How much should an AI agent cost?

Why AI Agent Pricing Breaks the Old SaaS Playbook

Ask a support agent a one-line question and it burns roughly 400 tokens. Hand it a hard ticket and the same agent may chew through 40,000 tokens across a dozen tool calls and three model retries. That hundredfold spread is why the SaaS pricing playbook — flat fees, per-seat licenses, simple tiers — falls apart the moment software starts doing the work itself.

The math underneath an agent is fundamentally different. Traditional SaaS charged for access: one user, one seat, one predictable cost. Agents autonomously execute multi-step workflows with wildly variable effort, and as billing platform Lago puts it, "the underlying cost structure is usage-shaped, so the price has to be too." A flat "unlimited" plan either bleeds margin on heavy users or quietly caps them later.

Per-seat pricing has a deeper problem: agents are designed to replace seats, not sit on them. If an agent handles the workload of three support reps, who exactly is the user you're paying for? Even products that look seat-based have backed into usage gates underneath. Cursor's $20/month Pro and $40/user/month Business plans are nominally per-seat, but hidden "Agent request" limits sit beneath the surface — and its shift away from "unlimited" plans triggered real customer churn debates, with Cursor investor Theo Browne noting the move away from loss leaders would "screw a lot of people."

So vendors have experimented with units that map to value instead of access:

• Outcomes — Intercom's Fin charges $0.99 per fully resolved customer issue, so you pay only when the work is actually done.
• Actions and credits — charging per workflow run or per credit consumed, which is easier to meter but harder for buyers to predict.
• Hybrids — a predictable base fee with a usage tail, the structure many billing platforms now recommend.
• Hours — Retool argues agents should be priced like contract workers against a mid-level analyst's $50–80/hour fully loaded cost.

Each unit trades off accuracy against intelligibility. Tokens are cost-accurate but opaque; outcomes are what customers actually want to pay for but the hardest to bill with certainty. As Lago's CEO warns, "get the unit wrong and you either bleed margin on every heavy user or price out the customers who'd become your best ones."

That leaves buyers with the core confusion: when you pay $20 a month, or $0.99 a resolution, or per credit — what are you actually paying for? Effort, output, or outcomes? The honest answer is that most pricing today is a proxy, and the proxies don't align. One practical lever you do control is how much work your agent wastes: agent costs scale with steps and tool calls, so equipping each agent with only the skills and tools it needs — something SkillzBench is building a free directory and MCP connector for — directly trims the bill. Understanding what each pricing model really charges for is the first step to paying less for the same work.

The Three Pricing Models (Plus One Wildcard) and What They Mean for Your Budget

AI agent pricing has abandoned the flat monthly fee, and for good reason: no two commands create the same amount of work for an agent. A hard support ticket can burn ~40,000 tokens across a dozen tool calls and three model retries, while a one-line question uses ~400 tokens — so the price has to be as usage-shaped as the underlying cost, as billing analysis from Lago puts it.

Three models now dominate the market. The first is outcome-based pricing, where you pay only for results. Intercom's Fin agent charges $0.99 per fully resolved customer issue, which makes budgeting trivially predictable — you never pay for a failed attempt, according to Chargebee's pricing playbook. The second is action- or credit-based pricing, where each workflow run or credit consumed is metered; n8n bills per workflow run, and many platforms wrap raw token costs into credit systems.

The third, and arguably the safest for buyers, is the hybrid: a predictable base fee plus a usage tail. Relevance pairs a flat fee with included seats and credits, while Lovable bundles credits into a recurring per-user fee with top-ups. Schematic recommends this structure precisely because it smooths the spikes of pure usage billing.

Then there's the wildcard. Retool argues that tokens, credits, and API calls "bear no relationship to actual output," and prices its agents by the hour instead — benchmarked against the $50–80/hour fully loaded cost of a mid-level analyst, so companies can evaluate agents like contract workers. It's worth noting this comes from a vendor selling hourly-priced agents, not neutral analysis.

• Developers tend to prefer pay-as-you-go per-call or per-token pricing, which mirrors the API economics they already know.
• Business buyers favor subscriptions with included credits plus metered overage — predictable enough to expense, flexible enough to scale.
• Enterprises gravitate toward committed usage tiers with true-up contracts, trading flexibility for negotiated rates.

Whichever model you choose, benchmark the cost against human labor, not other software. Toshiba's Copilot deployment for 10,000 employees saved 672,000 hours a year — roughly 323 full-time equivalents — which shows the right comparison is the cost of the work being replaced or multiplied, not last year's SaaS subscription.

One caveat: beware of hidden usage gates. Cursor's $20/month Pro and $40/user/month Business plans look seat-based but are gated by "Agent request" limits underneath, and the shift from "unlimited" plans triggered real customer churn. And because agent costs scale with every step and tool call, curating exactly which skills your agent can see directly reduces wasted spend — that's the problem SkillzBench is building for, with a free directory of verified, benchmarked agent skills and a free MCP connector to equip each agent with only what you choose.

The Hidden Costs Vendors Don't Put on the Pricing Page

The sticker price is only the beginning. AI agents introduce billing hazards that traditional SaaS never had to worry about — and most of them never appear on the pricing page.

The first hazard is concurrency. Because agents run tasks autonomously and in parallel, several tasks can draw down the same balance at once, effectively double-spending your credits before you notice, a problem that usage-billing analysis flags as unique to agentic workloads. Traditional per-seat software simply couldn't do this.

The second is fan-out. One user action — a single click or prompt — can cascade into many billable sub-events: tool calls, context loads, model retries. A hard support ticket can burn roughly 40,000 tokens across a dozen tool calls and three model retries, while a one-line question burns about 400, as billing research documents. Same interface, hundredfold cost difference.

The Replit incident made this painfully visible. A developer asked the agent to change a button's color, and the request cost around $1 because the agent processed the full chained context of the session — a small change, a large bill, as reported in pricing analysis. Multiply that across a team and a month, and "it's just a usage fee" becomes a budget problem.

Retries deserve their own scrutiny. When an agent fails mid-workflow and re-runs, who pays? The ethical line, according to billing experts: if a retry happens because of the vendor's own infrastructure, that's the vendor's cost to absorb — meter and bill only on completed, successful runs.

Then there's the moving target problem. Every major model provider repriced, deprecated a tier, or changed token-counting at least once in the past year. Your contract may hold, but the underlying unit economics can shift underneath it.

Before signing, ask:

• How do you handle concurrent tasks drawing from the same balance — are there spend caps or alerts?
• What counts as a billable event when one action fans into multiple sub-steps?
• Are failed runs and infrastructure-caused retries billed?
• What happens to my effective rate if the underlying model provider reprices mid-contract?
• Is there a real-time usage dashboard, or will I discover overage on the invoice?

One structural lever buyers control: agent costs scale with the number of steps and tool calls an agent performs. Curating exactly which skills and tools your agent can see cuts wasted work at the source. SkillzBench is building a free directory of vetted agent skills plus a free MCP connector that lets you equip each agent with only the capabilities you choose — so the fan-out you pay for is the fan-out you actually wanted.

Why the Sticker Price Is Only 20% of the Real Cost

The $0.99-per-resolution price tag looks like the decision. It isn't. The number on the invoice is just the entry fee — the real question is what you spend around it.

The data is sobering. According to McKinsey findings cited in industry analysis, roughly 8 in 10 companies using generative AI report no significant bottom-line impact, and about 90% of function-specific use cases remain stuck in pilot mode. Meanwhile, fewer than 1 in 5 organizations report a meaningful earnings lift from AI, even as adoption in at least one business function has climbed from 33% to 78% of organizations, per CFO-focused research from Launch Consulting.

Yet the winners are spectacular. Toshiba's Copilot deployment across 10,000 employees saved 672,000 hours per year — the equivalent of roughly 323 full-time employees — while Novo Nordisk cut clinical study report creation from 12 weeks to 10 minutes, according to documented AI ROI case studies. At the enterprise scale reported at Davos, Foxconn and BCG's agent ecosystem automated 80% of decision-making and unlocked around $800 million in value, while Fujitsu cut warehousing costs by $15 million.

So what separates the 20% who win from the 80% who don't? Not the tool. The transformation around it.

Launch Consulting puts it bluntly: "Buying the tool is the ante that gets you a seat at the table, not a credible bet on the outcome." Their recommendation is that roughly 80% of your early AI budget — the first 3 to 12 months — should go to process transformation and adoption, not the subscription itself. As they warn, "If your AI business case is almost entirely tool and token cost, it's likely a business case for disappointment."

In practice, that budget split funds work the invoice never shows:

• Redesigning the workflow the agent will run, not bolting the agent onto an unchanged process
• Training and change management so people actually trust and use the agent's output
• Curating what the agent can see and do, since agent costs scale with the number of steps and tool calls it performs
• Measuring against the cost of the human work being replaced — a mid-level analyst runs $50–80/hour fully loaded, per Retool's benchmark — not against last year's SaaS line items

That third item deserves emphasis. Because agent costs are usage-shaped — a hard support ticket can burn ~40,000 tokens across a dozen tool calls while a simple question burns ~400 — every unnecessary skill, tool, or step an agent can access inflates both cost and failure risk. This is where disciplined setup pays: platforms like SkillzBench are building free tooling to verify and benchmark agent skills, including an MCP connector that lets you equip each agent with only the skills you choose, so the agent never wanders through tools it doesn't need.

The sticker price gets you the software. The other 80% gets you the ROI.

Your Action Plan: Cut Agent Costs by Controlling What Your Agent Sees

The cheapest agent isn't the one with the lowest sticker price — it's the one that wastes the fewest steps getting work done. A hard support ticket can burn roughly 40,000 tokens across a dozen tool calls and three model retries, while a simple question needs only about 400, according to usage-billing analysis from Lago. If you're paying per action or per token, every wasted tool call comes straight out of your budget.

Start with the pricing model itself. Outcome-based pricing — like Intercom's Fin agent at $0.99 per fully resolved issue — ties your cost directly to value delivered, as Chargebee's pricing playbook documents. Where pure outcome pricing isn't available, hybrid models (a predictable base fee plus a usage tail) beat raw metered billing for most business buyers. And whatever the model, demand a real-time usage dashboard before you sign — surprise invoices destroy trust faster than any price increase.

Your negotiation checklist should include:

• Outcome or hybrid pricing over raw per-token billing, so cost tracks value rather than effort.
• Completed runs only — if a retry fails because of the vendor's infrastructure, that's their cost to absorb, not yours.
• Usage dashboards with deduplication, since multi-step agents can fan one action into many billable sub-events.
• Written notice of pricing changes — every major model provider repriced or changed tiers at least once in the past year.

Then attack the demand side: what your agent actually sees. Agent costs scale with steps and tool calls, so an agent loaded with dozens of skills and connectors will explore, test, and retry far more than one equipped with exactly what the task requires. Curation is a cost-control lever most teams never pull.

That's the thinking behind SkillzBench, which is building a free, independent directory of agent skills — each version-pinned, security-scanned, license-checked, and benchmarked in a sandbox so you can see the measured Lift before committing. Its free MCP connector (in development) will let you toggle skills and tools per agent, so each agent only sees what you've deliberately chosen for it. Given that published research found roughly 26% of skills carry at least one vulnerability, vetting isn't just about cost — it's about not wiring risky components into an autonomous system.

The payoff compounds. Robinhood scaled daily token usage tenfold while cutting AI costs 80%, according to documented case studies — proof that disciplined architecture beats brute-force spending. Fewer wasted calls, fewer wasted dollars, and an agent bill you can actually explain to your CFO.