A multi-brand car dealership had already tried an AI telephone responder. It answered, collected the caller’s details and passed them to a human. Its marketing manager closed the experiment with a sentence that works as an evaluation criterion: investing time and money in that was too limited to be worth it. A different principle drove the next project, and that principle governs how any knowledge base for AI agents performs: separate information that stays still from information that changes within minutes.
Used-car stock made the point in the meeting. A vehicle in a particular colour shows as available at nine in the morning and sells ten minutes later. An agent answering from a document uploaded last week promises a car that has already left the forecourt. An agent querying the business system at the moment of the question says no, which is the correct answer.
What belongs in a knowledge base for AI agents, and what gets read on demand
Stable information belongs inside: company identity, locations, procedures, service conditions. Stock, prices and progress statuses stay outside, read in real time through APIs.
Separating the two categories carries a contractual consequence that few people anticipate at proposal stage. Volatile information requires an integration towards whichever system holds it, and that integration has to be negotiated with whoever owns that system. At a mobility aids company the head of administration put it bluntly: data owners guard their systems closely, and her fear was that they would never grant write access. She was right, and the project started read-only.
Format is the second selection criterion. A knowledge base for AI agents accepts PDF, Word, plain text, spreadsheets and markdown, plus web sources that are refreshed or reread at regular intervals, typically once a day. Compare a catalogue exported as a marketing-designed PDF with the same catalogue exported as a table: the second produces better answers, because columns, headers and rows tell the agent how the information connects, whereas print layout hands it over in no particular order.
A knowledge base for AI agents has a shared area and a private one
Two areas make up a knowledge base for AI agents, and the split is where answer quality is decided. Documents loaded into the shared base are visible to every agent, including those created later. Anything loaded inside a single agent stays invisible to the others, and that agent gives it priority over the shared base.
One head of administration summarised the criterion better than any manual: whatever applies to everyone goes in the shared base, whatever needs specialising because it would otherwise cause confusion goes into the individual agent. On a company intranet, misplaced content costs a fruitless search. Here it costs a wrong answer given to a customer, delivered with exactly the confidence a correct one would have carried.
Volume logic runs counter to intuition. No limit applies to the number of documents in the shared base, while an individual agent’s private base is deliberately small, in the order of ten documents on entry plans and a few dozen on advanced configurations. Whoever builds the platform explains why: an agent performs poorly with many documents and much better with a few relevant ones. Technical limits do apply to individual file size and total storage, and both can be extended.
| Type of information | Where it lives | How it is updated |
|---|---|---|
| Identity, locations, opening hours, service conditions | Shared knowledge base | Uploaded directly by the client |
| Department procedures, service price lists, tone of voice | Private base of the individual agent | Uploaded directly, few selected documents |
| Availability, case status, order status | Source business system | Queried through APIs at each request |
| Recurring questions and approved answers | Question and answer area | Edited immediately, no retraining |
Why a complex process is split across several agents
Agent architecture decides how the information base is organised, because each agent carries its own document perimeter. Questions about the number of agents therefore come before questions about content.
One question arrives in every evaluation: does an artificial intelligence really need specialised sub-agents, or can a single agent handle everything? Limits provide the technical answer. A single agent does not reliably distinguish between two categories of request that sit far apart, and narrowing its perimeter is the only way to govern how it behaves.
Office organisation explains the criterion well. People work within their own areas because nobody holds every skill, and a flow built on specialised agents follows the same logic. One agent routes, another retrieves documents, another extracts variables from text, another handles recurring questions, while non-conversational blocks send an email or hand the conversation to a person.
Numbers cannot grow without consequences, though. Every hop between agents adds latency and a risk of the message being restated: a flow of medium complexity settles at five or six agents, not twenty. That was precisely the dealership manager’s concern, since a customer moves from new to used to rental to trade-in inside a single conversation, and the bouncing between agents must never reach them.
Who keeps a knowledge base for AI agents current after go-live
Almost every evaluation ends on the same question: if changes are needed later, can we make them ourselves or is your involvement always required? A correct answer separates two things clients tend to merge.
Content inside a knowledge base for AI agents is maintained by the client, after initial training. A document added, or a block of questions and answers edited, is picked up by the system and the agent aligns immediately, with no training cycle. Flow logic is different: deciding who answers what, and in which order, is built with the supplier and modified with the supplier, at least until the internal team has learned the method.
How much autonomy a correction allows also depends on the architecture chosen, which is less intuitive. During one demo an agent collected the caller’s phone number but omitted first and last name. Where behaviour is governed by an agent instructed in natural language, the fix is a line added to its instructions and the client applies it. Where the same field is governed by an automation block with fixed fields, no line exists to write and the correction happens in the mandatory field settings. Same fix, two different routes.
The routine that keeps the knowledge base current
Maintenance discipline deserves to be explicit from the first month. In organisations where the base stays reliable, updating is not a calendar activity: it is attached to the events that change the information. A new price list, an amended warranty condition, a site opening or closing, a rewritten procedure. Whoever signs off that change tells the content owner, who checks whether a document in the base needs replacing. Alongside that, a quarterly review of conversations where the agent handed over to a person gives the most reliable list of what is missing.
When the information base is incomplete or contradicts itself
A well-built knowledge base for AI agents anticipates missing data. The agent asks for it and carries on rather than stalling; where a request falls outside the perimeter it was instructed on, the flow hands over to a person, returning automatically to the agent if nobody is available. Edge cases measure the distance between a simple automated responder and agents built to handle real customer records.
Contradiction between documents is the more insidious problem, because it produces no visible error: it produces a plausible wrong answer. At a services company the manager spotted it before go-live, asking whether an agent learns from both sources when two pieces of content overlap. It does, which is why auditing existing documents comes before loading them.
One edge case observed on a third-party project marks the boundary. A company had outsourced the generation of self-billing invoices for purchases from abroad, three days of manual work a month. That system built each document from scratch on every run and produced errors reading VAT and distinguishing goods from services. A generative model is not deterministic, and offers no guarantee that the same document will be produced the same way every time. Splitting the roles works better: a field-based template in the business system produces the document, and a specialised agent does nothing but extract the variables from the text. Formal responsibility for the document stays with the accounting system, the only one able to defend it under a compliance review.
Checklist for preparing content before go-live
- Split the content inventory into two columns: stable information and information that changes during the day.
- For every volatile item, name the system that holds it and whoever authorises read access.
- Export documents in a structured format and avoid PDFs laid out for print.
- Look for contradictions between departmental procedures before loading, not afterwards.
- Define which topics stay outside the perimeter and on which channel a conversation passes to a person.
- Appoint an internal content owner, with the same logic applied to document management.
Companies skip that last point more often than any other. A knowledge base for AI agents with no internal owner ages within months, and the agent keeps answering confidently about conditions the company has already changed. Data stays in EU datacentres, with certifications in the ISO/IEC 27001 family and requirements aligned to NIS2 and to the General Data Protection Regulation: the same standards applied to the infrastructure of Aesir Srl, where we operate Tier IV datacentres with replication. The technical perimeter is defined; the content perimeter is defined by the client.
Questions from customer service teams
Does switching platform lose the work done on a knowledge base for AI agents?
No. Flows and content export, using the same procedure platforms follow when moving between major releases. What to verify when choosing is the export format and how readable it stays outside the product.
Does an agent created later inherit documents already loaded?
It inherits everything in the shared base, including documents loaded before it existed. Private documents belonging to other agents are not inherited and stay confined to the perimeter they were selected for.
How much work does content preparation take on the client side?
Initial analysis carries the weight, typically one or two days, defining perimeter, categories and sources. Loading itself is quick. Total configuration days vary considerably with the number of flows and integrations required.
Let’s talk
On the conversational automation projects we run at Aesir, technology takes up most of the discussion while something else decides the outcome: how clearly a company can describe itself in one voice. An agent does not create coherence where none existed. It brings to the surface, at the speed of a chat window, the divergences between departments that nobody had ever been forced to reconcile in writing. Anyone who designs a knowledge base for AI agents with that in mind starts months ahead, and the first agent goes live on a narrow perimeter rather than on everything at once.
If you would like to explore the subject or assess the situation in your own company, you can fill in the form at the bottom of this page or write to support@aesir-tech.it: we will arrange a free consultation and start from your numbers.