Why Agents Break the SaaS Contract
AI agent platform contracts are a genuinely new category, and treating them as ordinary software deals is the central buyer error. An agent does not just respond — it acts: it books, refunds, escalates, moves data and, increasingly, money. As agentic products shift from passive tools to autonomous actors, the contracting model is moving from SaaS toward something closer to outsourcing, with BPO-style clauses covering service definitions, outcome SLAs, indemnification, governance and audit. The familiar per-seat licence captures none of that. This is the frontier of the discipline we set out in the AI contract negotiation deep dive.
The Pricing Models
The market has not settled on one model — leading platforms run several at once. Salesforce Agentforce sells add-ons from $125 per user per month and an Agentforce 1 edition at $550 per user per month bundling 1 million Flex Credits a year; having dropped a flat $2 per conversation in May 2025, it now meters Flex Credits at about $0.10 per action ($500 per 100,000). Microsoft Copilot Studio sells credit packs of 25,000 for $200 per month plus seats at $21–$30 per user per month. And a fast-growing outcome-based camp — Sierra, Zendesk — charges only on success: Zendesk from $1.50 per automated resolution, with Sierra's pure outcome model helping it cross $150M+ ARR by early 2026.
| Platform | Model | Indicative price |
|---|---|---|
| Salesforce Agentforce | Seat + consumption | $125–$550/user/mo · ~$0.10/action |
| Microsoft Copilot Studio | Seat + credits | $21–$30/user/mo · $200/25,000 credits |
| Zendesk AI agents | Outcome-based | $1.50 per automated resolution |
| Sierra | Pure outcome-based | Per successful resolution |
The unit you pay for — seat, action, credit or resolution — determines whether the bill is predictable, and consumption models make high-volume workflows hard to forecast. It is the same billable-unit discipline that decides outcomes in conversational AI platform licensing and AI vendor benchmarking.
What Deployment Actually Costs
The per-seat headline understates the programme. A full Agentforce deployment commonly costs around $900,000–$1.5 million in year one, and a comparable Copilot Studio programme lands within roughly 15–20% of that, depending on Copilot licence alignment and Microsoft 365 tier. Because credit and per-action pricing scales with usage, the monthly total is unpredictable for high-volume workflows — so the year-one figure must be modelled against realistic action volumes, not a nominal seat count.
Model the action volume, not the seat count. A consumption-priced agent that looks cheap per action at pilot scale can become the largest line in the AI budget once it runs across the business.
The build-versus-buy maths here mirrors the lifecycle analysis in AI fine-tuning costs and contracts, and the portability concerns in multi-model AI strategy apply with equal force — an agent platform that locks your workflows in is expensive to leave.
The Liability Gap
The defining risk of an agent contract is the gap between what the agent can do and what the vendor will stand behind. Under most agentic agreements the customer bears the risk of the agent's actions: suppliers provide the software "as is", disclaim fitness for purpose, and cap liability at the fees paid over a defined period. That framework is hard to defend if an autonomous system moves money incorrectly or breaches a compliance rule — the loss can dwarf the annual fee that caps recovery.
The regulatory backdrop is shifting too. The EU Product Liability Directive, due for member-state implementation by 9 December 2026, explicitly treats software and AI as "products" and allows strict liability for a defective system. This is general information rather than legal advice — and it makes specialist legal review essential — but the commercial implication is clear: the default liability allocation in an agent contract is the most important term to renegotiate, and the one vendors most want left untouched.
Governance: Supervision and Audit
Because an agent acts, the contract has to govern how it acts — and standard SaaS terms are silent on this. The control that matters most is human-in-the-loop: a contractual requirement that high-stakes actions (moving money above a threshold, deleting records, issuing customer commitments) cannot execute without human approval. The right threshold is a commercial decision, but it must be written into the agreement and configurable by you, not the vendor.
Three further governance terms belong in any serious agent contract. Audit rights — a complete, tamper-evident log of every action the agent takes, retained and accessible for investigation. Explainability — the ability to reconstruct why the agent took a given action, which both regulators and your own risk function will demand. And supervision and kill-switch provisions — the right to suspend or constrain an agent's autonomy immediately if it misbehaves, without penalty. These are the operational equivalent of the indemnity and liability terms: they reduce the probability of the loss the liability cap would otherwise have to absorb, and they make an autonomous deployment defensible to the board.
From SaaS to Service: What to Negotiate
Contract for an agent the way you would contract for an outsourced function. Add service definitions and outcome-based SLAs that specify what the agent must achieve, not merely that the platform is available. Negotiate broader indemnification and a higher liability cap proportionate to the actions the agent can take — a refund agent that can move money should not be capped at a month's fees. Insist on governance and audit rights, explainability, and human-in-the-loop and supervision requirements so a high-stakes action cannot execute unchecked.
On the commercial side, pin the billable unit — credit, action or resolution — in writing, with a cap on consumption and a clear definition of what counts, then use term and volume as discount leverage. Buyers paying substantial sums, or able to draw on a wider supplier relationship, can realistically win these AI-specific terms; off-the-shelf, low-spend deployments will have less room. For the full clause set, work through the AI Procurement Checklist and the AI Contract Red Flags brief, benchmark the platforms via the Salesforce and Microsoft hubs, and request a confidential briefing before you delegate real authority to an agent.