Every per-client app fee makes you run the same quiet calculation. Will this tool save enough time on this client to cover its fee? For your biggest clients, easy yes. For the small write-up client, no, so you skip the tool and eat the hour of manual work instead.
In part 1 we unbundled an app subscription into five layers. This is the strangest one: the pricing wrapper, where your vendors quietly make decisions about your margin.
The decision tax
Per-client pricing forces a value judgment the vendor’s economics impose on you. Not all clients are created equal, but the fee pretends they are, so you adjudicate every small client against it, one by one, forever. That’s a standing tax on your attention, and it has nothing to do with what you actually consume.
You’ve run this movie before. QuickBooks Desktop’s perpetual licence gave you unlimited ledgers for one annual fee. Subscriptions ended that (suddenly every client had to justify a monthly charge individually) and the vendors patched the damage with tiers, Xero’s Ledger edition first, cut-down functionality at a write-up price. The tier system is the vendor admitting the unit is wrong. Tax software’s pay-per-return versus unlimited licence is the same adjudication in annual form.
Per-user pricing is the same defect, rotated. It punishes the part-timer, and it forces the admin-seat question: full freight for a user whose access is administrative, not productive, because most platforms won’t give you a free admin seat. Wesley launched in July at $96 per staff member per month with unlimited clients. A different unit, the same forced call: is this seat worth it?
The wrapper is a risk contract
Now add AI to the app, and the wrapper gets more interesting. AI is metered in tokens, the units consumption is measured in; every page the model reads and every email it drafts burns some, the way a meter turns while the lights are on. Your clients don’t consume tokens evenly, and token prices move.
So a fixed per-client fee on an AI-powered tool is a risk transfer. The vendor absorbs the volatility of what your work actually costs, sells you price certainty, and keeps a repricing option it can exercise whenever the math stops working. The option is already live: Intuit raised QuickBooks prices to reflect AI value, and Dext priced its AI model at £5 per client. And Claude’s newest model, Fable 5, runs only on metered, pay-per-use credits. Its included trial window has closed, and the model was never part of the standard subscription. Whoever controls the model choice controls your margin, and inside a wrapper, that’s never you.
Watch the meter
Here’s what I think happens next, offered as a watch-for rather than a promise: the unit of pricing drifts toward the unit of cost. The model market itself just showed the direction. In one July week, Meta started charging for its models for the first time, OpenAI priced GPT-5.6 in three consumption tiers, and Fable moved to pure usage credits. Usage-based pricing puts the write-up client whose work burns $5 of tokens a month and the group client that burns $500 inside the same contract with no adjudication at all. The decision tax simply disappears.
But be honest about the trade, because it cuts. Usage pricing is the vendor unbundling its own wrapper: keeping the margin on the harness and janitor layers from part 1, and handing the token volatility back to you. The first usage-based GitHub Copilot bills landed with June’s billing cycle at 10 to 50 times the old flat subscriptions. You can have perfect allocation or price certainty, not both. Most days I’d still take the floating price over the standing decision tax. A meter can be managed; the adjudication never ends.
Building inverts the contract entirely: floating costs, but you choose the model per task: a frontier model for the complex advisory analysis, a near-frontier value model like Grok 4.5 at $2 per million input tokens for production work, a cheap open-weight model (one you can run yourself) for internal tooling. Routing, not renting, becomes the skill.
One mirror to glance at before you resent any of this. Your own fixed monthly fee does to your clients exactly what the vendor’s wrapper does to you: you absorb the AI cost volatility, they get certainty, and you hold the repricing option. As the close reprices (the squeeze we’ve been writing about), it’s worth asking what your wrapper is worth, and when a client might first ask to see the meter.
Read the renewal like an accountant
A pricing page is a risk instrument, not a feature list. So at each renewal, ask the questions you’d ask of any contract that allocates risk: who carries the volatility, who holds the repricing option, and what does the pricing unit force you to decide: per client, per seat, forever?
None of that answers whether you should build instead. That depends on something no vendor can sell you, and it’s where part 3 finishes the series next Wednesday, with the build-readiness test and one question to ask for each of the five layers. Until then: in each of your subscriptions, who’s carrying the token risk, and what are you paying them for the privilege?
Reading your stack this way is the easy half. Deciding what to do about it, workflow by workflow, is the AI Practice Transformation program: four consecutive weeks, one day a week, live, working on your own firm’s operations rather than a case study. theaiaccountant.ai/transformation