meta spend. That is roughly $10,540 in monthly revenue before you subtract product cost, fees, and ads.
Now imagine two AI setups for the same job — watching margin, drafting support replies, checking ad delivery:
- Metered "chat with everything" tool: you pay per token. If a curious month of heavy prompting doubles your usage, a $120 plan quietly becomes $240+, and you can't point to a single order it saved.
- Outcome-anchored setup: you pay a flat subscription, and the AI's consequential actions wait for your approval. Spend is bounded by the plan, and every action ties to a task you signed off on.
At your scale, the metered tool's worst month can eat a full point of margin without producing anything you can name. The second setup can't surprise you the way one Microsoft employee surprised finance with a $28,000 month (Ynetnews).
If you want the deeper framing on how software-as-employee compares to a metered chatbot or a human contractor, our guide to AI employees for ecommerce walks the full landscape.
Usage-priced vs outcome-priced AI
The Microsoft story is really a pricing-model story. There are three shapes you'll meet as an operator.
Metered by token or seat. You pay for consumption or for a login, whether or not the work lands. This is where tokenmaxxing lives — cost decouples from value, exactly the failure Parikh's memo names (Ynetnews).
Priced per outcome. The support-AI category already works this way. Gorgias, for example, charges roughly $0.90 per fully resolved conversation on most plans and bills nothing when a human takes over (Gorgias). You pay for results, not activity.
Flat subscription with approval gates. You pay a predictable fee, and the AI stages consequential actions for your sign-off before anything executes. Your exposure is capped by the plan, and your review is the throttle.
The last two are what protect an operating store from a runaway bill. If you're weighing whether to build metered tooling yourself, our note on when to hire AI developers versus buy an AI employee is worth reading before you commit to a token meter.
How PodVector AI's Victor is structured to avoid this
Victor is an AI employee for ecommerce and print-on-demand stores, and its design directly answers the Microsoft failure mode. Victor is not a dashboard and not a metered chat window — it is a worker whose actions you approve.
Every write action Victor takes is approval-gated: it drafts or stages the change, and you approve before anything runs. That single design choice makes a $28,000-surprise-month structurally hard, because nothing consequential executes without you in the loop.
Victor integrates across the tools that actually move your margin — Shopify store operations, Meta Ads, Google Ads, Printify, Printful, Gelato, and Klaviyo — computes your true per-order profit, and saves reports and CSVs to a PodVector AI folder in your own Google Drive. It also drafts customer-support email for your approval before sending.
The point is coordination you can audit. When Victor looks at why last week's margin dipped, the work is tied to a decision you made, not tokens spent chasing a leaderboard. If support volume is your biggest AI question, our breakdown of the AI chatbot options for ecommerce pairs well with this cost lens.
What to actually do about your AI spend
- Meter your own usage first. Add a line to your monthly P&L for every AI tool, the way Microsoft belatedly did. You can't scrutinize what you don't track.
- Prefer outcome or flat pricing for anything customer-facing. Metered chat is fine for exploration; it is risky for production work where volume spikes.
- Insist on approval gates for money-moving actions. Refunds, budget changes, and sends should wait for a human. Every serious vendor — Google, Shopify, Gorgias — builds this in.
- Keep the work product in your own accounts. With more than 40% of agentic AI projects predicted to be canceled by end of 2027, per Gartner, prefer tools whose reports and changes live in your Shopify, Klaviyo, and Drive so the artifacts survive the vendor (Gartner).
Ready to put a per-order-profit view and approval-gated actions in front of your store? Start with PodVector AI.
FAQs
Why is Microsoft scrutinizing employee AI costs now?
After years of encouraging heavy internal AI use, Microsoft started measuring what it cost and found wide, sometimes extreme spending — one worker near $28,000 in 28 days and a company-wide median around $300 per 28-day period among reporters (Ynetnews). The scrutiny is about tying spend to real outcomes rather than raw usage.
What is "tokenmaxxing"?
It's employees inflating their token consumption with low-value prompts to look busy on a usage leaderboard. Microsoft's CoreAI leader publicly told staff "Tokenmaxxing is not what we are trying to maximize," steering them toward outcomes instead (Ynetnews).
Is AI actually cheaper than hiring a person?
Not automatically. An Nvidia VP noted that for his team "the cost of compute is far beyond the costs of the employees," and enterprise AI spend keeps rising even as per-token prices fall, because agentic tools consume far more tokens per task (Fortune). The honest comparison is cost per outcome, not cost per token.
How do I stop my own store's AI bill from ballooning?
Track every AI tool as a P&L line, favor outcome-based or flat pricing over metered chat for production work, and require approval gates on any action that moves money. Metered tools reward volume; a flat plan with sign-off caps your exposure the way Microsoft's runaway month was not capped.
How does an approval-gated AI employee change the cost math?
Because consequential actions wait for your approval, spend maps to work you authorized instead of open-ended token use. Victor by PodVector AI works this way across Shopify, Meta Ads, Google Ads, Printify, Printful, Gelato, and Klaviyo, computing true per-order profit and delivering reports to your own Google Drive — so you can see what each action was for, not just what it consumed.