The headline sounds like good news: AI model prices dropped 98% over the last four years. GPT-4-equivalent performance that cost $20 per million tokens in 2022 now runs about $0.40. You would expect enterprise AI bills to follow those prices down.

They went the other way. The average enterprise AI budget grew from $1.2 million per year in 2024 to $7 million in 2026. Uber burned through its entire 2026 AI coding budget by April. Microsoft revoked developer access to Claude Code months after enabling it, citing unexpected cost overruns. The FinOps Foundation’s 2026 State of FinOps report found that 73% of enterprises reported their AI costs exceeded original projections.

For accounting firms, this is not background noise. It is an active risk in your next renewal cycle, and most firms are not positioned to manage it.

Why Bills Tripled While Prices Fell

The math seems contradictory until you understand what changed about how AI gets used.

A simple chatbot query, a one-turn question and answer, costs pennies. The model takes your prompt, runs it once, returns a result. That was the original enterprise AI use case in 2023 and 2024: employees asking questions, getting answers, moving on.

Agentic AI works differently. A workflow agent takes a task, breaks it into steps, calls tools and external systems, reasons about intermediate results, retries when something fails, and loops until the task is complete. Each of those steps consumes tokens. An orchestrated agentic workflow in 2026 costs roughly $1.20 per execution on average, about 30 times more than a single-turn chatbot query.

You can see how this compounds. An accounting firm that deployed an AI assistant for staff questions had relatively predictable costs. The same firm that now runs AI agents through month-end close, document collection, and tax prep has a cost structure that scales with workflow volume, not just seat count.

The math that made the business case work at $X per seat per month breaks when you are paying per token per workflow.

The Pricing Model Shift Happening Now

Most enterprise AI contracts were written when the product was a chat interface with a fixed subscription. Vendors signed multi-year deals at predictable prices. Enterprises budgeted for them the same way they budgeted for other SaaS.

That model is changing because the underlying economics changed. When an AI provider builds an agentic product, their cost is not static per user. It scales with usage. A user who runs one agent workflow per day costs them roughly the same as a user who runs one query per day. A user who runs 200 agent workflows per day during tax season costs them 200 times more.

Salesforce responded to this by introducing resolution-based pricing for its Agentforce Help Agent: enterprises pay only when the AI autonomously resolves a customer issue from start to finish, not for queries that bounce to a human. Anthropic has moved to usage-based billing for enterprise customers. HubSpot is restructuring its AI features around outcome-based pricing.

The implication is clear: flat-fee AI subscriptions are becoming the exception. Token-based, usage-based, and outcome-based pricing are becoming the default.

For accounting firms that signed flat-fee contracts during 2024 or early 2025, this shift matters at renewal. You may be moving from a predictable $800 per user per month to a variable bill that spikes when usage spikes. And for accounting firms, usage spikes on a very predictable schedule.

Why Accounting Firms Are Particularly Exposed

Tax season creates a usage profile that most AI vendor contracts were not designed for.

From January through April, an accounting firm using AI for document collection, data entry, return preparation, and client communication might push through 10 to 20 times its off-peak workload. A firm that processes 2,000 returns during tax season is running AI workflows on 2,000 returns in four months, not spread evenly across 12.

Under a flat-fee model, this is fine. You pay the same whether your agents are busy or idle. Under usage-based pricing, you pay for what you use. Tax season bills could come in at three to five times what you pay in July.

This is not a hypothetical. It is already happening at firms that moved to newer AI platforms in 2025 and 2026. The bills in February and March looked nothing like the bills in August. In some cases, firms were not tracking usage at all until the invoice arrived.

If you are renewing an AI contract in the next six months, you need to know what pricing model you are agreeing to and what your peak-month usage actually looks like.

The Three-Step Audit You Should Run Before Your Next Renewal

The goal is to understand your cost exposure before the contract is signed, not after.

Step one: Map your AI workflows to usage categories.

List every AI tool your firm uses and categorize how each one is priced. Flat per-seat subscription. Per-query or per-token usage. Outcome-based pricing per resolution. Hybrid with a base fee plus usage overages.

For each tool, estimate the workflow volume difference between your slowest month (probably July or August) and your busiest month (probably February or March). This ratio is your peak exposure multiplier.

Step two: Request usage data from your current vendors.

Most AI vendors can export token consumption or workflow execution logs. Ask for your last 12 months of usage data and look at the monthly distribution. If your current contracts are flat-fee but the vendor tracks usage, they can show you what you would have paid under usage-based pricing.

If your vendor cannot or will not provide this data, that is itself useful information for your negotiation.

Step three: Model the contract scenarios.

Take your peak-month usage estimate and apply the pricing model from each vendor’s renewal offer. Calculate the annual cost under three scenarios: current average usage, 150% of current usage, and peak-season usage only. The gap between those scenarios is your budget risk.

If the gap is material, negotiate for usage caps, seasonal rate adjustments, or pre-purchased token blocks at a discount. Vendors are often willing to negotiate on peak pricing if you can show them a predictable volume projection and commit to a floor.

What Good AI Contract Terms Look Like

Firms that have navigated this already have a few things in common in how they structured their vendor agreements.

The first is a consumption cap with notification. They get an alert when usage hits 80% of their monthly or quarterly budget, and the vendor does not auto-charge beyond the cap without explicit approval. This gives the finance partner time to decide whether to expand the budget or throttle usage.

The second is a seasonal rate adjustment. Some vendors will offer lower per-token rates during high-volume months in exchange for a minimum annual commitment. This functions like peak pricing in reverse: you pay slightly more during slow months but get a discount when you need the capacity.

The third is a vendor audit right. The firm can request usage data at any time and validate that what they are being charged matches what their systems actually used. This matters more than it sounds because agentic systems can behave unexpectedly in edge cases, and an agent that gets stuck in a retry loop can consume tokens in ways the original workflow design did not anticipate.

What This Means for Firms Evaluating New AI Tools

If you are evaluating AI platforms for accounting workflows in the next six months, pricing model transparency should be a first-tier requirement, not an afterthought.

Ask every vendor three specific questions before the demo. First, how does pricing scale when usage doubles? Second, what happens to my bill during a peak month when workflow volume is five times higher than average? Third, what controls do I have to cap spend if an agent behaves unexpectedly?

Vendors who cannot answer these questions clearly are not ready to serve a firm with a highly seasonal business model. Accounting firms live in cycles. January through April is not like June through September. Any AI vendor who does not have a clean answer for how their pricing handles that cycle is selling you a contract they may not be able to honor at the price you expect.

The Practical Response

None of this means you should slow down AI adoption. The productivity gains are real. The capacity increases are measurable. The ROI is there for the workflows that matter most.

What it means is that you need to treat AI vendor contracts the way you treat any other variable-cost business relationship: with clear terms, usage visibility, and defined cost controls.

The firms that will get hurt are the ones who signed usage-based agreements without understanding their peak-season exposure, and who will see their February or March AI bill for the first time when they open the invoice.

The firms that will benefit are the ones who run the audit now, negotiate the contract terms that match their workflow profile, and build the monitoring into their operations before the next tax season.

If you want a structured way to map your firm’s AI workflow costs and model the different pricing scenarios, the Omni Audit for accounting and bookkeeping covers this as part of the workflow assessment. It takes 60 minutes and the output includes a cost model based on your actual workflow volumes, not industry averages. Book a session and we will build it with you.

The pricing shift is happening whether you audit your contracts or not. The only question is whether you find out in the negotiation room or on the invoice.