Questions to ask an AI support vendor, with Fini's answers

The questions that separate AI support vendors are about outcomes and control: how a resolution is defined and measured, what stops the AI from saying or doing the wrong thing, and what happens when it can’t help. Fini (usefini.com) answers each one below, with a link to the product docs that show how it works, so you can hold every vendor in your evaluation, Fini included, to the same standard. Use the questions as a script for demos and reference calls. For each one, ask the vendor to show you the answer in the product, not on a slide.

​

Outcomes and measurement

​

1. How do you define a resolution?

Why ask: “Resolved”, “automated”, “contained”, and “deflected” mean different things from vendor to vendor. A loose definition inflates every other number in the evaluation. Fini’s answer: A resolution counts only when Fini fully solves the issue, no human ever touches the conversation, and the customer does not reply for 72 hours. A conversation that reopens within 72 hours is not a resolution, and neither is anything a teammate replies to or touches, including flows where Fini leaves an internal note and a human sends the reply. In Analytics, every conversation has one status: Resolved by AI, Escalated to Human Team, or Waiting for Customer. Resolved by AI can appear before the 72-hour window closes, so the billed count can differ from Analytics in the short term. What counts as a resolution

​

2. Do you report resolution rate or deflection rate?

Why ask: Deflection usually counts any conversation that didn’t reach a human, including customers who gave up. Fini’s answer: Both are in the product, and they are defined openly. AI resolution rate counts only conversations with status Resolved by AI. Deflection rate is 100% - Human escalation rate, so conversations still waiting on the customer count as deflected. Fini’s position is that resolution rate is the metric that matters, and it is the one Fini bills on. Analytics, Resolution vs deflection

​

3. How do you measure accuracy, and what are your benchmarks?

Why ask: A high resolution rate means little if some of those answers are wrong. Fini’s answer: Fini’s stated benchmark across deployments is a 90% resolution rate at 99% accuracy, measured through human review of conversations. It is a benchmark, not a guarantee for every account: your numbers depend on your knowledge, rules, and traffic. In the product, you measure quality with Test Suite (conversation-based test cases that replay real customer messages and evaluate the new replies against your criteria), CSAT and sentiment in Analytics, and per-reply feedback in Inbox. How accuracy and resolution rate are measured

​

4. Can we see the reasoning behind any single conversation?

Why ask: When a regulator, auditor, or customer asks “why did the AI do that?”, you need an answer per conversation. Fini’s answer: Yes. Every conversation has an AI Steps trace: the user attributes fetched (with success or failure), every Intent Rule step that ran with pass or fail status, the generated answer, guardrail verdicts, and the reasoning behind its output tags. AI Steps trace, Inbox

​

Control and safety

​

5. What stops the AI from saying something it shouldn’t?

Fini’s answer: Guardrails check every generated reply before delivery: Internal reasoning leak, Banned terms, Confidential attributes, URL allowlist, AI disclosure, and up to 10 Custom rule checks per agent. When a check fails, Fini attempts one rewrite and checks again; if it is still unsafe, Fini sends a handoff message and escalates instead. Fini documents that guardrails are not a fail-closed security boundary, and shows each individual verdict in AI Steps so you can audit them.

​

6. How do you handle regulated workflows like refunds, disputes, or account changes?

Why ask: Free-form generation is the wrong tool for steps that move money or change accounts. Fini’s answer: Those workflows run as deterministic Intent Rules, behavior trees built from Check, Read, Form, Tool, and Reply nodes. The planner selects a rule from its description; once selected, the tree executes deterministically: given the same conversation context, it produces the same tree walk, the same Tool calls, and the same Reply instructions every time. Cancellation walkthrough

​

7. When and how does the AI escalate to a human?

Fini’s answer: The agent escalates when it lacks knowledge, an action fails, a policy or guardrail requires it, or the customer asks for a person. Inside a workflow, an Intent Rule branch hands off with a Reply that tells the customer a teammate will follow up. Articles can be marked for escalation with the article-level Escalation field, the Planning Prompt’s Escalation Topics route always-escalate categories to people, and for widget conversations a Business Rule with the On Escalation trigger creates the ticket in your helpdesk. Every escalation is tagged with a reason from a fixed taxonomy (Knowledge, Action, User, and Policy families), and your team receives the full conversation and context. Escalated conversations are never billed. Escalation and handoff, Escalation reasons

​

Capability

​

8. Can the agent take actions in our systems, or only answer questions?

Fini’s answer: It takes actions. Actions call your APIs (refunds, cancellations, address changes, card replacement), and User Attributes pull live customer context such as plan, status, or billing date into the conversation. Every call is recorded in the AI Steps trace. Card replacement walkthrough

​

9. Who keeps the knowledge base up to date?

Why ask: Most AI agents decay as policies change and nobody updates the content. Fini’s answer: Fini drafts knowledge from your sources and conversations and surfaces gaps and conflicts in the Review Queue. Background AI drafts always go to Review, so a person approves them before they reach customers. Magic Articles turn transcripts and notes into structured articles, flag duplicates, and land in Review or Published depending on your Status After Generation setting. Knowledge overview

​

10. Which channels and helpdesks do you support?

Fini’s answer: Chat (the Widget), Email, Voice, your Help Center, Slack, and the API, plus native deployment inside Zendesk, Intercom, Front, Gorgias, HubSpot, Salesforce, LiveChat, and Microsoft 365. One agent applies the same policies and the same audit trail across channels. Channel overview

​

11. Which languages do you support?

Fini’s answer: 130+ languages. Supported languages

​

Deployment

​

12. How long until we’re live, and how much engineering does it take?

Fini’s answer: The knowledge agent is live on day 1, agentic workflows by day 14, and full autonomy by day 30: live in 14 days, fully autonomous in 30. Rollout timeline, Quickstart

​

13. How do we test changes before customers see them?

Fini’s answer: Turn real conversations into Test Suite test cases with criteria for what should pass, and run them against draft Behavior, rules or articles before you publish: running with drafts does not publish them. Test Suite uses recorded or mock Action responses, so check live end-to-end Actions in a sandbox environment or a controlled pilot. Use Refine with AI to turn a bad reply into a proposed prompt, knowledge, or rule fix. After release, the Analytics pre-release validation workflow tells you whether to keep or roll back. To roll back a prompt, open History in Prompts, restore the last known-good published version as a draft, review it, then publish. Best practices for testing your agent

​

Commercials

​

14. What is your pricing model, and what do we pay for when the AI fails?

Fini’s answer: Fini pricing starts at $0.49 per resolution, a usage-based rate with no per-seat fees. See the pricing page for current plans and rates. Voice rates are listed on the pricing page. When the agent can’t solve an issue and escalates, you pay nothing for that conversation. Greetings, spam, and abandoned sessions are never billable. How Fini pricing works

​

15. Do you guarantee results?

Fini’s answer: Yes, through the 90-Day Money-Back Guarantee. Eligible new enterprise customers (1M+ annual tickets, at least 100,000 AI-handled interactions in the first 90 days) owe nothing for the first 90 days if automated resolution is below 90% of eligible queries, AI CSAT does not beat human CSAT, or average first response exceeds 30 seconds. Eligibility applies; see the guarantee terms (version dated 27/08/2025). 90-Day Money-Back Guarantee explained

​

Security and data

​

16. Which certifications do you hold?

Fini’s answer: Fini is SOC 2 Type II compliant, ISO/IEC 27001:2022 certified, PCI DSS Level 1 certified, HIPAA-compliant and BAA-eligible, with GDPR and CCPA compliance. Reports are shared through the Trust Center, and the PCI DSS attestation (AOC) through your Fini account team. Security questionnaire answers

​

17. Is our data used to train your models? Where is it stored?

Fini’s answer: Fini does not use customer data to train foundation models. Any per-customer learning happens only inside that customer’s environment and only with their written authorization. Data is encrypted with AES-256 at rest and TLS 1.2+ in transit, and stored in the US or the EU, as you designate. With EU data residency, data is stored and processed in the EU. An Azure-hosted deployment in your own tenant is available through Microsoft Marketplace. Data handling, residency and model training

​

18. Can we control who on our team has access?

Fini’s answer: Yes. Your team signs in through single sign-on with Okta, Google, Slack or Microsoft Entra ID, so access follows your identity provider, and the Trust Center lists role-based access control. API keys are scoped by permission. Okta SSO, API keys

​

Related