Which sources to connect to an AI Assistant (and which to filter out)

The sources that lift an AI assistant's answer rate the most are past conversations, the team's internal chat, and live customer data from the product and the CRM (customer relationship management, where account records live). Help center articles and SOPs (standard operating procedures, the written steps a team follows) cover the questions somebody already documented.

Everything else the team knows sits somewhere nobody indexed.

Most of what a support team knows has never been written down.

That gap is where answer rates stall. An assistant connected only to documentation answers documentation questions, and customers rarely ask those.

How much does each source actually move the number?

Answer rate, meaning the share of questions the assistant resolves without escalating and without saying it doesn't know, climbs in visible steps as sources get connected. These are the bands seen across Zentesimal accounts in production during 2026.

The bands move with the business. A product with a narrow feature set, few plan variations and a mostly self serve customer base sits at the top of each range. One where the answer to almost any question depends on the specific account, the contract or the integration sits at the bottom, and gains more from the customer data step than from anything else.

A team whose last month of tickets is mostly account lookups will land in the lower half of each band. One where most tickets ask how a feature works will land in the upper half.

The first row is the one most teams stop at, and it is the cheapest to connect by a wide margin. A crawl of the help center takes an afternoon. The rows above it take real work.

The last row is a different kind of step. The first three are about what the assistant knows. The fourth is about what it is allowed to change.

An assistant with full context can explain, in five accurate steps, how a customer updates their billing email. The customer still has to do it, and some of them come back when step three looks different on their screen. The last fifteen points, from 80 to 95%, live in the gap between explaining a change and applying it.

That gap is also where the risk sits, which is why each action gets a level of control that matches it. Updating a contact detail or resending an invoice can be checked by the assistant itself. A plan change passes fixed rules in code first. A large refund waits for a human to approve it.

More context, or cleaner context?

Past a certain point, adding raw material makes an assistant worse. These systems pull the closest match to a question out of everything connected, and a close match to a bad answer is still a bad answer.

In one production account, an unfiltered archive taught the assistant about a discount coupon that had been retired months earlier. It offered the coupon to a customer. The code failed, the customer came back asking why, and the conversation escalated to a human.

The archive was accurate when it was written. That is exactly what makes it dangerous.

Nothing in that thread was wrong at the time. The assistant learned a version of the product that no longer existed, and the failure showed up as a frustrated customer rather than as an error anyone could see in a log.

A small filtered archive answers more questions correctly than a complete one. Coverage is the easy part. Precision is what moves the number that matters.

How do you filter past conversations without losing what's useful?

Four filters do most of the work, and each one removes a different kind of noise.

One off cases

A thread resolving a bug caused by one account's custom integration teaches nothing about the product. It teaches about that account.

Private data

Names, emails, invoice numbers, anything that identifies a person or a company. Frameworks like the NIST AI Risk Management Framework, the US government's reference guide for managing AI risk, treat this as a design decision rather than a cleanup step.

Commercial exceptions

Discounts, refunds, custom terms, retention offers. A human made a judgment call with authority behind it, and a model reads that as precedent. The retired coupon above came through this gap.

Anything the product has made obsolete

Threads about features that changed, flows that were replaced, prices that moved. This filter needs a date, not a rule.

What survives those four is the part that transfers: how the product behaves, how accounts get managed, and the phrasing that actually worked.

When are past conversations worth less than they look?

An archive loses value fast when the product or the account management model has changed underneath it. Machine learning has a name for the general version of this problem, concept drift: the data keeps changing until the thing built on top of it no longer matches reality.

Three situations where an old archive stops being an asset:

• The product shipped enough changes that the answers describe something that no longer exists.
• The team's account management approach shifted, so the right handling of a situation changed even though the product didn't.
• The conversations come from a period when agents were less experienced, and the archive preserves their mistakes with the same confidence as everything else.

An archive is a snapshot of how the product used to work, answered by whoever happened to be on shift.

The alternative to discarding it is qualifying it. Tagging the Q&A pairs by customer segment, so the assistant retrieves against accounts that resemble the one in front of it, recovers a lot of what a blanket date cutoff would throw away.

How does a team's chat become a knowledge base nobody wrote?

Slack and Teams hold the answers that never made it into a doc, usually because the person who knew them answered in thirty seconds and moved on.

The catch is that a channel holds questions, wrong guesses, half answers and the correct one, all with equal weight. Something has to mark which is which.

The signal can be a reaction on the right reply, a postprocessing pass over the thread, or anything else that tells the assistant which answer was the correct one.

In Zentesimal deployments that signal is usually an emoji, because it costs the team nothing. Whatever form it takes, the marking happens as a side effect of how the team already works. No one writes documentation, and the knowledge base grows anyway.

What happens when the assistant doesn't know?

It asks the team, the same way a new hire would ask the person two desks over.

A customer asks something outside everything connected. Instead of guessing or escalating the whole conversation, the assistant posts the question internally, waits for the answer, and comes back to the customer with it.

An assistant that asks is worth more than an assistant that guesses.

Two things happen at once. The customer gets a real answer instead of a ticket number, and the archive gains one more entry that did not exist that morning.

Human oversight stops being a review queue and becomes part of how the system runs, which is roughly the direction the OECD's work on AI and work, the intergovernmental body's research hub, points to.

What does customer data change before the first message?

Plan, language, past conversations and product usage loaded before the first message change what the conversation costs the customer.

Without them, the assistant opens by asking which plan the customer is on. The customer already told someone that, twice, and now the conversation is about admin instead of the problem.

With them, the assistant goes straight to checking the account and answers for that specific plan, profile and situation. Resolution time drops because the back and forth that produced it never happens.

That difference is where most avoidable escalations live. The customer who escalates is rarely the one with the hard problem. It is the one who has explained their setup three times, which is the same reason escalation rate makes a poor success metric on its own.

Some of those conversations don't end in a ticket. They end in a renewal that doesn't happen.

The same context that answers also sells

An assistant that knows the plan, the usage and the history has everything it needs to spot an expansion moment or a churn risk while the conversation is happening.

A customer hitting a plan limit asks a technical question. The answer to that question and the upgrade conversation are the same conversation, and the context was already loaded.

In production, Zentesimal assistants sell over $1.5k a day on average, only through chat (internal Zentesimal figures, 2026). With the right sources and the right customer context, an assistant that only helps customers turns into a sales manager that converts users into customers. See how.

The connected context does not change between jobs. The same account history that resolves a support question is what makes a conversion moment visible, and what flags a retention risk early enough to matter.

What actually separates a 40% assistant from an 80% one?

Of everything that gets recommended when a team connects sources, three things carry most of the difference, and one popular one carries almost none.

A way to mark correct answers. Without it, every source is a pile of text where right and wrong weigh the same. The marking has to come from behaviour the team already has, not from a process nobody will follow.

A date on everything historical. The retired coupon is what an undated archive costs. This is the cheapest filter to build and the one most often skipped.

Account context before the first message. It is the single biggest jump in the table above, from the 50 to 65% band into 65 to 80%.

The one that moves least: expanding help center coverage. Writing more articles is the most visible response to a low answer rate and the one with the flattest return, because the questions that go unanswered are usually the ones nobody thought to document in the first place.

Give every customer a unique, personalised experience, with an AI manager for every user, at scale.

See how it works