Get FP&A best practices, research reports, and more delivered to your inbox.
To reduce AI costs without losing the productivity gains, first work out whether your AI bill is seat-heavy or usage-heavy, because the levers are different. Seat costs come down through license cleanup, tiering and renegotiation. Usage costs come down by routing routine work to cheaper models, caching repeated prompts and putting controls on agents.
The split matters more than any single tactic. In an illustration from our budgeting for AI spend guide, the same 30% model right-sizing cuts 2.4% of a seat-heavy $1 million bill and 18.6% of a usage-heavy one. Then protect the spend that produces the gains: your heaviest users and a controlled experimentation budget.
Bottom line: Sort your AI bill into seats and metered usage before you cut. Clean up and renegotiate seats, right-size and cache usage, and leave power users and a ring-fenced experiment budget alone, because that's usually where the productivity gains come from.
Is your AI bill seat-heavy or usage-heavy?
Look at how much of the bill moves with consumption. Seat-heavy bills are mostly fixed subscriptions: ChatGPT, Claude or Copilot seats, AI add-ons to existing tools, and coding assistants priced per user. Usage-heavy bills are mostly metered: API and model charges, tokens consumed by agents and workflows, and inference that runs inside your product.
To find your split, take the last three months of AI spend and tag each line as fixed or metered. Your provider exports show what was consumed and your general ledger shows what was booked; our guide to getting spend data from AI platforms covers the exports for Claude, ChatGPT and Cursor. If you've already split the bill by department, the work is mostly done: our guide to allocating AI costs across departments separates seats from usage as its first step.
Most companies land in between. A software company running AI features for customers will usually be usage-heavy; a company whose employees mostly use chat assistants will be seat-heavy. Many have a seat-heavy bill in G&A and a usage-heavy one in engineering and product, which is a reason to run the levers below by cost line, not for the company as a whole.
How much can model right-sizing save?
It depends on how much of your bill is metered. The guide's illustration holds everything else equal: two companies each spend $1.0 million on AI, and both find $230,000 of other savings. The only difference is how much of the bill moves with usage:
| Seat-heavy bill (profile A) | Usage-heavy bill (profile B) | |
|---|---|---|
| Metered share of the bill | 8% | 62% |
| Current AI spend | $1.0M | $1.0M |
| Other savings (seats, tiers, renegotiation, caching, agent controls) | -$230K | -$230K |
| Model right-sizing (30% of the metered share) | -$24K | -$186K |
| Spend after | $746K | $584K |
| Right-sizing as a share of total spend | 2.4% | 18.6% |
Illustrative scenario from Aleph's budgeting for AI spend guide, not a benchmark. Right-sizing is modeled as a flat 30% reduction on the metered share of the bill: 30% of 8% for profile A, and 30% of 62% for profile B.
For the seat-heavy company, a model-routing project would be a lot of engineering effort for $24,000. For the usage-heavy company, it's worth $186,000 a year and should be near the top of the list.
How do you cut AI seat costs?
Treat AI seats like any other software license: find what's unused, match the tier to the user, and negotiate from usage data.
- Remove inactive seats. Pull last-active dates from each provider's admin console and remove anyone who hasn't used a tool in 60 or 90 days. Check for people who've left the company: our guide notes that seat holders should be checked against the HR roster every month.
- Consolidate overlapping tools. Teams often pay for two or three chat assistants that do the same job. Pick a standard per use case and move people onto it at renewal.
- Match the tier to the user. Not everyone needs the enterprise or "max" tier. Keep premium tiers for people whose usage shows they hit the limits.
- Renegotiate at renewal with usage in hand. Know how many seats were active, how often they were used and what you'd move to instead. Decide in advance what you'd trade for a lower price, such as a longer term.
Watch for AI that arrives inside software you already buy, as a seat upgrade, a credit pool or a per-action charge. It won't show up as an AI vendor, so check your largest renewals for new AI terms before you sign.
How do you cut AI usage costs?
Usage costs respond to engineering levers, and finance's job is to know which ones exist and to ask for the numbers. The FinOps Foundation's FinOps for AI overview lists the main ones:
- Right-size the model. Define the quality the task needs, then use the least expensive model that meets it. Routine work such as classification, extraction and summarization often runs well on a smaller model.
- Route requests automatically. Model routing sends simple requests to small models and keeps frontier models for the work where the extra capability changes the outcome.
- Cache stable prompts. Prompt caching cuts the cost of repeated input, such as the same instructions or reference document sent with every request.
- Batch work that can wait. Overnight reports and bulk processing can run through lower-cost batch processing.
- Use tokens efficiently. Trim system prompts, cap output length and retrieve only the content the model needs.
- Put limits on agents. Approved model lists, quotas, rate limits and budget thresholds stop a runaway agent loop from doubling the bill overnight.
For what each model costs today, see the cost section of our comparison of the best LLMs for finance teams. Prices change often, so check them before you model the savings.
Which AI costs should you not cut?
Your heaviest users and a controlled experimentation budget. Both look like waste on a cost report, and both are often where the productivity gains come from.
AI usage is concentrated. In LayerX's 2026 data, cited in our budgeting for AI spend guide, the median enterprise AI user had 12 conversations, while the top 5% had 144 or more. Heavy token use can mean waste, but it can also mean someone has found a far more productive way to work. Before you cap a heavy user, look at what they're producing.
The same goes for experiments. Set aside a fixed budget for new tools, models and use cases, and judge it by what you learn and what graduates into real workflows, not by whether each experiment pays back immediately. A ring-fenced budget also keeps experiments from spreading across every team's card spend.
How do you cut AI costs without losing the productivity gains?
Decide what each AI investment is supposed to improve before you cut it, and watch that number after. If support AI was bought to lower cost per ticket, right-size its models and check that resolution time and cost per ticket hold. If they slip, the cut went too far. Our guide to showing the board your AI spend is worth it covers how to set that baseline.
Three rules keep cost reduction from eating the gains:
- Cut waste first, capability last. Inactive seats, duplicate tools and unmanaged agent loops cost money and produce nothing. Remove them before touching anything people use.
- Change one lever at a time. If you move a workflow to a cheaper model and add caching in the same week, you won't know which one hurt quality.
- Make spend visible before you charge it back. The FinOps Foundation recommends reporting through showback first, and notes that visibility alone often changes behavior. A monthly view of AI spend by team can trim more than a policy memo.
Worked example: trimming a $600,000 AI bill
Say a company spends $600,000 a year on AI: $250,000 on seats and subscriptions, $280,000 on API and agent usage, and $70,000 on coding tools.
Seats. An admin-console check finds 12% of chat seats inactive for 90 days, worth $30,000 a year. Two teams pay for overlapping assistants; consolidating at renewal saves another $15,000. Seat savings: $45,000, or 18% of the seat line.
Usage. Engineering estimates that most of the API usage is routine extraction and summarization that could run on a smaller model. Using the guide's 30% right-sizing assumption on the $280,000 usage line gives about $84,000. Caching and batching come on top, but measure them on your own workloads before you put them in the plan.
Coding tools. Left alone. The $70,000 is concentrated in a handful of engineers whose output the team can point to, which is the power-user case above.
The result is about $129,000 a year, roughly 21% of the bill, without cutting a single active user. The figures are illustrative; the order of operations is the point. Then build the lower run-rate into next year's number: our guide to forecasting AI spend covers how to forecast seats, usage and agents from their own drivers.
How should you set an AI cost reduction target?
Set it by line and by unit, not as one number for the whole bill. A single target such as "cut AI spend 15%" ignores that usage is supposed to grow when the business grows. If customer activity doubles, a usage line that holds flat isn't a saving; it's a sign something stopped working.
Use two kinds of target instead:
- Seats: an absolute target. Seat spend should track headcount and adoption. Set a target for active-seat utilization, such as every paid seat used in the last 30 days, and a dollar target for consolidation at each renewal.
- Usage: a unit-cost target. Measure cost per unit of work: cost per support ticket resolved, per document processed, per agent run or per customer request. Our guide forecasts these lines the same way, as volume times cost per run, so the target and the forecast use the same driver.
A unit-cost target lets you say something a board can check, for example: "cost per ticket fell 12% while ticket volume grew 30%." That's a saving even though the total bill went up.
How do you reduce AI costs inside your product?
Treat customer-facing AI as cost of goods sold and manage it as a gross margin line. Model inference, customer-triggered agents and AI-specific hosting rise with customer usage, so every saving per request goes to margin and grows with you.
The levers are the usage levers above, applied with more rigor: right-size the model for each feature, cache the instructions and reference data each request carries, and batch anything the customer isn't waiting for. Two more belong to product and pricing rather than engineering:
- Know the cost per customer action. If one feature costs ten times more per use than another, that's a pricing and packaging question as much as an engineering one.
- Match plans to usage. Usage limits or credit allowances in your own pricing keep a small number of heavy customers from eroding the margin on everyone else.
Who should own AI cost reduction?
Finance should own the number, and the teams that consume AI should own the levers. In practice that means three owners:
- Finance keeps one reconciled AI spend total, tags each line as seat or usage, sets targets and reports progress monthly. Finance also decides where each cost lands on the P&L; our guide to whether AI spend is COGS or OPEX covers the classification.
- IT and procurement own seat cleanup, tool consolidation and renewals.
- Engineering owns model routing, caching, batching and agent limits, because those are code changes.
Without a named owner for each lever, AI cost reduction becomes a one-off exercise before budget season, and the bill drifts back up within a few months.
Common mistakes when cutting AI costs
- Applying one percentage cut across the board. A flat 20% cut hits a seat-heavy team and a usage-heavy team in completely different ways.
- Capping heavy users without looking at their output. You may be cutting your most productive people.
- Renegotiating without usage data. Vendors know your utilization; walk in knowing it too.
- Counting only the obvious invoices. AI bundled into existing contracts, coding tools on cards and agent usage on cloud bills are easy to miss.
- Cutting before measuring value. Without a baseline, you can't tell whether a cut cost you anything.
Build an AI spend model with Aleph
Aleph connects your AI providers, general ledger and HRIS to Excel, so AI spend refreshes on its own, splits by team and seat or usage, and sits next to the rest of the P&L with variance against plan. Get the full method in our budgeting for AI spend guide, watch the webinar from AI spend to AI budget, or book a demo to see it on your own data.
Get FP&A best practices, research reports, and more delivered to your inbox.




