The Subscription Squeeze: Your Analytics Stack Is Getting More Expensive, and the Work Isn't Getting More Durable

I work from a shared office. Lately the background noise has changed. It used to be phones and coffee orders. Now it’s people ringing IT to ask for a bump in their AI tokens. “I’ve run out again.” “Can you raise my limit?” Same person, twice a week.

And here’s the thing about what those tokens are buying: I’ve watched what they build with them. Beautiful Claude artifacts, interactive charts and dashboards, generated in minutes. And then what? They’re a file. A one-off snapshot of a question asked once, with the data hard-coded at the moment it was generated. No refresh, no connection to the source, no governance. Next month the numbers move and that artifact is already wrong. The tokens produced something that looks like a dashboard and rots like a screenshot.

So the office is filling up with two kinds of stale: the Power BI dashboards nobody opened since March, and the Claude artifacts that were wrong by Tuesday. Different tools, same failure mode, and neither of them is what the person ringing IT actually needs.

Sitting there, I keep thinking of a particular business model. The first one’s cheap. The usage creeps up. The dependence becomes structural. And then the price goes up, because where exactly are you going to go?

I’m talking, of course, about enterprise software subscriptions. What did you think I meant?

The difference between that arrangement and a healthy purchase is simple: one of you owns something at the end. Right now, most organisations are renting their entire analytics capability (the licences, the compute, the tokens), and the only things they own outright are the artifacts nobody wanted anyway: the stale dashboards.

The price rises are not hypothetical

This is not a vibe. The invoices moved.

Power BI. In November 2024, Microsoft announced its first Power BI price increase in nearly a decade: Pro went from $10 to $14 per user per month, a 40% jump, effective April 2025. Forty percent is not a gentle nudge; it’s a repricing of what a seat is worth to you now that you’re locked in.

Microsoft 365. Announced December 2025 and effective 1 July 2026 per Microsoft’s own licensing notice: Office 365 E3 up 13%, Business Basic up 16%, Microsoft 365 Apps up 17%, and the frontline F1 SKU up 33% (43% for the no-Teams variant). When your per-seat line items rise mid-teens across the board, the stack you weren’t thinking about gets more expensive by default.

AI subscriptions. This one is the purest demonstration of the model. At DevDay on 29 September 2026, OpenAI launched a new $500/month ChatGPT Pro tier with the highest usage allowances and the fastest model access. And for the $200 plan? The price didn’t move. The usage allowance was cut in half, with a one-month grace period for existing subscribers, then the same $200 buys half as much. Anthropic and Google restructured their tiers in the same window. Note the mechanics: they didn’t even need to raise the sticker price. They just cut the dose. Same money, half the product, and nowhere to go, because by then your workflows live inside it.

And the usage-based side is worse, because it’s unbounded. I wrote recently about Uber’s CTO describing how coding-agent adoption burned the company’s entire 2026 AI budget in four months, at $500 to $2,000 per engineer per month. Agentic workflows consume multiples of the tokens anyone budgeted for. Total bills climb even where per-token prices fall. When the per-token price does rise, as the major providers eventually will once the price war settles, the effect is multiplicative, not additive.

This is the part of the market that behaves exactly like the business model I mentioned. Low-cost entry, rising dependence, structural lock-in, repricing at will. Microsoft didn’t raise Power BI 40% because costs forced it; it raised it because ten years of ecosystem investment made your reporting stack surgically attached to it.

Meanwhile, the thing you’re buying mostly rots

Here’s the uncomfortable symmetry. The spend is rising, and the outputs have a half-life.

Gartner has been measuring this gap for years. Their 2018 analysis found 87% of organisations have low BI and analytics maturity, and adoption has stayed stubbornly low ever since. Industry analyses routinely put unused dashboards somewhere between 60% and 80%. A dashboard is a snapshot of a business question, frozen on the day someone built it. The business moves; the dashboard doesn’t. Six months later it’s a museum exhibit that still costs a seat.

The broader SaaS numbers are just as damning. Zylo’s 2025 SaaS Management Index (40 million licences, $40 billion of managed spend) found average SaaS spend of $4,830 per employee (up 21.9% year over year), with the resulting waste at roughly $21 million per large organisation annually, most of it licences that go unused.

So the full picture: per-seat prices up sharply, consumption-based AI costs compounding, and more than half of what was already bought sitting idle. People in my shared office are ringing IT for token bumps on tools whose dashboards nobody opened last month.

I’m not against subscriptions. I’m against subscriptions for things that don’t accumulate. Rent is fine for the tool. It’s a terrible model for your business logic.

The stale dashboard problem is a definitional problem

Why do dashboards go stale? Not because the charts age. Because the definitions inside them were never owned.

Every dashboard embeds a pile of business logic: what counts as “active”, which region rolls into which, how revenue is recognised, which test records to exclude. When that logic lives inside one report file, it dies with the report. The next person rebuilds it, slightly differently, in a new dashboard, and now you have three versions of “revenue” and a meeting about which one the board should trust.

This is also exactly why dashboards are hard to “productionise”. There’s nothing to productionise; the logic isn’t anywhere durable. It’s trapped in the artifact.

The fix has a name, and it’s the least glamorous word in analytics: the semantic layer.

A semantic layer is where every business metric is defined exactly once, with its filters, its grain and its exclusions, in a model that tools consume rather than contain. In the Microsoft world, that’s a Fabric Data Warehouse or a Lakehouse (through its SQL endpoint) sitting on governed data: your tables, your views, your business rules defined once in T-SQL where every tool can reach them. Done properly:

• Excel connects straight to the Fabric Data Warehouse or Lakehouse’s SQL endpoint, the same way it connects to any database. The finance team’s pivot tables and the board report now agree, because they’re reading the same views.
• Power BI reports become thin, replaceable views over the warehouse. A stale report is now a fifteen-minute rebuild, not a two-week re-derivation.
• Fabric apps and custom apps query the same tables and views.
• AI agents get the benefit too: route an agent’s metric questions through the warehouse and it inherits your business rules for free. Anthropic’s own analytics write-up made this point: their agent’s accuracy jumped from 21% to over 95% not because of the model, but because of curated definitions and rules the model was made to consult. The semantic layer is the same discipline, whether the consumer is a person with a pivot table or an LLM.

The strategic property is this: the per-seat tools keep repricing, but a semantic layer in Fabric is an asset. It’s the one layer that survives tool churn. Swap the front end (Excel to Power BI, Power BI report to Fabric app, dashboard to agent) and the definitions stay put. You stop paying rent on your business logic and start owning it, and the next price increase lands on a commodity layer you can migrate, not on the accumulated meaning of your business.

Cheap dashboards are disposable. A semantic layer compounds. That’s the whole trade.

If you’re staying on Power BI: spend less on the same thing

Owning your definitions doesn’t mean paying more for seats. Some of the biggest Power BI savings are structural, and most organisations I visit haven’t done them:

• Audit before optimising. Zylo’s data says over half of licences go unused. Pull the Microsoft 365 admin centre’s Power BI activity metrics and find who hasn’t opened the service in 90 days. Reclaiming dormant Pro licences is free money.
• Check your E5 overlap. Microsoft 365 E5 and Office 365 E5 (annual billing) include Power BI Pro. Buying standalone Pro for E5 users is the most common double-purchase I see.
• Do the F64 breakeven maths. The per-seat trap is that viewers vastly outnumber creators. At Fabric F64 and above, report viewers need no paid licence at all; only creators do. With a one-year reserved instance, F64 runs roughly $5,000/month, which breaks even against $14 Pro seats at around 350 viewers. If you have hundreds of consumers, per-seat licensing is the expensive option; run the numbers for your region.
• But F64 isn’t the only exit from per-seat pricing. The F64 threshold only matters for the publish to the service model. The Power BI Embedded “app owns data” route has no such floor: viewers never need licences at any capacity level. A number of European vendors now offer branded, secure report portals with row-level security on Power BI Embedded, running on an F2 capacity (from a couple of hundred euros a month) instead of F64’s ~$8,400. If your viewers are external (clients, partners, franchises) or you mainly need a clean portal rather than org-wide Entra-governed distribution, an F2-plus-embedded setup can undercut the F64 breakeven by an order of magnitude. The semantic layer stays the same either way: the warehouse holds the definitions, the front end is disposable.
• Buy reserved capacity, pause the dev boxes. Reserved Fabric instances run about 40% below pay-as-you-go. Dev and test capacities can be paused outside working hours so you’re not paying 24/7 for an environment nobody uses at 2am.
• Consolidate workloads onto one capacity. If you’re using Fabric for pipelines or warehousing as well as Power BI, share one capacity across workloads instead of buying per-purpose.

None of that requires a re-platform. It requires someone to actually look at the licence report.

What I’d actually do

The price trajectory is now visible: per-seat up 40% where it was flat for a decade, suite prices up mid-teens, and AI plans where the price stayed put but the allowance was halved: an effective doubling. Meanwhile token consumption keeps compounding. And the artifacts those subscriptions produce are the least durable things in the building.

So the play is to invert where the money goes:

1. Define every important metric once, in a Fabric Data Warehouse or Lakehouse on governed data, not in a dashboard.
2. Make the tools thin. Reports, workbooks, and apps become views over the warehouse, cheap to rebuild and impossible to make disagree.
3. Cut the seat spend you don’t need (the list above) and stop buying licences for people who only ever consume definitions.
4. Budget AI consumption explicitly. Token burn is a real line item now, not a rounding error, and it’s the part of the bill growing fastest.
5. Use deterministic AI where the job is maths. Once the warehouse is up and the definitions are settled, don’t reach for an LLM when the question is a calculation. Forecasting revenue, scoring risk, predicting churn: that’s classic machine learning territory, and it’s dramatically cheaper than tokens. A gradient-boosted model retrained on a schedule costs fractions of a cent per prediction and gives the same answer every time. Save the expensive, non-deterministic language models for the work that actually needs language and judgement: the “why did this number move” questions, the ad-hoc investigation, the explanations. Paying language-model prices for arithmetic is how the token bill gets out of control in the first place.

The subscriptions will keep going up. That’s not a reason to panic; it’s a reason to make sure that the thing you’re paying for is the tools, not the intelligence. The intelligence should already be yours.

---

If your dashboards are multiplying faster than your trust in them, that’s the conversation I like having. Book a free consultation and we’ll map where your business logic actually lives.