Buyer Guide

Enterprise AI chatbot platforms: the requirements checklist

Enterprise chatbot platform checklist: SSO, PII redaction, audit logs, handoff, data residency, languages, analytics and pricing, plus what to ask.

Author
By the Open Team
|Updated November 18, 2026|10 min read

An enterprise chatbot platform has to pass the same review as any system that touches customer data: identity, privacy, auditability, residency and cost control. It also has to resolve customer issues well enough that the business case holds. This checklist covers the eight requirements that decide most enterprise evaluations, why each one matters, and the exact questions to put to vendors.

Use it as the skeleton of your RFP. Copy the "what to ask" lines straight into the questionnaire.

TL;DR

  • Security teams block deals on four things: SSO, PII handling, audit logs and where the data lives. Get answers in writing before the demo stage.
  • Ask how the bot decides to hand off to a human, and what the agent receives. Confidence thresholds and full context transfer are the difference between help and a dead end.
  • "Supports 100 languages" means little until you test your own customers' dialects and mixed-language messages.
  • Analytics should explain individual decisions (sources, confidence, reasoning) as well as report resolution rates.
  • Pricing model shapes behaviour. Per-resolution pricing ties cost to outcomes; check what counts as a resolution and whether handoffs are billed.

The checklist at a glance

#RequirementOwner on your sideDeal-breaker if missing?
1Single sign-on (SSO) and role-based accessIT / securityUsually yes
2PII redactionSecurity / DPOYes in finance, health, telecom
3Audit logsSecurity / complianceUsually yes
4Human handoffSupport opsYes
5Data residencyLegal / DPOYes in the EU and much of the GCC
6LanguagesSupport ops / regional leadsDepends on markets
7Analytics and observabilityCX leadershipYes for scale-up
8Pricing modelFinance / procurementShapes the business case

1. SSO and role-based access

Why it matters. Your support platform will have dozens or hundreds of users: agents, team leads, content editors, BPO staff. Without SSO, every one of them is a separate password and a separate offboarding step. Role-based access decides who can change the bot's knowledge, who can read transcripts, and who can export data.

What to ask:

  • Which identity providers do you support (Okta, Microsoft Entra ID, Google), and over which protocol (SAML 2.0, OIDC)?
  • Is SSO included in the standard price or an enterprise add-on?
  • Do you support SCIM for automatic provisioning and deprovisioning?
  • Which roles exist out of the box, and can we define custom roles that limit transcript access?

2. PII redaction

Why it matters. Customers paste card numbers, ID numbers, addresses and health details into chat and email. That data then flows into transcripts, analytics, model prompts and any tool connected to the platform. Under GDPR, data minimisation is a principle (Article 5), so carrying personal data into systems that do not need it is a risk you have to justify.

What to ask:

  • What types of personal data do you detect out of the box, and can we add custom patterns (for example a national ID format or an internal account number)?
  • Is redaction applied before data reaches the language model, before storage, or both?
  • Can authorised agents still see the original where they need it, and is that access logged?
  • Is redaction applied to voice transcripts as well as text?

What Open publishes. Open's public pages describe optional PII redaction that detects and redacts or masks personal data. A ready-made PII redaction workflow covers tickets and call transcripts in Zendesk, Intercom, Freshdesk, Salesforce, HubSpot and Twilio Flex.

3. Audit logs

Why it matters. When something goes wrong (a wrong answer, a changed setting, a data request from a regulator) you need to show who did what and when. For an AI agent there are two layers: admin actions (sign-ins, settings, role changes, knowledge edits) and AI decisions (what the bot answered, from which source, and why).

What to ask:

  • Which events are logged: sign-ins, setting changes, role changes, knowledge base edits, data exports?
  • How long are logs retained, and can we export them to our SIEM?
  • For every AI response, can we see the sources used and the confidence score?
  • Can we replay a full conversation as the AI saw it?

What Open publishes. Open's healthcare page states that audit logs record sign-ins and changes to settings and roles. The Agent 5 page describes full reasoning traces, confidence scores and source citations for every response, plus conversation replays.

4. Human handoff

Why it matters. Every AI agent will meet questions it should not answer. What happens next decides whether customers trust it. A bad handoff loses the context and makes the customer repeat themselves. A good one passes the full conversation, a summary and the fields the bot already collected into the queue your team works from.

What to ask:

  • What triggers a handoff: customer request, low confidence, topic rules, sentiment?
  • Can the customer reach a human at any time by asking?
  • Does the agent receive the full transcript and a summary inside our existing helpdesk?
  • Are handed-off conversations billed?
  • What happens outside business hours: ticket creation, callback, or an honest "we'll reply by email"?

What Open publishes. Open routes uncertain answers to a human using confidence thresholds, and hands off with the conversation context. Open runs inside your existing helpdesk, so the handoff lands in the same queue. Handoffs to a human are not billed, per the pricing page.

5. Data residency

Why it matters. Where customer data is stored and processed affects which laws apply and how much legal work a deployment needs. GDPR does not require data to stay in the EU, but Chapter V restricts transfers to countries without an adequacy decision unless safeguards such as standard contractual clauses are in place. Saudi Arabia's Personal Data Protection Law also sets conditions on transferring personal data outside the Kingdom; our guide to the Saudi PDPL and AI customer service covers those rules and the draft in-Kingdom storage requirement. For a chatbot, the question covers the vendor's own storage and every sub-processor, including the language model provider.

What to ask:

  • In which region is our data stored at rest, and in which regions is it processed?
  • Which sub-processors handle conversation data, including model providers, and where are they?
  • Can we choose a region, and is that a contract term or a configuration?
  • How are transfers outside the region covered in your DPA?
  • Is customer data used to train your models or any third party's models?

What Open publishes. Open's financial services page states that customer data is hosted in the EU (AWS Frankfurt), that transfers outside the EEA follow Open's DPA, and that customer data is never used to train Open's models. Its GDPR page adds that some sub-processors and third-party AI models process data outside the EU under Standard Contractual Clauses. Open's security documentation is at trust.open.cx.

6. Languages

Why it matters. A European or GCC support team rarely works in one language. The bot has to understand Gulf Arabic and Modern Standard Arabic, Dutch and Flemish, and the customer who switches to English halfway through a message. It also has to answer in the right register, which is a harder test than understanding.

What to ask:

  • Which languages do you support for understanding, and which for generating replies?
  • Do we need separate knowledge bases per language, or does one source answer in every language?
  • How do you handle dialects and mixed-language messages?
  • Can we review and correct tone per language?

Test it yourself. Pull 50 real conversations per language, including the messy ones, and run them through each shortlisted platform. Score accuracy and tone separately. For Arabic, our Arabic AI chatbot guide has a per-dialect test plan.

What Open publishes. Open's healthcare page states that Agent 5 supports 100+ languages, with automatic language detection.

7. Analytics and observability

Why it matters. "Deflection rate" alone hides failures. A bot can deflect a customer by giving up on them. Enterprise teams need to see resolution rate, and they need to inspect why the bot did what it did, so they can fix the knowledge gap behind a bad answer. Wrong answers are the costliest failure; our list of AI hallucination examples shows what they look like in support.

What to ask:

  • How do you define a resolution, and is that definition in the contract?
  • Can we see per-intent and per-channel resolution rates?
  • Do reports stay in our helpdesk's native reporting, or only in your dashboard?
  • Does the platform flag knowledge gaps (questions it could not answer) automatically?
  • Can we evaluate changes against past conversations before they go live?

What Open publishes. The Agent 5 page lists real-time metrics for resolution rate, accuracy, response time and customer satisfaction, plus conversation replays and error detection. Because Open runs inside your helpdesk, native helpdesk reporting stays intact.

8. Pricing model

Why it matters. Pricing model decides who carries the risk. Per-seat pricing charges whether the bot works or not. Per-message or per-session pricing charges for activity. Per-resolution pricing charges only when the bot closes the issue, which aligns cost with value but makes the definition of "resolution" the most important line in the contract.

ModelYou pay forWatch out for
Per seatAgent licencesPaying for AI that does not resolve anything
Per session or messageConversations startedCost grows with volume even when the bot fails
Per resolution or outcomeConversations the AI closesA loose resolution definition; minimums; whether handoffs count

What to ask:

  • What exactly counts as a billable resolution, and over what time window?
  • Are handoffs to a human billed?
  • Are there platform fees, seat fees, minimums or overage rates on top?
  • How are telephony and messaging carrier costs passed through?

Market examples (checked 18 November 2026). Intercom prices Fin at $0.99 per outcome, with a 50-outcome monthly minimum when used with a non-Intercom helpdesk. Zendesk bills its AI agents per automated resolution, which a large language model verifies, and does not state a per-resolution price in its help article on the topic (see our Zendesk AI pricing breakdown). Freshdesk's Freddy AI Agent is billed per session, at $49 per 100 sessions after the first 500. For more vendors side by side, see AI agent pricing compared, and to turn a quote into a business case, AI customer service ROI.

What Open publishes. Open charges $0.75–$0.90 per resolution depending on monthly volume, with no seat fees. Handoffs to a human are not billed, and carrier and telephony costs are passed through at cost. See pricing.

How to run the evaluation

  1. Send the security questions first. SSO, PII, audit logs and residency answers come back in writing before you spend time on demos. Ask each vendor for their security documentation and DPA.
  2. Test on your own conversations. Give each shortlisted vendor the same set of real, anonymised conversations across your top intents and languages. Compare resolution and accuracy, not demo polish.
  3. Start with one narrow workflow. Email is often the easiest place to begin, for example a workflow to automate customer email replies with AI inside your existing helpdesk, before you move to chat, WhatsApp and voice.
  4. Run a paid or guaranteed pilot on live traffic. Route 10% of one channel to the bot for two to four weeks. Measure resolution rate, CSAT and handoff quality against the human baseline.
  5. Read the contract for the resolution definition. It decides your invoice.

When a different platform is the better fit

If you run Intercom end to end and only need chat deflection, Fin keeps you on one vendor. If your use case is internal (employee IT or HR questions rather than customer support), an employee-service platform may suit you better. If you are still choosing the helpdesk itself, settle that first; our comparison of Help Scout alternatives is one place to start. Open is built for customer support teams that need one agent across chat, email, WhatsApp and voice, inside the helpdesk they already use.

Sources

Frequently Asked Questions

An enterprise chatbot platform is software for running AI agents that talk to customers or employees at scale, with the controls a large organisation needs: SSO, access roles, audit logs, data residency options and integrations with existing systems. For customer support, it also needs reliable handoff to human agents and reporting on resolution rates.

Include security and privacy questions (SSO, PII redaction, audit logs, data residency, model training on your data), operational questions (handoff, languages, integrations with your helpdesk), analytics requirements, and the pricing model with a written definition of a billable resolution. Ask every vendor to test on the same set of your real conversations.

No. GDPR allows transfers outside the EU when the destination has an adequacy decision or when safeguards such as standard contractual clauses are in place. Many enterprises still prefer EU hosting because it reduces the legal work, so ask where data is stored, where it is processed and which sub-processors are involved.

Convert every quote into cost per resolved conversation at your expected volume and resolution rate. Per-seat and per-session prices look cheap until you divide by the conversations the bot closes. Then check minimums, overage rates and whether handoffs are billed.