Most knowledge bases get written from the inside out. Whoever knows the program best sits down with the policy document and produces articles in the order the policy is organized. The result is accurate and largely unread, because the article set mirrors the rulebook instead of the phone queue. A call center knowledge base built the other way around starts with what callers already said out loud. This post walks the pipeline: pull transcripts, extract the questions people actually asked, cluster those questions into topics, rank the topics by measured demand, then write in that order.
Nothing happens until recording and transcription are configured
The pipeline depends on transcripts existing, and transcripts are a configuration decision made long before anyone plans a call center knowledge base. In Amazon Connect, a "Set recording and analytics behavior" block has to sit in the contact flow before the contact reaches an agent. Voice recording can be set to agent and customer, agent only, or customer only. Chat transcripts are recorded automatically once enhanced contact monitoring is enabled. Reviewers need security profile permissions before they can open a recording at all.
Two practical consequences follow. First, placement is retroactive in only one direction: if the block was added late or sits after the routing decision, the contacts before that point produce nothing you can mine. Check placement before you promise anyone a content plan. Second, someone has to be permitted to read what gets recorded. Decide early who reviews transcripts, under which security profile, and for what stated purpose.
On the state education benefits program Mpathic operates, call recording and transcription were already part of the reported operating set, used for management review, quality assurance, coaching and up-skilling. The content work did not require new instrumentation. It required a second use for instrumentation that was already running.
Extract questions, not summaries
A summary tells you what a call was about. A question tells you what the caller wanted to know, which is the only unit an article can answer. Those are different outputs, and teams default to the first because it is easier to write.
The discipline is narrow. Read the transcript and write down the caller's ask as an interrogative sentence, in the caller's vocabulary rather than the program's. If the caller says "does the money expire," that is the question. It does not become "clarification of fund expiration policy." Keep the caller's noun for the thing, even when the program calls it something else, because the caller's noun is what gets typed into a search box.
One contact usually yields more than one question. Across the phase, the team reviewed transcripts and extracted thousands of questions. That ratio is itself a finding. A contact typically carries a primary ask plus one or two follow-ons, and a call center knowledge base written from call summaries only ever answers the primary ask.
Deciding a cluster is really one topic
Grouping is where a question list turns into an article plan, and where most of the judgment lives. Use one test: do these questions resolve to the same answer, stated by the same owner, under the same rule? If yes, they are one topic with many phrasings. If the answer forks depending on who is asking, they are not.
That fork is common in a program serving several constituencies at once. The state education benefits program supports families applying for benefits, enrolled students, schools and education providers, and program vendors, each with different questions and different stakes. A question about payment timing has a different answer for a family than for a vendor. Same words, two topics.
The phase closed with dozens of topics and articles from those questions. Articles outnumber topics because some topics split by audience and others split by step in a process that a single article would have crowded.
What ranking a call center knowledge base by demand means
Demand gets counted at the topic level, never at the phrasing level. A topic reached by 40 different wordings is one topic with heavy demand, not 40 small ones, and counting phrasings would push you toward writing many thin articles about the same rule. Rank by how many contacts touched the topic. Write in that order. Stop when the count flattens out, and leave the tail unwritten until it earns attention.
Published standards do not settle this for you. ISO 18295-1:2017 specifies service requirements for customer contact centres and includes performance metric specifications, and it applies to in-house and outsourced centres across sectors. What it cannot tell you is which articles your callers need first. That ranking is local measurement, and it changes as a program changes.
Public answers need the program owner's signature
Program rules belong to the program owner, and an article in a public call center knowledge base that states a rule is program communication whether or not it looks like it. Build the sign-off loop before you build the queue of drafts.
The loop that worked: the team drafted from the ranked topic list, the program owner reviewed each draft for rule accuracy, and anything public-facing required the owner's approval before it went live. That mirrors how the training content was built in the first place, through a train-the-trainer engagement with the program administrator to develop program-specific scripts, FAQs and procedures, after which Mpathic owned ongoing training, quality assurance and compliance monitoring. The state agency contact center case study describes the knowledge base Mpathic designed and developed for that program, serving both as a single accurate reference for agents and as the basis for self-service.
Start with one week
You do not need a program to test the premise. Take one week of transcripts, extract the questions rather than the summaries, cluster them with the same-answer test, and rank the topics by how many contacts touched each one. Then compare your top five against your five most-read articles. If the lists match, your call center knowledge base is already pointed at demand. If they do not, you have found the gap between what your team knows and what your callers ask, plus the order to fix it in.
If you want help building that pipeline into a live operation, talk to our team. Mpathic delivers knowledge base design alongside customer support and CCaaS implementation work, and publishes operating detail in our premium resources.
Frequently asked questions
What is a call center knowledge base?+
A call center knowledge base is the structured set of articles agents use as an authoritative reference during contacts, and which customers can often search themselves. On the state education benefits program described above, Mpathic designed and developed one that served both purposes: a single accurate reference for program rules for agents, and the basis for self-service.
How do you build a knowledge base for a call center?+
Build it from recorded demand rather than from the policy document. Confirm recording and transcription are configured in the contact flow, extract the questions callers actually asked, cluster the questions into topics using a same-answer test, rank the topics by how many contacts touched each, and write in that order with the program owner approving each draft.
What should go in a call center knowledge base first?+
The topic the most contacts touched. Counting demand at the topic level, not by individual phrasing, keeps you from writing many thin articles about one rule. Write down the ranked list, work from the top, and leave the long tail unwritten until contact volume justifies it.
Do you need call recording to build a knowledge base?+
You need recording or transcription configured before the contacts you want to mine, since neither is retroactive. In Amazon Connect the recording and analytics block must sit in the contact flow before the contact reaches an agent, and reviewers need security profile permissions to open recordings.
Who should approve knowledge base articles?+
The owner of the rules being stated. In a program engagement that is the program or product owner, not the contact center. Draft from the ranked topic list, route each draft to the owner for accuracy review, and require the owner's approval on anything customer-facing, because a public article stating a rule functions as program communication.
