Why grounded knowledge and governed agents matter more than another chatbot

Over the past year, millions of dollars in enterprise AI licenses turned into shelfware, or expensive tools to write emails other people’s AI will read.

Teams experimented with the prompt box for a couple of weeks, got generic advice that didn’t match their daily reality, and closed the tab. Or worse, followed advice from the web instead of your company policies. 

The issue isn’t intelligence; it’s the context. Without your content, playbooks, and SOPs, an AI model is just an eager intern who skipped onboarding… fast, confident, and guessing wildly. Ask it how to handle a disputed refund and you’ll get an answer. Just not always your answer.

Enterprise AI only works when it knows what your company knows. That’s why we built Docebo Knowledge and Docebo Agents.

Knowledge gives your learners trusted answers grounded in your own content, with the permissions you set. Agents take those answers and do the work that follows, from building lessons to following up after training. Both now live in Docebo.

The intern doesn’t know where anything is

Every company runs on knowledge that lives in many places. Policies sit in SharePoint, procedures in Confluence, SOWs in the shared drive, and product updates in a Slack thread from last Tuesday. In our world, there’s also a lot of training content.

Take a guess how much of the content across Docebo systems is in the format of ‘standard’ files – like documents, videos, or PDFs.

About 25%.

The other 75% is SCORM and xAPI packages, some of them 20 years old. Most knowledge tools are like document readers. If the answer lives in part three of a ten-part program on operating a nuclear reactor, built as a SCORM package a decade ago, a standard KM system can’t read it. So the intern (your AI) guesses.

The rest of the content has its own problems. Connecting SharePoint is the easy part. Deciding which policies are current, which documents are authoritative, and which old drafts should stay out of an answer is the challenge.

Then there’s access. AI doesn’t care about access rights unless it’s already perfect in the host-system. Does your company have perfect access setups across all your platforms? How about direct control of which Confluence space or Google Drive is being searched? We think knowledge managers (you) should have that power to filter. 

“Mostly right” doesn’t work at work

AI is probabilistic. Most enterprise software exists because organizations need answers that are deterministic and auditable. You can’t mostly, sort of, kind of know that everyone completed their KYC training. If you’re a major bank, one hallucination is millions of dollars in fines.

The bar isn’t an answer that sounds right. It’s the right answer for this person, in this place, with their training.

Jim, working on a warehouse floor, badly cuts himself. Sally pulls out her phone and types, “Jim just opened up his finger, I can’t find my manager, what do I do?”

A generic tool gives her a first-aid summary from the internet. Our system knows if I’m certified in first aid, where I’m asking from, and the procedure that I should follow according to my SCORM training material.

What we actually built

Knowledge comes first. Docebo Knowledge connects the content in Docebo, including those SCORM and xAPI packages, with sources like SharePoint, Confluence, Google Drive, and Slack. It inherits the permissions from each source system, so people only see what they’re already allowed to see.

And it doesn’t stop at the answer. It points people to the course, certification, or next step that follows. It cites where the information came from so people can check against it. 

Then Agents. If Knowledge is the memory, Docebo Agents are the follow-through.

And they know who they’re working for. Every agent works from the learning data already in Docebo, so the Tutor Agent knows what you’ve completed and the Reinforcement Agent knows what you got wrong last time.

We launched four pre-built agents (and custom agent building) on day one, with more to come. 

  • The Authoring Agent builds the lesson. Give it a document or a SharePoint folder and you get a lesson with interactions and assessments already built in. You can then iterate with the agent, have it create code for you, and build incredible learning experiences – fast. 
  • The Tutor Agent sits inside a course and can be limited to the information you choose. When a learner gets stuck, it helps them work through the question instead of simply handing them an answer.
  • The Reinforcement Agent picks up after the course ends, asks follow up questions, and battles the forgetting curve — because finishing a course and remembering are two different things.
  • The Marketplace Agent helps admins find relevant off-the-shelf courses in our catalog. They can describe what their learners need and get help narrowing the options, rather than search through thousands of packages themselves.

When those don’t cover it, admins can build their own agents in AgentHub without writing code. Every agent runs inside boundaries you set: which agents run, who can use them, and which actions need a person to approve them first. Every action is logged. Agents run on AI credits, and admins can deeply control where they’re using them with our AI Credits Manager, so nobody gets a surprise bill.

We built it with you

More than 50 customers piloted Agents before launch. One transportation company finished an estimated six months of course creation in five weeks. A quick-service restaurant operator saves store managers about 10 hours a month on training follow-up. A retail company that onboards 2,500 new hires a year built its own onboarding agent in AgentHub.

Proximity beats capability

In an average 1,000-person company, maybe 400 people log in to training in a given month. That’s been the baseline for years, long before AI. Training is a quarterly behavior. Knowledge is a daily one.

An AI tool earns its place when someone uses it to solve a real problem and comes back because it helped. That’s the test we’re setting for Knowledge and Agents.

If you’re using them, tell us where they’ve saved you time and where you’re still doing the work yourself. Tag us on social, post in the Docebo Community, or send us a DM. That’s what this intern needs to learn next.