Start with the contradiction that should keep you up at night.
Microsoft put Copilot everywhere. Windows 11, Word, Excel, Outlook, Teams, the taskbar — the company describes it as almost impossible to avoid. And fewer than 4.5% of Microsoft 365 customers pay for it. Of those who do, only 20 to 30% use it weekly. Do the arithmetic: roughly 1% of the base is a weekly active premium user.
Now hold that against a smaller, quieter number. Somewhere, a PM opens a finance email on a Tuesday morning. Their AI feature — shipped, loved, "included in Pro" — is costing about $22,000 a month. That's roughly 14% of the revenue from the customers using it. A few power users are eating 40% of the bill.
One company can't get anyone to pay. The other is getting paid and losing money anyway. Most readers assume these are opposite problems. They're the same problem, seen from two ends: pricing was never treated as a product decision, so ubiquity and adoption both arrived without margin attached.
Last week I covered the vendor side of the SaaS repricing scramble — SAP and Microsoft rewriting list prices while the ground shifted under them. This week the story moves to the buyer's desk and the builder's desk, where monetization actually fails: the moment a feature ships on the wrong model, and the moment someone finally reads the invoice.
Here's the statistic everyone quotes and almost nobody metabolizes. Three of four SaaS companies have shipped generative AI features. Only 15% are successfully charging for them.
That 85% gap is not a to-do item waiting for finance. It's evidence of a decision — or the absence of one. PMs shipped AI features on pricing models that were structurally guaranteed to lose money, then handed the bill downstream.
The mechanism is boring and brutal. Traditional SaaS ran on near-zero marginal cost: the 10,000th action cost the same as the first, which was nothing. Per-seat pricing worked for two decades on exactly that assumption.
Every metered model call breaks it. The 10,000th call to a frontier model costs roughly what the first one did — a real cost per invocation, not a rounding error. AI features are compressing gross margins from the classic 80–90% down to 50–60%, and lower if the unit economics aren't nailed.
So when you price a feature flat and "include it in Pro," you've made a specific bet: that usage stays predictable and cheap. The Tuesday email is what it looks like when that bet loses.
James Colgan, an operating partner at Cathay Innovation who spent time at Microsoft and Slack, put the inversion plainly: "In the age of AI, where your cost structure is different, your most engaged user could end up being your most expensive one."
In the age of AI, where your cost structure is different, your most engaged user could end up being your most expensive one.
Sit with how much that breaks. For twenty years, heavy engagement was unambiguously good — stickier accounts, better retention, better net dollar retention. Now the customer sending 500 queries a day is actively eroding your margin, and a flat seat fee collects the same money from them as from someone who logs in twice a month.
This isn't a fringe case. Usage variance in AI features runs far wider than in classic SaaS — a heavy user can burn 100 times the compute of a light one. Your best, most engaged account is often the one quietly bleeding you.
That's the finance email decoded. The feature works. Users love it. And the person who chose "flat, included in Pro" chose a model with a built-in failure mode nobody instrumented for.
Now return to Microsoft, because it kills the obvious escape hatch. The tempting reading of the 15% problem is that it's a distribution problem — ship AI into everything and payment follows.
Microsoft has more distribution than anyone alive. Half a billion commercial seats. Copilot bolted into every app people already open. And the paid weekly-active number is around 1%.
Ubiquity is not demand. "Ship it into everything" is not a pricing strategy. Businesses are asking one question — does this justify another $30 per user per month on top of what we already pay? — and mostly answering "not yet." Microsoft's own Copilot leadership reportedly believes the product still needs to "earn the right to exist."
So distribution doesn't save you and adoption doesn't save you. What's left is the actual work: deciding what a customer pays for, tied to what the action actually costs — and revisiting that decision as fast as the product changes.
Which is exactly why this becomes the PM's job and no one else's.
Prittam Bagani, VP of Product Management at Chargebee, makes the timing argument: because AI capabilities evolve so fast, pricing has to evolve at the same pace. He points at OpenAI's pricing page as the live example of how often these decisions now need revisiting.
Finance can't move at that cadence. Finance runs quarterly change control. If your product ships new AI capability every few weeks, whoever owns pricing has to understand what the product can do at any given moment — and that's product's knowledge, not commercial's.
There's a second reason it lands on product, and it's the more fundamental one. Only product knows the true cost-per-action. Every AI query is compute, retrieval, infrastructure, and third-party calls hidden inside each request. The person who scoped the feature is the only one who can name that number before launch.
Notice this is the same blind spot showing up on two sides of the house. On the engineering side, multi-model routing carries a re-evaluation tax with no budget line; on the revenue side, per-user AI cost carries no pricing line. Same unowned cost, two symptoms.
Look at what disciplined cost ownership actually does. One practitioner consolidating an AI stack across a 120-vendor enterprise pulled 90 days of invoices, went through them one by one, and cut the annual run-rate from £50k to £20k — about 60% of spend was redundant or overlapping, much of it shadow tools used by fewer than five people. He was auditing cost from the buyer's side.
The PM staring at the Tuesday email is doing the identical job from the seller's side: knowing, per action, what you're paying to deliver — and pricing so the meter and the cost move together.
The good news is the levers exist. Usage-based, credits, per-query, hybrid, outcome — Bessemer's line is that your charge metric is a strategic statement, not just a billing decision. Replit went from roughly $2M to $144M ARR while moving off flat plans to usage-based, dragging gross margin out of negative territory into the 20–30% range because the charge metric finally reflected what heavy users cost.
But you only get to pull those levers if you built the instrument first.
Here's the discipline, and it costs a day. Even if you price flat, instrument per-user AI usage before launch. Know who your power users are. Set a norm so you can have the "your usage is well above average" conversation instead of discovering the outliers in an invoice.
Teams that ship flat include without usage analytics don't just get a surprise bill. They get a surprise bill with no lever to pull — they can't see who's driving cost, can't throttle the outliers, can't reprice what they never measured.
That's the whole argument. The 15% statistic isn't finance's failure to monetize later. It's product's failure to decide earlier.
Every "successful" AI launch that shipped on flat pricing with no per-action cost instrumentation is a margin liability wearing a growth win's clothes. The finance email is just the day the disguise comes off.
Instrument the cost before you ship, or read about it on a Tuesday.