How to reduce AI costs for work?
Match the model to the task instead of defaulting to the biggest one. Shopify cut costs 68% and ran 2.2x faster by fine-tuning a 32-billion-parameter open model for one narrow job. The other big levers: shrink your prompts, cache repeated context, batch non-urgent work, and stop paying for seats nobody opens.
Why — the first-principles explanation
You pay for tokens, not for answers. Every word in your prompt, your pasted document, your system instructions, and the model's reply gets billed. This single fact explains almost every AI bill that comes in higher than expected: teams stuff entire knowledge bases into every request, then pay for that context on every single call, forever. Retrieving the three relevant paragraphs instead of pasting all 300 pages isn't a clever optimization — it's the difference between a bill that scales and one that doesn't.
The second lever is that model capability is priced steeply, and most work doesn't need the top of the range. Frontier models cost dramatically more per token than small ones, and firms route everything to the frontier because it's simpler. But classifying a support ticket, extracting a date, or tagging a product doesn't require a model that can do competition math. Shopify's Flow team is the cleanest public proof: they fine-tuned Qwen3-32B for one specific job and got 2.2x the speed at 68% lower cost than their previous approach — while improving accuracy, because a model trained on your task beats a generalist at your task.
The third lever is time. Providers price urgency: cached context is far cheaper than fresh context, and batch processing — where you accept results later instead of instantly — carries steep discounts. A nightly categorization job has no business paying the interactive rate. Ask of every workload: does a human need this in two seconds, or does a dashboard need it by morning?
And the lever nobody wants to look at: seats. In most companies the largest AI line item isn't inference at all — it's per-user subscriptions bought during a pilot and never reconciled against usage. Pull the login report before you optimize a single prompt. Fifty unused seats will dwarf whatever you save by trimming system prompts.
An example that makes it click
It's a taxi meter, not a monthly bus pass. You're charged for distance, so people who assume it's a flat fee end up horrified. And the biggest waste isn't the trips you take — it's the luggage you bring. Hauling your whole filing cabinet in the back seat on every single ride, just in case someone asks about the Henderson account, means you pay to move that cabinet every time.
So: take the one folder you need, not the cabinet. Take a hatchback to pick up milk instead of a limousine. And if the delivery isn't urgent, send it on the overnight truck for a fraction of the price. Then go check how many limousines are idling in the parking lot with nobody in them — that's your seat licenses.
How to do it
- Audit seats first: pull the usage report and cancel licenses for anyone who hasn't logged in for 30 days. This is usually the single largest saving and takes an afternoon.
- Measure before you cut: find your top 3 workloads by token spend. Most teams discover one runaway job is the majority of the bill.
- Right-size the model: move classification, extraction, tagging, and routing to a smaller or fine-tuned model; reserve frontier models for open-ended reasoning.
- Shrink the context: retrieve only the relevant passages instead of pasting whole documents. Trim bloated system prompts — you pay for them on every call.
- Cache what repeats: if the same instructions or reference documents lead every request, use prompt caching so you aren't billed full price for identical context.
- Batch what can wait: route non-interactive jobs — nightly reports, bulk tagging, backfills — to batch processing rather than real-time endpoints.
- Cap and alert: set hard spend limits and per-key budgets so a looping agent can't quietly run up a five-figure bill over a weekend.
- Re-check quarterly: prices and model tiers change often, so verify current rates on the provider's official pricing page rather than trusting last quarter's spreadsheet.
Key facts
- Shopify's fine-tuned Qwen3-32B workflow agent runs 2.2x faster and 68% cheaper than the model it replaced, while improving correctness on its target task.
- Shopify improved syntactic correctness by 22 points and semantic correctness by 13 points simply by changing the output format the model generated (Python instead of JSON) — an accuracy gain with no extra compute.
- Billing is per token across input and output, so system prompts and pasted context are charged on every call, not once.
- Shopify retrains its production model weekly on real merchant data, and found a 35% gap between benchmark scores and real-world success — meaning cheaper models must be validated on your actual traffic, not benchmarks.
- Provider pricing, model tiers, caching discounts, and batch discounts change frequently; verify against official pricing pages (as of 2026-07).
▶ The 60-second explainer (script)
Want to cut your AI bill at work? Start with the one fact that explains almost every surprise invoice: you pay per token, not per answer. Every word in your prompt, every pasted document, every system instruction — all billed, on every single call. So teams that stuff a whole knowledge base into each request pay for that knowledge base forever. Retrieve the three paragraphs you need, not the three hundred pages. Second: stop sending everything to the biggest model. Frontier models cost far more per token, and most work doesn't need them. Classifying a ticket or extracting a date does not require a model that can do olympiad math. Shopify proved this publicly — they fine-tuned a thirty-two billion parameter open model for one narrow job and got two-point-two times the speed at sixty-eight percent lower cost, and it was more accurate, because a model trained on your task beats a generalist at your task. Third: providers charge for urgency. Cached context is cheaper than fresh context. Batch processing, where you get results later instead of instantly, is heavily discounted. Your nightly tagging job has no business paying the interactive rate. But here's the one nobody wants to hear. In most companies the biggest AI line item isn't inference at all — it's seats. Licenses bought during a pilot that nobody ever opens. Before you optimize a single prompt, pull the login report. Fifty dead seats will dwarf everything else you do.
What authoritative sources say
People also ask
What's the fastest way to cut an AI bill?
Cancel unused seats. It requires no engineering, takes an afternoon, and in most organizations it's a bigger line item than inference itself.
Are cheaper models actually good enough?
For narrow, well-defined tasks, frequently yes — Shopify's fine-tuned 32B model beat their previous setup on cost, speed, and accuracy. Validate on your real traffic, not benchmarks; Shopify found a 35% gap between the two.
Does a longer prompt really cost more?
Yes, on every single call. A bloated system prompt is a fixed tax on all your traffic, which is why trimming context and retrieving only relevant passages compounds.
Should I self-host to save money?
Only at high, steady volume. Below that, GPU capacity sits idle and you've traded a per-token bill for a fixed one plus engineering time — often a worse deal.
How do I stop runaway agent costs?
Set hard spend caps and per-key budgets with alerts. A looping agent can generate an enormous bill over a weekend, and no amount of prompt tuning protects you from that.