What Sierra AI agents do in 2026
Sierra is the conversational AI company founded by Bret Taylor and Clay Bavor. It sells one thing: a branded AI agent that talks to your customers over chat, voice, email, SMS and messaging apps, and resolves the request instead of deflecting it. Its customers are consumer brands with high contact volume, the kind of companies whose support queue is mostly order changes, returns, bookings and billing questions.
The important word is resolves. A Sierra agent connects to your order system, CRM and billing stack and takes the action itself. Sierra's own Agent SDK page gives "updating a subscription or submitting a warranty" as examples, and its AWS Marketplace listing defines an outcome as a result such as "processing a return, resolving a case, or completing a transaction". Every one of those touches money. A return triggers a refund. A cancellation reverses a charge. A goodwill gesture is a credit with a dollar value.
Sierra builds these agents with you. The Agent SDK composes an agent from skills such as triage, respond and confirm, connects it to internal systems, and adds "clear objectives and guardrails" to keep it within policy and brand tone. There is also a no-code Agent Studio for operations teams. What none of the public material describes is a dollar ceiling, which is the question a finance team asks first.
Sierra AI pricing is outcome-based and negotiated
Sierra publishes no price list. On 12 October 2026 sierra.ai/pricing returned HTTP 404, the same response as a randomly generated path on the same site. Pricing is set per contract and tied to outcomes: you pay when the agent resolves the conversation, and nothing when it hands the customer to a person.
The one public price document is Sierra's AWS Marketplace listing. It shows a Sierra Platform dimension at 1,000,000 USD for a 12-month contract and a Sierra Outcomes dimension at 0.01 USD per unit for outcomes beyond the contract, both described as set through a private offer, and it states that "all fees are non-cancellable and non-refundable except as required by law". Treat those as listing placeholders rather than a quote. We break down the numbers and the third-party estimates in our Sierra AI pricing guide.
The pricing model matters for spend control in one specific way. Sierra charges you for a resolved return, and the refund that resolved it is a separate cost that never appears on Sierra's invoice. The meter counts outcomes. It has no idea whether an outcome cost you 12 dollars or 1,200.
We read Sierra's tau2-bench to see how agent policy is enforced
Sierra's production SDK documentation is not public. developers.sierra.ai redirects every path to a login page, and docs.sierra.ai returns the same 47,006-byte page for its root and for a random control path, so it is a catch-all, not a documentation set. There is no Sierra SDK on npm or PyPI.
Sierra does publish one body of agent code: tau2-bench, the benchmark from Sierra's research team that the industry uses to test customer service agents. It is MIT licensed on GitHub. On 12 October 2026 we pulled commit 4ce7c03 (7 October 2026) and parsed the tool files with Python's ast module, executing nothing. Each domain pairs a policy document the model reads with a set of tools, and each tool is tagged READ or WRITE in code.
Across the three customer service domains, airline, retail and telecom, there are 43 tools, 19 of them WRITE, and 12 of those move money: they charge a card, reverse a payment, refund a price difference or issue a credit. The policies set three hard numeric limits: a 100 dollar per passenger certificate for a cancelled flight, a 50 dollar per passenger certificate for a delay, and a 2 GB maximum on a data refuel. None of the three is checked in the tool code. send_certificate accepts any integer amount, and its docstring says only "Be careful!".
The airline policy says it outright: "The API does not check that cancellation rules are met, so the agent must make sure the rules apply before calling the API!" The cancellation tool reverses every payment on the reservation without reading the 24-hour, insurance or business-class rules at all.
Why a policy the model reads is not a spending limit
This is a benchmark, not Sierra's production code, and we are not claiming Sierra ships agents without controls. The finding is about the architecture Sierra's own researchers chose to model a customer service agent, which is the same one every LLM agent platform uses: the rules are text in the context window, and the tools execute whatever arguments the model passes. Compliance is a property of the model's behavior, and behavior varies.
Sierra measured that variance itself. The original tau-bench paper (Yao et al., 2024) introduced the pass^k metric because a single success rate hides inconsistency, and reported that gpt-4o succeeded on fewer than 50% of tasks, with pass^8 below 25% in the retail domain. An agent that follows the refund policy on seven of eight attempts has still broken it once, and at contact center volume once is a daily number.
The tools in the benchmark are not careless either. Across the 19 write tools we counted 40 raise statements. They check that an order exists, that it is in the right status, that a gift card has enough balance, that a payment method belongs to the user. That is the control result: the code validates state carefully and leaves money policy to the model. A spending limit belongs in that first category, enforced by code the agent cannot talk its way past.
What a finance team adds before a Sierra agent issues refunds
Before an agent gets a tool that refunds, credits or cancels, a controller will ask for controls that sit outside the model and that a persuasive customer cannot rewrite:
- A per-transaction ceiling on refunds and goodwill credits, by reason code, so a 100 dollar certificate rule is a number the system rejects above, not a sentence the model remembers.
- A daily and monthly budget for each agent, cumulative across every channel it serves, so a bad prompt or a social-engineering pattern cannot drain the month in an afternoon.
- An approval threshold that routes refunds above a set amount to a supervisor in Slack or by email before money moves, with the customer told a person is confirming.
- Counterparty and velocity rules, such as one goodwill credit per customer per 30 days, which a model reading a fresh conversation cannot see.
- An audit trail that records the agent, the conversation, the amount and the rule that allowed it, ready for chargeback disputes and quarter-end review.
These are ordinary finance controls. The difference with an AI agent is that they have to be enforced at the moment of the API call, because there is no human in the loop to apply judgment afterward.
How AgentsPay works alongside Sierra
Keep Sierra as the agent your customers talk to. AgentsPay sits where the money leaves. Each Sierra agent that can refund, credit or charge gets its own AgentsPay identity with a hard per-agent budget and per-transaction cap, reason-code and counterparty rules, and approval thresholds that send larger refunds to a person before they are issued. Every decision lands in an exportable audit trail that finance and the dispute team can read. The pattern is the same one we describe for AI agent refunds in general.
Wiring is one call. The action that issues the refund or credit asks AgentsPay first with the agent, amount, customer and reason, and proceeds only on an approval. The policy is the same for agents built on other platforms, which matters when a single brand runs a Sierra agent on the front line and back-office agents in Salesforce Agentforce, Zendesk or Intercom Fin.