If Your AI Bill Doubles, Which Workflows Do You Keep?

Spend 30 minutes on your AI spend this week and a price hike costs you a line item instead of a workflow you now depend on.

If you've quietly built AI into daily operations over the past year — call summaries, quote follow-ups, inbox triage, product descriptions — the bill is probably small enough that nobody looks at it. That's fine while it's small. It's a problem if the price per token changes and you have no idea which of those workflows would still be worth running.

This post is a budgeting exercise, not a prediction. We're not claiming prices will rise, and we have no source that says they will. We're saying you should know what happens if they do.

Why anyone is even asking this question

Two dated news items, plus one opinion of ours. None of the three is a forecast; all three are reasons to run the exercise.

First, power. TechCrunch reported on 14 August 2026 that hyperscalers may come to regret leaning on natural gas for data-center power if a new forecast proves correct. Inference runs on electricity, so power sits somewhere underneath what you pay per token. How much of your price it actually drives is not something we can source, and that article doesn't put a number on it for you.

Second, vendors are treating token cost as a headline feature. TechCrunch reported on 13 August 2026 that Writer launched a new model and an upgraded harness explicitly aimed at containing token costs. Our read — opinion, not something that article says — is that "we'll help you spend less" becomes a selling point when buyers start feeling the meter.

Third, and this is our opinion, clearly labeled as such: the current price of frontier AI looks to us like land-grab pricing, with vendors competing for developers and workloads. We'd rather plan as if today's per-token price is the floor than assume it keeps falling. We can't source that, and nobody should budget off it. It's a reason to hedge, not a number.

The practical response isn't panic. It's what you'd do with any supplier who might raise rates: know your exposure, know your alternatives, and don't build anything that only works at today's price.

The one question that ranks every AI workflow

For each place you use AI, answer this: what is one run worth, and what does one run cost? Both halves come from your own numbers — your average job value and your own provider bill — not from us.

Take a missed-call text-back. If it recovers one booked job a month, compare the margin on that job to the entire month of tokens that workflow burned. Our expectation is that the division comes out badly lopsided in the workflow's favor for most service businesses, but check it against your own average ticket instead of taking our word for it. Lopsided ratios are the ones that shrug off a price increase, and they're the ones you protect.

Now take a workflow that summarizes every email in a shared inbox — newsletters and vendor spam included. It's worth close to nothing per run and it costs the same per run as everything else. That's the kind that stops making sense first, and arguably shouldn't be running now.

If you have both kinds running and have never sorted them, sorting them is the whole exercise. High value per run: protect it. Low value per run and high volume: make it cheaper or turn it off.

A worked example (invented numbers — not a client, not a benchmark)

Every figure below is made up for illustration. It is not a case study, not a client, and not a benchmark of what anything costs. It's a template for the arithmetic. Fill in your own numbers from your actual bill.

Say a hypothetical nine-person home-services company runs three AI workflows in a month:

Workflow	Volume/mo	Share of AI bill	Value per run
Missed-call text-back + qualifying questions	600 calls	15%	Very high — recovers jobs
Quote follow-up emails drafted for review	250 quotes	15%	High
Tech notes + photos turned into a job summary and invoice lines	700 jobs	70%	Moderate, saves admin time

All figures in that table are hypothetical. Assume the total bill is $140/month and the admin time saved is roughly 18 hours at a fully loaded $28/hour — about $504 of labor in this made-up scenario. At $140, it's an easy yes. At $280, it's still an easy yes.

Here's the part owners miss. The third workflow is 70% of the spend and it scales with job volume. If this imaginary business grows 60% next year and prices double, that line goes from about $98/month to roughly $314/month. Invented inputs, real arithmetic: price risk and growth risk multiply. The workflow you're least attached to is the one carrying the exposure.

That's the output of the exercise. Not "AI is expensive," but "one specific job is 70% of my bill and it grows with the business, so that's the one I optimize."

Four ways to cap the bill this week

1. Put a hard cap on every key, per workflow. If your provider offers spend limits and budget alerts, turn them on today — check its current docs, because those features and their limits change. Then give each automation its own API key so the bill breaks down by job instead of arriving as one number. If your provider doesn't offer caps, put a counter in the workflow itself that stops after N runs a day. A cap that fires and pages you beats a surprise invoice.

2. Route boring work to the cheap model. Classification, tagging, extraction, routing, "is this a complaint or a quote request" — these don't need your most expensive model. Customer-facing prose and genuine judgment calls do. Split the two. Before you switch, take 50 real examples, run them through both models, and have a human mark each one right or wrong. If the cheap model matches on 50 out of 50 boring inputs, that's your answer for that job — and only for that job.

3. Send fewer tokens per run. You pay for what goes in as well as what comes out. Stop pasting whole PDFs when three fields will do. Cap output length. Batch ten items into one call where the task allows it. Strip boilerplate footers and signature blocks out of emails before they hit the model. None of it is glamorous, and in our experience it's usually the biggest single lever on a bill dominated by one high-volume job.

4. Keep the model swappable. The expensive mistake isn't paying more per token — it's building something you can't move. Keep the model name and prompt in one place, not hardcoded across fifteen steps, so switching provider is a config change plus an evaluation run rather than a rebuild. n8n makes the harder version of this argument: token prices won't increase if you host your own LLMs (undated post; read it as an argument, not a quote). Self-hosting is a real option for high-volume, low-judgment tasks, and it carries its own costs — hardware, ops, and someone who owns it. We haven't benchmarked the crossover point, and it depends on your volume and hardware, so for most nine-person businesses we'd treat self-hosting as the plan B you keep credible rather than the plan A. That's our view, not the article's.

What we'd actually build

For a business shaped like the example above, the build is a metering layer, not a platform: separate keys per workflow, a small usage log (date, workflow, tokens in, tokens out, estimated cost) written to a sheet or a table, a weekly summary that lands in the owner's inbox, and a hard stop on each key. Then one model-routing decision — cheap model for classification and extraction, better model for anything a customer reads.

We'd scope that as an afternoon of setup rather than a project. That's our scoping estimate, not a quote, and it moves with how many workflows you're already running. What it buys you is the ability to answer "what would a price change do to us?" in about a minute, with a real number.

Then re-run the question quarterly: at 2x, which workflows do I keep? At 4x? Write the answers down now, while nothing is on fire. That list is your plan, and it's cheap to make.

One housekeeping note. Any specific token price you find in a blog post — including ours, which is why there isn't one here — has a shelf life measured in weeks. Check the provider's own pricing page before you put a number in a budget.

---

Want the ROI math before you commit? Our AI Readiness & ROI Calculator is a free five-minute assessment that projects your return over 24 months, so you can see whether an automation still pays for itself with room to spare.