Knowledge management for customer service where every answer is cited and articles a policy change contradicts are held from the bot within minutes
Policy changes reach every article in a day, and the bot never answers from a stale page.
See one case, screen by screen ↓Knowledge management for customer service that answers advisor and customer questions from your policies with citations, and flags stale articles. Seven agents answer with every sentence cited, catch the articles a policy change contradicts and hold them from the bot, trace who was already told the wrong thing, find missing articles and draft the fixes. The named owner approves every change.
How to choose AI solutions for back-office and enterprise work →
Updated
A policy changes in Commercial or Legal, and customer service hears about it in the weekly ops brief, from a complaint or from an advisor's flag, often days after it took effect. The knowledge team then searches the help desk by keyword for every article that retells the old rule, and keyword search misses the paraphrases: “a month”, “four weeks”, a row in a table. Each fix is edited, reviewed and published by hand while the backlog of reviews keeps growing.
Meanwhile advisors read the old article aloud and the chat bot quotes it word for word. Nobody can say how many customers were already told the wrong thing, and a tribunal has held a company liable for its chatbot's wrong answer. The gaps show up the same way — advisors improvising, searches with no result, escalations — and the answer the best advisors found never makes it into an article.
Estimated split of a 23-day median from a policy change to every affected article fixed, by hand and with the solution, in days.
Seven agents answer with citations, catch the articles a policy change breaks, and draft the fix; the named owner approves.
Kestrel & Vane Group runs one Customer Care operation for its stores, online shop and paid Vane Plus membership: 6,140 articles on 12 shelves, resting on 186 policies, in 11 languages. Returns Policy v7 took effect on Monday 12 October. Commercial uploaded it on Wednesday at 09:28. Here is what happens next, screen by screen, in the working solution.
Priya owns the Returns & refunds and Vane Plus membership shelves. Her library shows 5,412 articles fresh, 611 stale and due for review, 117 in conflict and held from the bot, and 23 drafts waiting for owners, 4 overdue. Each article is a spine coloured fresh, stale, in conflict or draft waiting. The feed on the right shows what the agents did since 07:58 this morning.
Today in the library: “Returns Policy v7 broke 3 articles. Uploaded 09:28, effective since Monday. KB-1182, KB-1207 and KB-0954 now contradict §3.2 and §3.4. All three are held from the bot.”
The Librarian diffs the policy: §3.2 now gives Vane Plus members 45 days to return opened items, including audio products; §3.4 cuts the restocking fee from 15% to 10%, with no fee for members within 14 days; §3.6 adds gift receipts. It re-reads the 14 articles anchored to §3: 3 are in conflict, 11 agree. KB-1207's table row says members have “a month”, and the Librarian reads that as 30 days.
“Effective Mon 12 Oct, uploaded Wed 14 Oct 09:28. For 2 days and 9 hours the library and the bot ran on v6.”
Every sentence of “Returning opened headphones and earbuds” is linked to the clause it rests on. Two are marked against v7: the 30-day window for everyone and the 15% restocking fee. The article is held from the bot; advisors still see it, marked. It was quoted in 312 bot answers this week and viewed 6,240 times in 30 days.
“v7 §3.2 gives Vane Plus members 45 days. This sentence says 30 days for everyone.”
The Impact agent reads the answer log from Monday 00:00 to 09:31: 312 bot answers and 146 advisor views quoted the wrong sentence. 87 Vane Plus members asking about day 31 to 45 were told no. Checking each delivery date, 79 can still return and 8 are past day 45, so goodwill needs approval. It drafts the correction for Sofia Brandt, team lead for Returns & Refunds.
The Impact agent never sends to a customer. Nothing goes out until the team lead approves.
The Writer drafts fix D-4471: two of six sentences rewritten from v7 §3.2 and §3.4, four kept, the reading level unchanged. The Checker then answers 11 real questions from this week — 8 bot chats and 3 advisor searches — from the draft alone, and compares each answer with the policy. 11 of 11 are correct, every sentence cited, so the draft goes to Priya.
Earlier that morning the Checker sent D-4476 back: 5 of 6 questions passed, and “Do old gift cards expire?” failed.
Rhys Calloway calls about Aurel ANC-700 headphones delivered on 6 September, day 38. The bot told him 30 days on Monday; the advisor's view flags him as one of the 87. The suggested answer reads his membership from the customer record and cites Returns Policy v7 §3.1, §3.2 and §3.4, with a version to say aloud.
Say it like this: “Good news, Mr Calloway: as a Vane Plus member you have 45 days from delivery to return opened headphones. Yours arrived on 6 September, so you have until 21 October. A 10% restocking fee applies because it is past 14 days, unless they are faulty.”
On the shelf, the same question returns in 1.8 seconds: 6,140 articles and 186 policies searched, 14 passages read. KB-1182 is in conflict, so it is not quoted. The Citation checker reads every sentence against its passage, confidence is 0.96, and the spines the answer read light up on the shelf.
“Yes, for Vane Plus members. Members can return opened headphones and earbuds within 45 days of delivery. Customers without Vane Plus keep 30 days.”
Daniel owns the chat bot's nine rules. He tests “What is the restocking fee on opened earbuds?”: the bot says it is an AI in its first message, finds KB-0954 in conflict, refuses to quote it and answers from Returns Policy v7 §3.4 instead. This week the rules stopped or handed off 1,906 of 107,270 bot answers.
“The bot is never more confident than the article it reads.”
Priya reads the redline, the clause it now matches and the replay, and approves with the reason code “Policy change”. KB-1182 v13 is published to advisors in 11 languages, and the bot may quote it again. The person who drafted can never approve, and nothing is published without its named owner.
“Priya Nair · Signs as the named owner · 09:42”
The audit trail holds each step in order: Elena Varga's upload at 09:28, the sweep and the hold at 09:31, the trace at 09:33, the replay at 09:34, and the decisions that follow. Entries are hash-chained so nothing can be changed afterwards, and an evidence pack per article exports every version, the clauses it rested on, the approvals and the answers that quoted it.
“Every published version, decision, rule change and agent run, chained so that nothing can be changed afterwards.”
The same policy change, seen by the five people who carry it — what their week looked like, and what it looks like now.
Seven agents answer with citations, catch the articles a policy change breaks, and draft the fix; the named owner approves.
Answers advisors and the bot in plain words from published articles and policies, with every sentence cited. It drops any source that is held or stale before it writes.
Reads each sentence of every answer against its passage and removes what is not supported. This week it removed 2,311 unsupported sentences, 1.1% of sentences.
Links every article sentence to the policy clause it rests on. When a policy version lands, it diffs the clauses, re-reads the anchored sentences by meaning and holds any that now contradict it from the bot.
Traces every answer that quoted a held sentence, finds the customers who acted on it, checks their orders and drafts a correction for the team lead.
Clusters contacts that got no answer, low-confidence answers, “didn't help” flags and escalations, and finds the advisors who resolved them anyway.
Drafts fixes and new articles in the house structure from the policy clause and advisors' resolution notes, changing only what the policy changed.
Re-asks the real customer questions and the article's regression set against each draft, on its own, before the owner sees it.
Publishes; team lead approves corrections. The agents propose; a named person decides.
Ask about any article, policy, gap or answer — or say what to hold. Every answer points to its source.
Which articles did Returns Policy v7 break?
Three of the 14 articles anchored to §3 now contradict it: KB-1182 and KB-1207 still give members 30 days for opened items, against 45 days in §3.2; KB-0954 still says a 15% restocking fee, against 10% in §3.4. All three were held from the bot at 09:31, and fixes D-4471 to D-4473 are waiting for Priya Nair.
How many customers did KB-1182 mislead this week?
87 Vane Plus members. KB-1182 was quoted in 312 bot answers and 146 advisor views from Monday 00:00 to 09:31; 87 of those were members asking about day 31 to 45 who were told no. 79 can still return; 8 are now past day 45. A correction is drafted for Sofia Brandt.
Draft the missing article on gifts without a receipt
It is already drafted: D-4480, from 14 advisors' resolved cases and the new Returns Policy v7 §3.6. 1,284 contacts asked this in 30 days and 708 bot chats ended with no answer. The Checker replayed 9 real questions; 9 answered correctly.
Hold every article citing Payments Terms v6 §7.2 from the bot
Done. 2 articles rest on §7.2: KB-5174 “Gift cards: balance and expiry” was already held, and KB-5192 “Promotional codes” is now held as well. Advisors still see both, marked as held. The change is logged and Lena Fischer is told.
13 screens from the working solution, on its sample data. Pick one to see it large.
How confident an answer must be, what makes an article stale and how fast a fix is due are settings, saved as versions and on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Bot confidence floor | 0.80 | Below it the bot hands to a person |
| Advisor confidence floor | 0.70 | Below it advisors see “no answer” |
| Checker pass mark | 100% | Every replayed question must be answered correctly |
| Questions replayed per draft | Up to 25 real questions + the regression set | Each answered from the draft alone |
| Stale after | A cited clause changed, or the review date passed | Review cycle per shelf: 90 · 180 · 365 days |
| Bot hands off refunds over | $500 | Changed from $750 by the bot product owner; rules versioned |
| Fix for a conflict due in | 1 business day | New article from a gap 5 business days · review renewal 10 business days · customer correction 2 business days · safety gap 4 hours |
| Notifications | Owner digest at 08:00; conflicts at once | SLA warning at the warn mark · weekly report to the CCO, Mondays 07:00 |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Gartner press releases 2024-12-09 — 61% of service leaders have article backlogs; 2024-08-19 — 14% of issues fully resolved in self-service, and in 43% of failures customers could not find relevant content). The tribunal case is 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, February 2024). EU AI Act Article 50, which asks AI systems that talk with people to say they are AI, applies from 2 August 2026. People and companies named on this page are fictional — characters and sample data in the working solution.
Knowledge management for customer service that answers advisor and customer questions from your policies with citations, and flags stale articles. Seven agents answer with every sentence cited, catch the articles a policy change contradicts and hold them from the bot, trace who was already told the wrong thing, find missing articles and draft the fixes. The named owner approves every change.
The Librarian agent links every article sentence to the policy clause it rests on. When a new version lands it compares the policy clause by clause and re-reads only the sentences anchored to what changed, by meaning rather than keywords, so “a month” is caught as surely as “30 days”. In the working solution it swept 3 minutes after the upload and held 3 of 14 articles.
Not once the library knows the article is wrong. An article that is stale or in conflict with a policy version in the library is held from the bot, which answers from the policy itself or hands to a person. The library can only catch a change once the new version is uploaded: in the working solution, Returns Policy v7 arrived two days after it took effect, and the bot ran on v6 until the Librarian swept 3 minutes after the upload.
Yes. The Impact agent traces every bot answer and advisor view that quoted a held sentence, matches them to customers and checks their orders. In the working solution it found 312 bot answers and 146 advisor views, 87 members told no, 79 who could still return and 8 needing goodwill, and drafted a correction the team lead approves.
The Gap agent clusters contacts that got no answer, low-confidence answers, “didn't help” flags and escalations, and finds the advisors who resolved them anyway. The Writer drafts the article from their resolution notes and credits them, as knowledge-centered service intends. Safety topics are routed to the safety owner and never drafted.
Yes. The Checker agent re-asks the real customer questions that exposed the problem, plus the article's regression set, against the draft alone. A draft must answer every question correctly to reach its owner; one that fails goes back to the Writer with the failing question.
Yes. The agents draft and recommend; they never publish, release a held article or contact a customer. The named shelf owner approves every article with a reason code, the person who drafted cannot approve, the team lead approves customer corrections, and the bot product owner changes the bot's rules as versions.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your help desk articles, policy library and a sample of contacts — and it goes live once every quality gate has passed. We will walk you through it on your own knowledge base first.
Invalid claims stopped before payment, and supplier defects charged back with proof.
1-min demoReturns ManagementFraud held before the refund; every item sent to the lane that recovers the most.
1-min demoChargeback ManagementMore chargebacks won, weak ones accepted fast, and no dispute left to expire.
1-min demoWe'll run Customer Service Knowledge Management on a sample of your own articles, policies and contacts.