AI chatbot vendor due-diligence checklist

30 questions to ask any AI chatbot vendor — before signing

Buyer-side procurement checklist for AI chatbot vendors. We answer every question for SLAtech transparently and tell you how to verify each claim against any vendor. Pairs with eval scoreboard, filterable vendor compare, and TCO calculator.

1. Compliance posture

If you serve regulated industries (Med, Legal, Edu, EU customers), compliance is a gating concern, not a nice-to-have.

  1. Are you HIPAA-compliant? Will you sign a BAA?
    Why it matters: Without BAA, your firm assumes liability for any PHI processed by the vendor.
    SLAtech answer: Yes — SLAtech Medical ships a BAA-eligible single-tenant option. BAA executed before any PHI ingest.
    How to verify: Ask the vendor to send you a sample BAA. If they say "contact sales" instead of sharing — that's a signal.
  2. Where is data stored? Can I require EU residency?
    Why it matters: GDPR + the 2025 Schrems-III decision tightened EU-to-US transfer requirements. Non-EU vendors carry transfer-risk.
    SLAtech answer: Hosting in your own jurisdiction is available on request (Enterprise deployments).
    How to verify: Look for explicit data-center geography on /trust/ or /security/. If the page only says "global cloud" without specifics — escalate.
  3. Do you have SOC 2 Type II or ISO 27001?
    Why it matters: Annual independent audit signals real security investment vs marketing claims.
    SLAtech answer: SOC 2 Type II audit in progress (closure Q3 2026). ISO 27001 timeline TBD based on EU enterprise traction.
    How to verify: Ask for the audit report (under NDA). "We're working on it" forever = red flag.
  4. Will you delete my data on contract end? Within how many days?
    Why it matters: Data-retention without explicit deletion clause leaves a forever-copy of your customer PII at the vendor.
    SLAtech answer: 30-day grace period after termination, then full deletion. Audit log of deletion provided.
    How to verify: Get this in writing in the MSA. "As required by applicable law" alone is not enough.

2. Eval / quality measurement

Any vendor can claim "powered by a leading LLM" or "94% accuracy". Real signal is a published, reproducible eval methodology.

  1. What eval methodology do you use? Is it published?
    Why it matters: Vendors who can't describe their eval process likely don't run one regularly.
    SLAtech answer: 200-question per-vertical sealed eval harness, LLM-as-Judge scoring on factuality/hallucination/confidence axes. Methodology open-sourced.
    How to verify: Ask vendor to share a sample eval report. If they refuse "trade secret" — they're hiding something.
  2. What's your bot containment rate (% of conversations resolved without human handoff)?
    Why it matters: Containment rate drives ROI — a 50% containment vs 80% containment is a 30-percentage-point cost-of-support difference.
    SLAtech answer: 65-80% baseline depending on vertical. Med 78%, Legal 73%, General-Business 72%. Tracked per-tenant in dashboard.
    How to verify: Run a 100-conversation trial against your own traffic. Compare vendor's claim vs reality.
  3. How often do you re-run evals after model updates?
    Why it matters: Vendors who don't re-eval after model swaps risk silent quality drift.
    SLAtech answer: Every model swap triggers full eval rerun. Eval-score history published in changelog (/en/changelog/).
    How to verify: Check vendor's changelog / release notes for eval-rerun mentions. Absence = signal.

3. Multilingual + locale depth

If you operate in Israel, EU, or anywhere outside US-English, locale depth is a hidden differentiator.

  1. Do you support Hebrew RTL with proper bidi rendering?
    Why it matters: Hebrew RTL done badly produces garbled forms, broken date pickers, mis-aligned buttons. Most vendors only auto-translate, not localise.
    SLAtech answer: Hebrew RTL first-class, bidi-aware UI primitives, date/time/currency format locale-aware. Tested on every release.
    How to verify: Spin up vendor's demo in Hebrew. Send a message; check if response renders correctly, if date format is dd/mm/yyyy not mm/dd.
  2. Do you support Russian with native-locale phone/date/currency formatting?
    Why it matters: Russian-speaking diaspora is significant in Israel, Germany, Cyprus, Estonia. English-only chatbots leak this traffic.
    SLAtech answer: Russian first-class language, phone format +7 / +972 / EU formats auto-detected, ruble currency available.
    How to verify: Set vendor's demo language to Russian, test phone-number entry. Auto-formatting? Translation from English to Russian buttons?

4. Data portability + vendor lock-in

Lock-in is a common vendor strategy. Look for explicit data-portability commitments.

  1. Can I export all my conversation logs in standard format?
    Why it matters: Without export, switching vendors means losing all conversation history + training-signal investment.
    SLAtech answer: JSON Lines export of every conversation, 24-hour SLA on export request. No "per-export fee".
    How to verify: Ask vendor to demo export of 100 sample conversations. If they require a "data services SOW" — lock-in red flag.
  2. Can I export my knowledge-base / FAQ / training data?
    Why it matters: If knowledge-base is locked to vendor's schema, you're paying for the schema-conversion every switch.
    SLAtech answer: Markdown export of knowledge-base, source-of-truth stored in Git repo on customer side if customer prefers.
    How to verify: Ask vendor to demo knowledge-base export to standard format (Markdown, JSON, CSV).
  3. Will you charge me for switching off (egress fees, exit fees)?
    Why it matters: Some vendors hide vendor lock-in via prohibitive egress / exit fees.
    SLAtech answer: Zero egress fee. Zero exit fee. Documented in pricing page (/en/pricing/) and in MSA.
    How to verify: Get "no egress fee" in writing in the MSA. "Contact sales for pricing" = signal.

5. Pricing transparency

Per-conversation pricing with hidden overage is the dominant lock-in vector.

  1. Is your pricing published on the website?
    Why it matters: "Contact sales for pricing" alone signals enterprise-lock-in motion. Buyers should be able to budget without a sales call.
    SLAtech answer: Full pricing matrix at /en/pricing/ with live API-driven numbers. No "contact sales" gate for Starter/Pro tiers.
    How to verify: Check vendor's pricing page. If you can't get a number without email-gate — that's how procurement gets stuck.
  2. What's the cost per conversation in overage?
    Why it matters: Overage rate is where surprise bills hide. Some vendors charge 10× the base rate.
    SLAtech answer: Overage rate published per plan tier. Hard cap option available — bot stops vs unlimited overage.
    How to verify: Run TCO calculator (/en/tco-calculator/) with vendor's overage rate. Compare to base subscription.

6. Implementation timeline + ongoing support

Most enterprise chatbot deployments take 6-12 weeks with a forced implementation-consultant SOW.

  1. How long is a typical implementation? Is the consultant SOW required?
    Why it matters: Forced consultant SOW often costs more than the first year of subscription.
    SLAtech answer: Self-serve setup in 30-45 minutes for Starter/Pro. Custom Solutions tier offers white-glove (optional, not required).
    How to verify: Sign up for vendor's free trial. If setup wizard exists end-to-end, that's a signal of no-consultant motion.
  2. Who do I email for support? What's the response SLA?
    Why it matters: Tier-1 support routing to a ticket queue with 72-hour first-response is a common SaaS pattern. Mid-market expects faster.
    SLAtech answer: Direct email <a href="mailto:info@slatech.ai">info@slatech.ai</a> for founder reachability. 24-hour first response. Premium tiers get Slack Connect.
    How to verify: Email the vendor a pre-sales question. How long until response? Is it a human reply or canned template?

7. Roadmap + company stability

AI vendor space is volatile. Buying a chatbot from a pre-funded startup that goes under in 18 months is a risk.

  1. Is the company profitable? When was the last funding round?
    Why it matters: Pre-revenue startups burning through $20M / quarter risk shutdown / acquihire vs delivering roadmap.
    SLAtech answer: Bootstrapped (no VC) — sustainable from subscription revenue since 2025. Founder Emil Slavin (see /en/founder/).
    How to verify: Check Crunchbase / PitchBook for funding history. Bootstrapped or profitable = different risk profile than $200M Series C.
  2. What's your public roadmap?
    Why it matters: Vendors who don't publish a roadmap leave you guessing whether a feature is coming next quarter or next year.
    SLAtech answer: Public roadmap at /en/roadmap/ — current quarter shipping, next quarter committed, future quarter directional.
    How to verify: Look for public roadmap page. "Contact sales" or "we don't publish roadmap" = signal.

Want SLAtech's answers in a PDF?

Send us the email @ info@slatech.ai — we'll send a filled-in version for your procurement team. No marketing automation, no drip campaign — just the document.