Building a Knowledge Base That Makes AI Call Analysis Accurate

A knowledge base for AI call analysis is the set of facts about your business that the AI reads before it judges a sales call: what you sell, what it costs, who it’s for, how your sales process is supposed to run and what a good call sounds like. Without it, even a strong model scores your calls against a generic idea of “good sales.” It will flag a rep for not discussing price when your process deliberately saves price for the second call. It will praise a call that sailed through discovery without noticing the prospect was nowhere near your ideal customer.

Most of the complaints teams have about AI call analysis, “the scores don’t match what my sales manager thinks,” “the summaries miss the point,” “it keeps recommending things we don’t do,” come down to missing context, not a weak model. This post covers what to put in the knowledge base, how to write entries the AI can actually use, the mistakes that make analysis worse, and how to check whether it’s working.

Why context matters more than the model

Think about how you’d onboard a new sales manager. You wouldn’t hand them a stack of recordings on day one and ask for scores. You’d explain the offer, the price, the typical buyer, the stages of your call, the objections you hear most and what separates your best closer from the rest. Only then would their call reviews be worth reading.

AI analysis is the same. A model can transcribe, summarize and spot patterns very well. What it can’t do is know that your $12,000 program has a payment plan, that you don’t work with pre-revenue founders, that setter calls should end with a booking rather than a pitch, or that “we’ll send you the contract” is how your closers handle a yes. That’s the knowledge base’s job.

What belongs in the knowledge base

Keep it to the facts that change how a call should be judged. Here’s the set we’d start with for a high-ticket business.

Company and offer

• What you sell, in two or three sentences, the way your best closer would explain it
• Price points, payment plans and any deposits or guarantees
• What’s included and, importantly, what isn’t
• Typical results and timelines you’re comfortable claiming

Ideal customer and disqualifiers

• Who the offer is for: industry, size, stage, revenue range
• Hard disqualifiers: who you won’t sell to, and why
• Soft signals of a poor fit that a rep should probe

Sales process

• The stages of each call type, in order, with the goal of each stage
• What should happen at the end of a setter call versus a closer call
• When price is supposed to come up
• How a yes is closed: contract, payment link, onboarding call

Objections and positioning

• The five or six objections you hear most, in buyers’ words, and how your process expects them to be handled
• How you position against the alternatives prospects mention, including doing nothing

If you’ve built an objection library from real calls, that’s the best source for this section. See building an objection library from lost sales calls.

How to write entries the AI can use

The format matters as much as the content. A few rules from experience:

Write facts and rules, not marketing copy. “Our transformational program unlocks your potential” tells the AI nothing. “12-week group program, $9,500 or 3 payments of $3,400, for agency owners doing $30k to $150k a month” tells it exactly what to check.

Be explicit about what good looks like. “Closers should present the price only after the prospect has agreed the gap between their current and desired revenue is worth solving” is a rule the AI can score against.

State the exceptions. If price comes up early because the prospect asks, say what the rep should do. Otherwise the AI will flag every early price mention as a deviation.

One topic per entry. An entry about pricing and an entry about the ideal customer are easier to keep current than one giant document.

Keep entries short and dense. Many systems cap entry length. Even when they don’t, a concise entry is more reliably applied than a long one where the key rule is buried in paragraph six.

Update when the business changes. A new price or a new guarantee should hit the knowledge base the same day it hits the sales script.

Mistakes that make analysis worse

• Pasting the whole sales script. A 10-page script contains too many details of equal weight. Extract the stages and the rules instead.
• Contradicting your own process. If one entry says “never discount” and another describes a “fast-action bonus,” the AI has no way to know which wins.
• Describing the ideal call, not the real one. If your best closers routinely skip a stage because it doesn’t work, write the process they actually follow, or the AI will mark your best calls down.
• Leaving out disqualifiers. Without them, the AI can’t tell a well-run call with a bad-fit lead from a badly run call with a good one. That’s the distinction your marketing team needs most.
• Treating it as a one-time setup. A knowledge base written at launch and never touched will be wrong within a quarter.

Separate setter calls from closer calls

If your team has setters and closers, a single rubric for every call will produce bad scores for both. A setter’s 15-minute qualification call shouldn’t be judged on whether it handled price objections. A closer’s strategy call shouldn’t be judged on whether it booked a next meeting.

Give each call type its own section in the knowledge base, with its own stages and success criteria, and make sure calls are labelled by type so the right rules apply. In Callix, Call Types live in the same place, under Settings → Knowledge base → Call Types, and a call takes its type from the booked event type, a rep’s default or a manual pick.

How it works in Callix

In Callix, the knowledge base sits under Settings → Knowledge base. Company Info and Context entries shape call analysis, rep scorecards and dial analysis, so a call is judged against your own offer, pricing and process. Each entry can hold up to 2,000 characters, which is a useful forcing function for the “short and dense” rule above. An analysis that used the knowledge base shows a Knowledge base applied badge, so you can tell at a glance which calls were judged with your context and which weren’t.

You don’t have to write entries in the settings screen. You can add them from AI chat, where Callix AI drafts the entry and you confirm it before it’s saved, which is handy right after a call review when you realize the analysis is missing a rule. You can also add entries from the API at /api/v1/knowledge-base, or with the manage_knowledge_base tool from an AI assistant connected to the Callix MCP server. That last route suits teams who keep their offer details in a doc and want an assistant to keep the knowledge base in sync with it.

A worked example: before and after

Say you sell a $15,000 consulting engagement. Before adding context, the AI’s analysis of a closer call says:

• “Rep did not discuss pricing until minute 41. Recommend introducing price earlier to qualify budget.”
• “Prospect mentioned they are pre-revenue. Rep handled this well by focusing on vision.”
• Score: 82.

Your sales manager disagrees on both points. Price at minute 41 is your process. And a pre-revenue prospect shouldn’t have been on the call at all; the rep should have politely ended it.

You add three entries: an offer entry with the price and payment terms, a process entry saying price is presented only after the gap is agreed (typically past the 35-minute mark), and an ideal-customer entry stating that you work only with businesses above $500,000 in annual revenue.

The re-analysis says:

• “Price introduced after gap agreement, consistent with process.”
• “Prospect is pre-revenue, which is outside the ideal customer profile. Rep continued the full pitch instead of disqualifying. Flag lead source for review.”
• Score: 58.

Now the score matches the manager’s judgment, and there’s a new, useful signal: the lead source that booked a pre-revenue prospect. That’s a marketing problem the uninformed analysis hid.

How to test whether your knowledge base is working

Don’t trust it blindly. Run a simple calibration:

1. Pick 10 recent calls your sales manager has already reviewed and scored by hand.
2. Compare the AI’s scores and notes to the manager’s. Look for patterns in disagreement rather than individual misses.
3. For each pattern, ask what fact the AI was missing. Usually it’s a process rule, a disqualifier or an exception.
4. Add or fix the entry, re-run the analysis on the same 10 calls, and compare again.
5. Repeat monthly with fresh calls.

Two or three rounds usually close most of the gap. When they don’t, the disagreement is often worth discussing as a team; sometimes the AI is pointing at something the manager had stopped noticing.

Keeping it current

Assign an owner, typically the sales manager or whoever owns the script. Review the knowledge base:

• Whenever the offer, price, guarantee or process changes, the same day.
• Monthly, alongside the calibration check above.
• After any surprising analysis, in either direction. A call that scored suspiciously high or low is often a sign of a missing or outdated entry.

AI call analysis gets talked about as a model problem. In practice it’s mostly a documentation problem, and that’s good news, because documentation is something your team fully controls. For how better analysis feeds back into marketing, see using AI to summarize sales calls into marketing insights.