Documentation

AI assistant

Improving answer quality

A repeatable workflow to diagnose bad or missing AI answers, the quality tools in the dashboard, common failure modes with fixes, and a checklist to run before going live.

Last updated: 2026-09-29

Almost every bad answer comes from one of a handful of causes, and each has a specific fix. Follow this page in order: find what the AI was allowed to know, fix the source, then test again.

Diagnosis workflow#

  1. Reproduce. Open the conversation in the inbox and note the customer's exact words and channel. Send the same message from the web widget or a test number.
  2. Was the AI even allowed to answer? If there was no reply at all, run the silence checks below. See How the AI works for the gate.
  3. Find the source of truth. Ask: where should this fact live? Text policy or contact: knowledge base. Price, stock or menu: catalog. Branch address and hours or an offer: business records. Delivery fee: delivery settings. Tone or format: assistant identity.
  4. Check that the source is correct and ready. Is the file ready? Is the fact stated plainly, once, without a contradicting copy?
  5. Fix the source, not the prompt. Add or correct the fact. Only then adjust rules.
  6. Retest with the original wording and two paraphrases, including a dialect version.

If the AI did not answer#

CheckWhere
Disable AI replies onSettings → Inbox settings → AI Assist
No AI Assist add-on or subscription inactiveSettings → Plan & usage
Credits and wallet emptySettings → Plan & usage. See Credits and usage
No brand in the assistant identityAI identity: save with a brand. See Assistant identity and scope
Conversation Escalated, assigned to a person, or assistant pausedInbox conversation header
WhatsApp AI automation offIntegrations, WhatsApp card
Customer sent only emoji, or an automatic reply loopBy design, the AI stops after a few

Quality tools in the dashboard#

  • Inbox. Read the customer message, the AI reply and any handoff summary. Escalated conversations show why the AI stopped.
  • Automatic quality review. On the AI identity page, turn on Automatic quality review. Set Replies reviewed (%) (default 5) and a Daily cap (default 20). Each review uses credits.
  • AI quality report. In Reports, open AI quality. It shows the average score, the number of reviews under 60, issue types such as hallucination, and recent reviews with an Open conversation link. See Reports and analytics.
  • Learned from chats. Human-handled conversations become suggested knowledge for you to approve. See Knowledge base.
  • Copilot. Shows draft answers to your operators from the same knowledge, revealing gaps quickly.

Common failure modes and fixes#

What you seeLikely causeFix
"I could not verify..." replyThe AI's draft was not supported by tool resultsAdd the missing fact to the KB, catalog or records
AI says it does not knowFact is missing, file not ready, or wording far from the documentWrite the fact as text with the customer's words; wait for ready
Wrong priceOld price in a KB file, or product entry outdatedRemove prices from KB; correct the catalog
Says you are open or closed wronglyWorking hours or timezone wrong, or no published branchFix hours in Environment and publish branch records
Contradicting answersTwo versions of a policyDelete the old file
Ignores your ruleThe rule conflicts with a platform rule, or is vagueRewrite as one short, checkable instruction; remove conflicting demands
Wrong language or dialectNo language rule, or Arabic names not aliasedAdd a language rule and aliases for products and areas
Refuses a topicOff-topic for your businessAdd real knowledge for the topic, or accept the refusal
Escalates oftenRepeated unverifiable answers, or customers requesting peopleFill the fact gaps and review the handoff summaries
Cannot quote productsGeneral environment without a catalogUse a Store or Restaurant environment and add products
Order not created after "yes"Quote expired after 15 minutes, a newer message arrived, or the reply had extra wordsAsk the customer to confirm again. See Orders and payment links
Quiet after a handoffThe conversation is still Escalated or assignedReopen it and remove the assignment before handing back

Pre-launch checklist#

Work through this once before you turn the AI loose on customers.

  • [ ] AI Assist add-on is active and credits are available.
  • [ ] Assistant identity has a name and a brand, and is saved.
  • [ ] Tone, style and 3 to 6 short rules are set.
  • [ ] Knowledge base has your policies, contact details and payment methods, all showing ready.
  • [ ] No prices, stock or hours in KB files.
  • [ ] Catalog or restaurant menu is complete, with Arabic names and aliases.
  • [ ] Branches and offers are published in Business records.
  • [ ] Working hours, timezone and delivery settings are correct.
  • [ ] WhatsApp AI automation is On, and other channels are connected.
  • [ ] You tested: an FAQ, a price question, a delivery fee, an order end to end, and "I want to talk to a person".
  • [ ] Your team knows how to take over, pause the assistant and approve orders.
  • [ ] Automatic quality review is on at a low sample rate.

Keep improving#

Once a week, open the AI quality report, read the lowest-scoring conversations, and fix the source for each one. Review Learned from chats items and publish the good ones. Repeat until the same problems stop appearing.

Improving answer quality | Sela docs | Sela