Ada vs Sierra: Who approves support actions, and who owns the handoff?

Selecting tools for complex support SOPs requires separating approval of individual actions from approval of agent releases. Early-access eligibility for Ada Consult affects collaboration pilots, while Sierra’s release governance suits shared maintenance. Both still require validation of exception recovery and handoff ownership.

Source: docs.ada.cx
November 18, 2026
6 min

A customer requests a replacement order. After checking inventory, the AI discovers that a manager must approve it, but by the time approval arrives, the original item is out of stock. When comparing Ada and Sierra, the key questions are what that approval actually permits, who continues the work after a human takes over, and how completed actions can be remedied. Both connect to business systems. Selection should start with this chain of exceptions rather than the fluency of the answer.

If your team wants to break support SOPs into explicit steps and explore collaboration in which AI asks a colleague for approval before continuing, Ada merits a pilot, subject to access to Consult. If several teams maintain an agent together and especially need pre-release review, gradual rollouts, and version tracking, Sierra’s public governance mechanisms deserve priority in acceptance testing. This assessment reflects official information, not a hands-on reliability ranking.

Put approval before the specific action

Ada’s current Playbooks use structured steps: ASK gathers information, SET writes variables, RUN executes API tools or hands off to a human, and IF/ELSE handles branches. This lets teams examine order confirmation, eligibility checks, exception requests, and replacement execution separately. Coaching is not considered during Playbook execution. Internal behavior must be changed in the Playbook itself; adding a global instruction does not establish that the workflow has been fixed.

Generating steps does not mean the systems are connected. The getting-started documentation says newly generated API tools are disabled by default and use placeholder URLs. Actual endpoints, authentication, inputs, and outputs still need configuration before launch. Operations can review the process, but engineering remains responsible for the boundaries of actual writes.

Ada’s combination-announcement FAQ explicitly labels Consult as early access: AI can ask a colleague for approval or input and then finish the task itself. That fits this scenario, but procurement must confirm account eligibility, waiting timeouts, and what happens after rejection. The announcement also says existing products, agreements, and support remain unchanged and that using Dialpad alongside Ada is not required. Future integration cannot be treated as an already unified product and contract.

Sierra’s Agent Studio likewise supports generating journeys from SOPs, invoking integrations, and inspecting traces of decisions, tool calls, and responses. Both can orchestrate cross-system work. Ada should not be portrayed as chat-only, nor should Sierra’s visual interface be read as eliminating engineering maintenance.

Maintenance staffing belongs in the cost comparison for these workflows. Teams that change rules frequently need someone to assess whether existing tests still cover new exceptions. When an internal order system changes fields, someone must update the interface and error handling. This article does not calculate outcome-based charges without comparable contract scopes, because one fewer human reply does not prove there was one fewer round of rework.

Release approval and order approval are different

Sierra’s release-governance explanation describes clear launch controls: teams can require simulations to pass, appoint a Reviewer to examine changes, and release gradually to a subset of customer traffic. Reviewers can inspect line-by-line diffs, comment, and request changes. That is valuable when several departments jointly modify support rules.

These approvals apply to releasing an agent version, however, not to a manager approving every replacement. The public materials reviewed do not provide enough detail to establish how Sierra implements per-order approval and subsequent resumption in this scenario. Ask the vendor to demonstrate it rather than concluding that it is unsupported. Ada Playbook edits made through MCP also enter a change set before testing and promotion, so version control is not exclusive to Sierra.

Both demonstrations should answer whether a manager approves a particular item on a particular order or a broadly worded instruction. If inventory, the address, or compensation changes after approval, is approval requested again? These are recommended acceptance criteria, not verified default behaviors. The approval record must be traceable to the write that actually occurred.

Do not leave ownership unresolved after handoff

Ada’s flow configuration contains an easily missed detail: when a customer requests a human, the default is to exit the Playbook, clear progress, and never return. Enabling Stay in this Playbook instead acknowledges the request, continues the current flow, and follows up later. Selection demonstrations should check the actual configuration. “Supports handoff” alone does not mean the agent immediately stops.

Sierra’s Agent SDK says escalation routes the conversation to the appropriate team member and automatically generates a detailed summary. That helps, but acceptance testing must still establish whether the recipient sees verified information, completed actions, and outstanding work. If a replacement order already exists, a summary saying only “order issue” must not lead the receiving agent to create another.

Assign three areas of responsibility in advance: the business owner defines exceptions and approval scope; the integration owner checks actual system state and handles duplicate requests; and the receiving team owns unfinished work and customer notifications. Products can pass context, but the organization must still decide who follows up when nobody accepts the case, approval times out, or an API response is inconclusive.

The handoff demonstration should also include periods when no support staff are online. The customer-facing message, destination queue, follow-up channel, and owner’s acknowledgment should form one traceable record. If the system merely creates a ticket with no owner, automation may have ended the conversation, but the business problem has not been taken up. Count that as unfinished instead of concealing it with fluent wording.

Choose the pilot through exception scenarios, not feature counts

Give both vendors the same set of fictional orders: a normal replacement, manager rejection, approval timeout, stock loss after approval, a successful write with a lost response, and a customer requesting a human midway through. Check branches with simulations first, then verify writes and handoffs in an isolated test environment. This article has not run those tests for readers and does not recommend using real customer orders for initial validation.

Record “action not executed,” “success confirmed,” and “outcome unknown” separately. Require a state check for unknown outcomes before deciding to retry. Also check who cancels or corrects replacement orders already created after an older agent version is restored. Sierra’s immutable release snapshots support version rollback, but reverting agent configuration must not be treated as a guarantee of reversing external business transactions.

Before the pilot, define pass criteria as verifiable results: no write occurs without approval; rejection does not cause execution through another path; the same request does not create duplicate replacements; and the recipient can identify stale information. On failure, preserve inputs, approval scope, API results, and the person handling the case. This distinguishes rule errors, integration errors, and handoff omissions and identifies what to change next.

If the main problem is continuing work after waiting for a colleague’s approval, first confirm eligibility to pilot Ada Consult, then validate the connection between approval and execution. If shared rule changes are causing rework, prioritize testing Sierra’s review, simulations, and gradual release. If your team has not even agreed who takes exception cases, establish ownership before expanding either product’s authority to act automatically.

Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.

Original source:docs.ada.cx