Requirement 1: per-client isolation

Define the boundary for client records, agent configuration, knowledge and staff access. Ask how permissions apply to lists, searches, exports and direct record access. Separate the agency operator's privileges from the client user's privileges.

In distinct authorized test workspaces, create synthetic contacts with recognizable names. Sign in as one client's user, open its contact and then request the other client's record through the supported views and interfaces. Keep the actor, operation and response with the evidence.

Pass: the user's own contact opens and the unrelated contact is denied without revealing its name or details. Fail: the unrelated contact's name appears in a search result, even if opening it is blocked. Keep any untested access path open for review.

Requirement 2: white-label coverage

A client may need a branded workspace without requiring branded account emails. Write down the surfaces in the agreement: logo, favicon, colors, login, invitations, notifications and domains. This gives the evaluator a finite journey to inspect.

Apply the permitted branding in a test setup, then follow an invitation through login, ordinary use and password recovery. Save what the client sees on each required surface, including messages generated along the way.

Pass: an account-recovery email displays the agreed sender identity when that surface is required. Fail: it arrives with the platform's default identity. Record exceptions against the agreement so a surface the client never requested does not decide the comparison.

Requirement 3: knowledge freshness and authority

Start with the knowledge readiness guide: name the source owner, approval rule and refresh process. Distinguish static policy from prices or availability that need a live lookup. Specify who handles a conflict or unavailable source.

Use a synthetic policy saying enquiries close on Friday. Change it through the supported process to Thursday, wait for the stated refresh, then ask the same question. Inspect the source used for the answer and repeat with that source unavailable.

Pass: the answer changes to Thursday and the unavailable-source case follows the agreed fallback. Fail: the agent keeps offering Friday after the refresh. Record the source version and time alongside the answer so the policy owner can investigate.

Requirement 4: human handoff with usable context

Ask a question outside the agent's authority, correct a collected phone number and request a human. Then open the request through the receiving staff account. This tests the point where a conversation becomes someone's work.

Pass: the staff request contains the corrected number, unresolved question and attempted action status, with a receiving owner. Fail: staff see the original number or must reconstruct the unresolved decision from a transcript.

Use the answer boundary and handoff worksheet to define what the customer hears before staff accept the request. Include the unavailable path and the team's working window. Save the staff view with the test record, and agree who owns requests that remain unaccepted.

Requirement 5: workflow triggers and recovery

Identify the event, schedule, webhook or manual action the service needs, along with its payload, published version and recovery owner. Agree the duplicate rule before trying the workflow. Specify which steps, if any, may send during the authorized evaluation.

Deliver a synthetic booking event twice and interrupt the destination response after a write. Inspect saved progress and the authoritative booking record before deciding whether to replay. The webhook retry method helps prepare this uncertain-result case.

Pass: repeated delivery leaves one booking, and recovery identifies that booking before another write. Fail: two bookings appear. Preserve the trigger and destination identifiers with the observation; a simulated run should be marked as such in the worksheet.

Requirement 6: provider and channel choice

Which provider and channel combination does the client actually need? List telephony, speech and model responsibilities where relevant, number ownership, language review and dependency fallback. Ask what can change without rebuilding the service.

Run an authorized conversation through the intended provider path to a test destination. Include an interruption and a correction. In a separate controlled failure case, make a required dependency unavailable and inspect the conversation and saved request together.

Pass: during the outage the customer hears the agreed fallback and staff receive an unresolved request. Fail: the agent announces a confirmed booking while the destination remains unavailable. Record the provider configuration and test date; if evaluation uses another path, leave the intended one unrun.

Requirement 7: cost transparency

Use the service-cost worksheet to separate agency support work from platform access, metered usage and external provider charges. Identify the payer and unit for every component, including unsuccessful attempts and retries where they are billed.

Build ordinary, quiet and support-heavy workload scenarios from stated assumptions. Reconcile an authorized usage sample against the written charging definitions. Trace a retry charge back to its attempt and ask how allowance overruns or disputes would be handled.

Pass: that retry has an identifiable usage entry, applicable unit and named payer. Fail: it appears only in an unexplained aggregate charge. Where pricing is unpublished, keep the commercial requirement unresolved until a quote supplies the missing terms.

Requirement 8: usable export and offboarding

Plan the exit while ownership is easy to agree. List the records and configuration the receiving operator needs: contacts, activity, conversations, knowledge references and workflow definitions. Identify exclusions, provider-account owners, retention requirements and who can transfer or remove access.

Rehearse the handover with synthetic linked records and a corrected contact. Request the supported export and have the next operator inspect identifiers, relationships, timestamps and final values. Then rehearse a cutoff with a delayed job and an uncertain external result. Decide who resolves each pending action, stops schedules, changes credentials and ends continuing charges. Keep secret transfer separate from the ordinary record export.

Pass: the receiver can connect the corrected contact to its activity, and the delayed job is stopped or explicitly transferred to a named owner. Fail: the export loses that link or the job runs after the agreed cutoff.

The offboarding inventory and evidence checklist provides the account review. Save evidence of the final access and charge dispositions. If a needed category cannot transfer, agree a workable alternative before accepting the platform.

Requirement 9: API and MCP access

Specify the operation, client and workspace you need before evaluating API or MCP access. Identify its authentication boundary, supported inputs and observable result. Ask whether the intended account has execution access and whether the need is an agent action, external tool import or developer integration.

Inspect the supported interface documentation, then perform an authorized operation with synthetic inputs. Retain its returned identifier and read the result back through the relevant staff view. Include a rejected input in the evaluation.

Pass: a supported task-creation operation returns an identifier that opens the expected synthetic task. Fail: the response reports success but the task cannot be found. Where runtime access is unavailable, record the documentation reference and leave execution unrun.

Worksheet: compare evidence against the same requirements

Copy the table for each candidate and set priorities before testing. Record a pass, fail or unrun result with the observation, evidence reference and remaining question in the empty final column. A failed essential boundary should stay visible rather than disappear into a combined score.

Name the reviewer, configuration, provider, connection owner and test date. Adapt the pass conditions to the service brief, then use the same conditions for every candidate.

RequirementPass condition to adaptPriorityResult and evidence
Per-client isolationOwn record opens; unrelated name and details stay hiddenSet before testing
White-label coverageRequired recovery email uses the agreed identitySet before testing
Knowledge freshnessChanged policy reaches the answer; unavailable source triggers fallbackSet before testing
Human handoffStaff receive the corrected number and unresolved taskSet before testing
Workflow triggers and recoveryRepeated delivery and recovery leave one bookingSet before testing
Provider and channel choiceOutage produces the agreed fallback and unresolved requestSet before testing
Cost transparencyRetry charge traces to its attempt, unit and payerSet before testing
Usable export and offboardingLinked records remain usable; delayed jobs stop or transferSet before testing
API and MCP accessReturned identifier opens the expected saved taskSet before testing

Where Vora fits this evaluation

Vora is the operating system for AI agencies, in early access through the waitlist. It supports agent creation, duplication, versions and configuration diagnostics, including draft voice-agent creation from a website. Its native CRM includes contacts, companies, activities, pipelines, deals and tasks. Its native inbox brings together website chat, SMS and email.

Chat creation, workflows, lead generation, native email outreach, client-workspace creation and per-client branding controls are in preview. Branding controls cover logos, favicon and colors. Workflow preview includes trigger choices and execution history, while accepted work and simulated sends remain distinct from delivered results.

The Vora voice MCP developer reference describes registered voice tool wrappers and the checks required before integration. Use it to scope the voice-interface evaluation and identify the account access needed for your operation.

Vora pricing is not public yet. Leave the cost row open until you have the terms for the proposed configuration. Vora's tenant isolation, export and offboarding behavior isn't covered here; run the tests above on it like any other platform.

If an essential requirement remains unmet, retain the current tool or narrow the service. The agency software stack guide helps compare that choice with an integrated workspace. Bring the worksheet when evaluating Vora so the conversation starts with the job and its evidence.