AI Agent Pricing Models Explained: Per-Agent, Per-Task, Per-Seat

Why AI agent pricing is being rewritten in 2026

Software-as-a-service pricing has three properties that made per-seat billing work for two decades: software had near-zero marginal cost, every user consumed roughly similar value, and humans did the work. AI agents break all three.

First, agents have real marginal cost: every interaction consumes LLM tokens, tools, and compute. Second, one agent can replace ten users — so charging per seat no longer maps to value. Third, agents do the work, not the user. Billing a company for ten "seats" when only one person supervises a fleet of agents makes no sense on either side of the table.

That statistic is the whole industry waking up to the fact that the economic model underneath SaaS has shifted. In response, vendors are experimenting — sometimes aggressively — with five distinct approaches. This post breaks down each one, the math behind it, and when to pick it.

Per-seat pricing

Still the most common model in 2026, per-seat pricing charges a fixed fee per named user per month. The company's bill scales with headcount.

When per-seat still works

• The agent is a co-pilot (Copilot, Glean Assist, Notion AI) that augments a specific user.
• Each user gets meaningful individual value — not shared across the org.
• Usage is predictable and roughly equal across users.

When per-seat fails

• The agent is autonomous — it does work without a human operator.
• One agent serves an entire department or customer base.
• The work volume is wildly uneven between users.

Typical 2026 range: USD 25 to USD 60 per user per month for horizontal co-pilots; USD 100 to USD 400 per user per month for vertical, specialised agents used by professionals (legal, medical, financial).

Per-agent pricing

Here the customer pays a fixed fee per deployed agent, regardless of how many users interact with it or how many tasks it runs within a soft cap. This is increasingly common for productised agents that act as "virtual employees."

Why vendors like it

• Maps directly to the "AI teammate" mental model customers already have.
• Predictable revenue per unit.
• Upsell is clear: need more capacity, spin up another agent.

Why buyers like it

• Price comparable to hiring a person — easy internal sell.
• Decouples cost from headcount.
• Simple budgeting: you know how many agents you have.

The risk: per-agent pricing conceals the marginal cost of tokens and integrations. Vendors absorb that on their margin until one "agent" running at 10x expected volume turns unprofitable. Most vendors therefore publish per-agent prices with soft caps on monthly tasks or tokens.

Per-task (per-action) pricing

Usage-based. The customer pays each time the agent completes a defined action — a ticket resolved, an email drafted, a meeting booked, a lead enriched. This is the model that most closely mirrors infrastructure-style AI APIs.

Strengths

• Direct alignment between spend and workload — you pay for what you use.
• Scales gracefully with seasonality.
• Easy to compare across vendors.

Weaknesses

• Hard to budget if your volume is volatile.
• Vendors are tempted to over-count "tasks" — watch the definition carefully.
• Customer psychology: users hesitate to invoke the agent because each call feels like a meter running.

Typical 2026 rates: USD 0.08 to USD 0.60 per task for lightweight actions (email drafts, lookups), USD 2 to USD 15 per task for heavier workflows (full research reports, multi-tool orchestration). Sales and customer service agents usually sit in the USD 0.30 to USD 3 per-conversation range.

Outcome-based pricing

The hottest — and most contested — model of 2026. The customer pays when a defined business outcome is achieved. No outcome, no charge.

Where outcome pricing has landed

• Customer support: per-resolution, typically USD 0.80 – USD 3.50 per fully-resolved ticket.
• Sales agents: per-qualified-meeting, typically USD 40 – USD 180 per meeting booked.
• Recruiting agents: per-screened or per-hired candidate, usually USD 15 – USD 800 depending on scope.
• Collections agents: percentage of recovered amount, typically 5 – 12%.

Outcome pricing is attractive because it aligns incentives, but it breaks in three places: attribution (was it the agent or something else?), definition (what exactly counts as "resolved"?), and averages (does the math work at your volume?). Vendors who succeed with it usually also publish a platform-minimum to protect themselves from customers who never reach threshold volume.

Hybrid and platform-fee models

Most mature vendors now blend models. The typical 2026 enterprise structure looks like this:

1. Platform fee — fixed monthly cost that covers access, onboarding, base support.
2. Usage allowance — a bucket of tasks, tokens, or agent-hours included in the platform fee.
3. Overage rate — price per task beyond the included bucket.
4. Outcome bonus or discount — an adjustment tied to achieved outcomes.

This is effectively the model telco and cloud providers converged on 20 years ago. It gives vendors floor revenue, gives customers predictability, and lets the relationship flex with usage.

How custom AI agencies price (including Bananalabs)

Custom AI agencies sit in a different category from SaaS vendors because the deliverable is different — not access to a shared platform, but a bespoke agent owned by the customer, integrated with their systems, maintained over time.

The standard agency structure

1. Discovery and scoping engagement. Typically a fixed fee or absorbed into build.
2. Build fee. Fixed-scope deliverable covering design, development, evaluation, integration, and handover.
3. Operating retainer. Monthly fee covering monitoring, prompt maintenance, integration upkeep, model updates, incident response, and agreed enhancement hours.
4. Pass-through costs. LLM tokens, vector storage, and hosting billed at cost or bundled.

The advantage for buyers is that the deliverable is owned, not rented. The business logic, prompts, integrations, and data belong to the customer. Unlike SaaS, there is no "platform lock-in" beyond the agency retainer you can replace with an internal team once the system is stable.

Bananalabs does not publish rate cards because every build is scoped to a specific business context. But the 24-month cost structure for a typical client looks nothing like "USD X per seat per month" — it looks like a deliverable with a known build cost and a predictable monthly operating cost. We believe this is the most honest structure for a bespoke agent. For more on how we work, see custom AI agents vs off-the-shelf tools.

Side-by-side comparison of 2026 AI agent pricing models

How to pick the right AI agent pricing model

Picking the right model depends on three variables: who uses the agent, how volume behaves, and whether outcomes are attributable. Walk through this decision framework.

Step 1: Who uses the agent?

• Named humans (employees): per-seat or hybrid.
• End customers (e.g. support chatbot): per-task, per-resolution, or hybrid.
• Nobody — it runs autonomously: per-agent or outcome.

Step 2: How does volume behave?

• Predictable and steady: per-seat or per-agent win.
• Spiky or seasonal: per-task aligns cost to reality.
• Scale-dependent ROI: hybrid with committed-use discount.

Step 3: Are outcomes attributable?

• Yes, cleanly: outcome-based pricing is the most aligned.
• Partially: hybrid with an outcome bonus.
• No: stick with per-task or per-agent.

Step 4: Do you need the agent to be custom?

If yes — if the agent has to deeply integrate with your stack, handle your specific data, or carry your brand — then platform pricing usually hits a ceiling. A fixed-scope build plus retainer becomes the more efficient model. This is the sweet spot for custom AI agencies.

What to watch for in AI agent contracts

Whatever model you pick, the contract matters as much as the sticker. Watch for:

• Task definition. What exactly counts as a "task"? A complete interaction, or every tool call?
• Cap structure. Are there soft caps that convert to hard overages?
• Model change rights. Can the vendor switch underlying models mid-contract, and what happens to quality and cost?
• Data ownership. Who owns the conversations, the training data, and the fine-tuned weights?
• Exit provisions. Can you extract your data and prompts if you leave?
• SLA and uptime. Most AI agent SLAs are weaker than standard SaaS.
• Price review clauses. Many 2026 AI contracts include annual uplifts explicitly tied to model cost increases.

Where AI agent pricing is going next

1. The decline of per-seat continues

Per-seat will not disappear for co-pilots, but for autonomous agents it is already falling below 30% of the market. Expect by 2028 for per-seat to be the minority model overall.

2. Outcome pricing matures with better attribution

Attribution infrastructure — event tracking, causal analysis — is the unlock for outcome-based pricing. As that tooling improves, outcome deals will move from 5% of the market to 15-25% by 2028.

3. Consumption pricing splits the bundle

Expect to see more vendors itemise tokens, tool calls, and storage separately. That is uncomfortable for buyers but accurate to how the underlying cost works.

4. Performance tiers emerge

Vendors will offer "fast / cheap / accurate" tiers that are essentially different model routes. Buyers will pick the tier by task type. This is already visible with major frontier labs.

5. Committed-use discounts normalise

AWS-style reservation pricing — commit to X tokens per year, get a 20–40% discount — is showing up across all tiers of AI agent vendor. If you have steady volume, you will leave money on the table by not asking for it.

The bottom line on AI agent pricing models

The right pricing model is the one that aligns your cost with the value you actually get — not the model that was convenient for the vendor's last product. In 2026, that usually means moving off pure per-seat, considering per-agent or per-task for productised agents, and preferring build-plus-retainer for anything custom.

The worst trap is picking a model because "that is how we always buy software." AI agents do not behave like software; they behave like hired work. Price them like it. For a deeper dive on the economics behind these models, read our pieces on the hidden costs of building AI agents and real 2026 AI agent ROI numbers.