The dangerous part of AI agent pricing is not that vendors charge money. The dangerous part is that each pricing model quietly tells you what the vendor wants your architecture to become.
If you want the commercial pricing and cost-modeling page, start with AI agent pricing without the platform math trap. This article is the longer buyer guide behind that page: pricing models, hidden cost layers, and the architecture choices that change the three-year curve.
Per-resolution pricing pushes volume through the vendor. Per-action pricing asks you to count every step the agent takes. Platform fees make the first year feel predictable, then the migration cost becomes the renewal trap. A managed custom build moves spend toward implementation and operation, but gives you the chance to own the system.
This guide compares the pricing models buyers are seeing in 2026 and explains when each one makes sense.
AI Agent Pricing Models Compared
| Pricing model | Public examples | How it works | Best fit | Main risk |
|---|---|---|---|---|
| Per resolution | Intercom Fin lists $0.99 per resolution on its pricing page |
Pay when the AI resolves a customer conversation | Low to mid volume support teams with standard workflows | Outcome definition and volume spikes |
| Per conversation | Salesforce lists Agentforce Conversations at $2 per conversation on its pricing page |
Pay for each conversation package | Salesforce-heavy teams that want a native buying path | Cost rises with conversation volume |
| Per action or credit | Salesforce lists Flex Credits at $500 per 100k credits |
Pay for discrete agent actions | Teams with mixed employee and customer-facing agents | Harder forecasting if actions per case vary |
| Outcome-based enterprise | Sierra explains its outcome-based pricing model | Pay when the agent completes a defined business outcome | Large brands with clear support outcomes | Vendor controls outcome measurement |
| Platform fee plus usage | Common in enterprise AI support contracts | Annual platform access plus usage | Larger teams that want vendor-managed operations | High first-year cost and lock-in |
| Managed custom build | OpenNash model | Pay for discovery, build, deployment, monitoring, and improvement | Complex workflows, regulated teams, high volume | Requires a serious implementation partner |
The headline rate is only a starting point. A $0.99 resolution can be cheap if the agent solves simple tickets. It can be expensive if the workflow needs custom integrations, human review, compliance logging, and weekly tuning outside the vendor's standard package.
The Year-One Cost Stack
Most buyers compare vendor pricing as if the AI agent were a subscription. In practice, the first-year cost stack has several layers.
| Cost layer | What to ask |
|---|---|
| License or usage | What exactly is metered: conversation, resolution, action, seat, or outcome? |
| Minimum commitment | Is there an annual minimum, prepaid wallet, or contract floor? |
| Implementation | Who maps the workflow, connects systems, and tests edge cases? |
| Integrations | Are custom API connections included or billed as services? |
| Knowledge maintenance | Who updates source docs, procedures, and retrieval quality? |
| Human review | How are escalations, approvals, and QA queues staffed? |
| Analytics | Can you inspect traces, tool calls, cost, quality, and outcomes? |
| Migration | What do you keep if you leave? Code, data, prompts, logs, procedures? |
The migration row is where many buyers undercount risk. If the agent's memory, conversation data, procedures, and analytics live only inside a vendor platform, the contract is not just a subscription. It becomes the operating system for your customer workflow.
When Platform Pricing Wins
Platform pricing can be the right call.
Choose a platform when:
- The use case is common: password resets, order status, return policy, appointment scheduling.
- Your helpdesk or CRM already holds most of the workflow.
- Low volume makes per-resolution pricing affordable.
- Speed matters more than system ownership.
- The vendor's native integrations are enough.
- You can accept the vendor's definition of a resolved outcome.
This is why Intercom Fin, Zendesk AI Agents, Salesforce Agentforce, Sierra, Ada, Decagon, and similar products keep winning early deployments. They give buyers a way to start without staffing an internal AI platform team.
Zendesk's 2026 packaging shift, described in its AI agent packaging announcement, also shows where the market is going: AI agent capability is moving closer to the core support platform. That is good for standard service operations. It is less good when the workflow is cross-system, regulated, or strategically important enough that you should own it.
When Custom Pricing Wins
Custom pricing starts to win when the workflow has high volume or high complexity.
Common signals:
- You handle thousands of interactions per month.
- The agent needs to read from and write to several systems.
- Your policies are specific to your business.
- You need complete audit logs.
- The agent must support regulated workflows.
- Human handoff needs context, not just escalation.
- The system should become an owned asset.
At that point, a managed custom build can be more rational than a metered platform. You pay for the engineering and operating model up front, then the marginal cost of each additional interaction is mostly model, hosting, and monitoring cost.
That does not mean custom is always cheaper. It means the cost curve is different. With a platform, better automation can mean more billable resolutions. With an owned agent, better automation usually lowers effective cost per completed workflow.
A Practical Buying Rule
Use this simple threshold:
| Situation | Better first move |
|---|---|
| Under 2,000 simple monthly interactions | Buy a platform |
| 2,000 to 10,000 interactions with moderate workflow complexity | Compare platform TCO against managed custom build |
| 10,000+ interactions or regulated workflows | Seriously model custom ownership |
| Cross-system workflow with write actions | Start with consulting and architecture |
| Unclear process, no owner, no success metric | Wait and map the workflow first |
The middle band is where most buyers make the wrong decision. They see a low unit price, ignore the workflow cost, then discover that the agent cannot handle the exceptions that make the work expensive in the first place.
Where OpenNash Fits
OpenNash is not trying to be the cheapest chatbot. The fit is companies that need a production AI agent that works across their actual business systems and keeps improving after launch.
The OpenNash model usually makes sense when buyers want:
- Flat managed implementation instead of pure usage billing
- Custom tools and integrations
- Inspectable prompts, traces, and procedures
- Evals before and after launch
- Human review paths
- Single-tenant ownership of code, data, and workflow logic
- A partner who stays through operation
Start with the AI agent pricing service page if you need a cost model for a real workflow. Use AI agent consulting if you are still deciding what to build, or custom AI agent development if you already know the workflow requires an owned system. Read the Agentforce vs Sierra vs Decagon comparison if you are already comparing vendors. Bring one workflow, current monthly volume, current human cost, and the systems involved. That is enough to build the first useful cost model.