Concept buildA worked example of how we would build this for a typical business. Not client work: no client, no client data, no achieved results.
AI Chatbots
Concept buildA concept build for a B2B software company whose help centre is thorough but hard to search. The copilot answers product questions from the published documentation, shows the passages each answer came from, and opens a support ticket, carrying the question and the sources it tried, whenever it is not confident.
The brief
A typical B2B SaaS company with a few hundred documentation pages still receives a steady stream of “how do I…” tickets whose answers are already written down. Customers search in their own words rather than the product’s feature names, keyword search misses, and the ticket lands with a support engineer who replies with a docs link. Generic chat widgets make it worse when they answer confidently from the wrong product version or describe a setting that does not exist; one wrong answer about permissions or billing costs more trust than ten slow ones.
Explainer
Explainer · 7 slides · about 46 s
Concept buildConcept build: Cited Docs Copilot for a B2B SaaS Help Centre
Concept build, explained in 7 slides: a concept build for a B2B software company whose help centre is thorough but hard to search.
Slide 1 of 7
01 · AI chatbots
Cited Docs Copilot for a B2B SaaS Help Centre
Concept build: Cited Docs Copilot for a B2B SaaS Help Centre.
02 · The problem
What a typical B2B SaaS company is up against
A typical B2B SaaS company with a few hundred documentation pages still receives a steady stream of “how do I…” tickets whose…
03 · The approach
How we would build it, in 5 steps
We would build it in 5 steps, from question bank from real tickets to a docs-gap report for the writers.
04 · Architecture
How the pieces connect
How the parts connect, from the first input to the person who reviews the result.
05 · Demo · sample run
A how-to question, then one the docs cannot answer
A sample run, step by step, including where a person steps in.
06 · Typical targets
What we would measure it against
Typical targets, each with the basis it would be measured against.
07 · Next step
Scope a build like this for your business
Book a strategy call to scope a build like this.
Architecture
Architecture of the Cited Docs Copilot for a B2B SaaS Help Centre concept: Help widget → Hybrid search (question); Help-centre CMS → Hybrid search (indexed pages); Hybrid search → Cited-answer LLM (top passages); Cited-answer LLM → Support team (ticket + context); Support team → Docs-gap report (weekly gaps).
Diagram with 6 parts: Help widget (source), Help-centre CMS (source), Hybrid search (processing), Cited-answer LLM (model), Support team (human), Docs-gap report (system of record).
Connections:
Sample run
User: how do I set up SSO?Routine
Hybrid search returns three passages from the SSO setup guide.
Answer with citation chipsRoutine
A step-by-step answer; each step links to the section it came from.
Follow-up: SSO for a sub-account?Needs care
Retrieved passages are weak and disagree, so confidence is low.
Copilot says it is not sureNeeds care
No answer is generated; it offers to open a ticket instead.
Ticket opened for supportA person takes over
The question, the passages tried and the page URL go to the help desk.
Build plan
We pull about sixty days of “how do I” tickets and search logs and turn them into a 150-question evaluation set, each question paired with the doc page that answers it. That set defines “working”: retrieval and answers are scored against it before launch and again after every docs or prompt change.
The copilot indexes the published help centre straight from the CMS, chunked by heading, with product area, plan and version stored on every chunk. A docs change re-indexes only the pages that changed, so answers follow the documentation on the day it is published. Internal and draft pages are never indexed.
Each question runs keyword (BM25) and embedding search together and keeps the strongest passages. The model may answer only from those passages and must cite them, so every answer shows citation chips that open the exact section. If nothing relevant comes back, no answer is generated.
When retrieval scores are weak, the passages disagree, or the question touches billing or account security, the copilot says it is not sure and offers to open a ticket in the helpdesk. The ticket carries the question, the passages tried and the page the customer was on, so support starts from context.
Questions that end in a ticket are grouped weekly into a short report: missing pages, outdated screenshots, features customers call by another name. Fixing the docs is what reduces tickets over time, so the copilot doubles as the documentation team’s list of what to write next.
Typical targets
Typical target
Every answer cites the documentation it came from
Basis: Design property: an answer without a retrieved source is blocked before it is shown
Typical target
20–35% fewer “how do I” tickets
Basis: Compared with the 60-day pre-launch baseline for that ticket category, measured during the pilot
Typical target
80–90% of evaluation questions retrieve the right page in the top three results
Basis: Retrieval hit rate on the 150-question set; the exact launch threshold is agreed before go-live
Typical target
Low-confidence questions reach support with full context
Basis: Design property: the ticket includes the question, the passages tried and the page URL
Your project
Questions taken from your tickets and transcripts, each paired with the source that answers it, with a pass threshold agreed before launch.
The knowledge index, model accounts and channel numbers are set up under your business, so the data and the configuration stay yours.
Message templates, consent wording and widget placement checked against each channel’s rules where your customers are.
Handoff and resolution tracked per intent; scope widens only when the numbers hold for a full week.
Where this fits
FAQ
No. The copilot is a concept build, and no company’s documentation sits behind it. We wrote it to show how we would approach citations, evaluation and ticket handoff for a typical B2B SaaS product; the demo questions are invented, and the ticket-reduction range is a target with its basis, not a measured outcome.
Answers come only from the indexed docs, and the index refreshes from the CMS whenever a page is published. Each chunk carries its product version, so a customer on an older release is matched to the right page, and the evaluation set re-runs after each docs release to catch regressions.
Not in this design. It reads public documentation only. Questions about a specific account, invoice or permission go to support as a ticket, because they need a person who can see that account.
Yes, once they are published. Release notes are indexed as their own source type with the version and date, so a question about a recent change can cite the note, and when the main guide catches up the guide wins. Unpublished roadmap items are never indexed, so the copilot cannot reveal plans.
The evaluation set re-runs after every docs or prompt release, and a drop below the agreed retrieval threshold blocks that release from going live. Between releases a weekly sample of real conversations is reviewed, and any answer a customer marks unhelpful is added to the evaluation set.
Worked examples
Concept builds are worked examples, not client work.
Concept buildAI Chatbots
A concept build showing how we would build a support bot for an online electronics store whose helpdesk fills with the same few questions. The bot would answer order, delivery, returns and product-fit questions from the store’s catalogue, policies and live order data, on the website and WhatsApp, and pass everything else to the helpdesk with the conversation attached.
Concept buildAI Chatbots
A concept build for a direct-to-consumer brand in India whose WhatsApp inbox fills with the same two questions: where is my order, and how do I return it. The assistant answers in English, Hindi or Hinglish from live order data and the brand’s own returns policy, and hands anything unusual to a person with the whole thread attached.
Concept buildAI Chatbots
A concept build for a boutique hotel in the UAE whose front desk answers the same guest questions in several languages, day and night. The concierge replies on WhatsApp and the hotel website in the guest’s language, answers from the hotel’s own guide, turns requests into housekeeping tickets with the room number, and puts complaints in front of the duty manager straight away.
Next step
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment
