A trusted knowledge layer stopped being a search product this week. On October 28, Docebo launched Knowledge and Agents, a layer that indexes company documents and then adds a second kind of record that search never held: what each person knows, what they are learning, and where their skills gaps sit. The launch matters because the research behind it points the same way. Only 34% of agentic AI projects reach production on average, and the firms that do reach production hold stronger knowledge, not larger models.
What Docebo actually shipped
Docebo, listed in Toronto and on Nasdaq, said Knowledge turns scattered enterprise information into a trusted knowledge layer connected to learning, in its October 28 release. Learners can search training content and connected systems in one place. The layer indexes content from Docebo, including SCORM and xAPI packages, alongside approved sources such as SharePoint, Confluence, Google Drive, and Slack. Answers arrive with source links, and a reader can move from an answer to the course, certification, practice, or workflow that follows.
The second half of the launch is the shift. Docebo said most agents can reach historical company data, but few know a user's learning history and skills gaps, or the organization's skills taxonomies. Its agents run on that skills and learning data alongside organizational knowledge. Four prebuilt agents shipped at launch: an Authoring Agent that turns source material into structured courses, a Tutor Agent that answers questions inside a course from course material and approved knowledge, a Reinforcement Agent that runs spaced practice after a course ends, and a Marketplace Agent that compares third party courses for admins.
The chief executive put the claim plainly. "Every organization is about to run two workforces: its people and the agents working alongside them," said Alessio Artuffo, President and Chief Executive Officer of Docebo, in the release. "Both need the same things to perform well: trusted knowledge and an accurate picture of skills. That's the foundation Docebo is now built on."
Controls shipped with the agents, not after them. Knowledge inherits the permissions set in each connected source system, so learners and their agents see only content they are authorized to access. Admins decide which agents run, who can use them, and which actions need human approval before they happen. Every action is logged, giving security and compliance teams an audit trail, according to the release. The platform also connects through Model Context Protocol to Microsoft Copilot, ChatGPT, and Claude, so learning and knowledge can surface inside assistants employees already use.
The pilot results are vendor reported and should be read that way. Docebo said more than 50 customers piloted Agents before launch. A global restaurant operator used the Tutor Agent to resolve 25% of frontline training questions without an administrator, and the Reinforcement Agent to save store managers about 10 hours a month on training follow up. An electronics distributor cut roughly five hours a month of manual curation for each administrator. A retail company that onboards 2,500 new hires a year built a custom onboarding agent which the company expects will give sales managers one to two hours back per new hire. That last figure is an expectation, not a measured result, and the release gives no error rates for any of the pilots.
The evidence: knowledge, not models, is the production gap
The strongest survey evidence this month comes with a caveat that belongs in the text. MIT Technology Review Insights published Connecting AI Agents to Enterprise Knowledge on October 5. The report was produced by the publication's custom content arm, not its editorial staff, in partnership with Neo4j, and it surveyed 300 data, AI, and other technology executives. It is sponsored research. Its figures describe that respondent base, and they show association, not causation.
Within those limits, the pattern is consistent. On average, only around a third, 34%, of organizations' agentic AI projects make it into production. A small group the report calls production leaders, where an average of 61% of projects advance beyond pilot, have stronger knowledge capabilities than the rest, especially in semantics. Data fragmentation, the inadequate sharing of data across systems, was the most commonly cited top challenge to expanding agents' access to knowledge, at 55%. Production leaders, by contrast, were more likely to name security and privacy as a major concern, at 72% of that group.
The report's useful contribution is its definition. It gauges knowledge in three kinds: semantic knowledge, what data means inside one organization; episodic memory, what has happened; and procedural knowledge, how work is done. Executives in the survey expect the biggest impact to come from strengthening the structural foundation between the organization's data and its agents, and the experts interviewed for the report see a knowledge layer as a prime way to do that. Planned investment follows the same line: ingestion pipelines, AI ready APIs, and retrieval augmented generation, plus AI evaluation agents and knowledge graphs.
Read against Docebo, the survey explains the launch and also exposes what the launch adds. Semantic, episodic, and procedural knowledge are all records about the work. Skills data is a record about the worker. No category in the MIT framework covers it. Docebo has added a fourth kind of knowledge to the layer, and the fourth kind is the one that changes fastest, is most often inferred rather than stated, and is about a person.
The skills record is the weakest record in the building
A second October study shows why that fourth record cannot be taken on trust. edX Enterprise published The State of Enterprise Learning 2026 on October 8. Research firm NewtonX surveyed 507 verified learning and workforce strategy leaders at organizations with 500 or more employees, fielded from June to July 2026, with 10 in depth interviews. AI capabilities are an active workforce priority for 83% of respondents, but only 29% say their companies have built mature AI capabilities. Nearly four in five employers, 78%, lack a strong mandate to lead AI upskilling, and just 26% are highly confident their learning function can deliver the capabilities the workforce needs.
The study identifies AI transformation leaders as 15% of companies, those combining mature capability with high confidence in delivery. Its analysis found mandate strength accounted for 59% of the combined influence of four factors on leader status, while formal ownership of AI upskilling accounted for just 1%. Chris Kelly, Strategic Insights and Analytics Director at NewtonX and one of the study's authors, said: "Companies burn a lot of energy debating whether AI upskilling belongs to HR, L&D, or IT, but that decision alone builds nothing."
That finding cuts against a simple reading of the Docebo launch. Connecting knowledge to skills data does not create a mandate, a budget, or a measure of whether capability actually changed. It creates a record that an agent will act on. If the skills taxonomy is a year old, if a gap was inferred from a course a person clicked rather than work a person did, if a certification expired and the record did not, the agent personalizes on a false picture and does so at scale, politely, and with a citation to the record that is wrong. A document can be superseded and dated. A statement about a person's capability needs the same treatment, plus a person who can contest it.
A trusted knowledge layer must pass four tests, not one
The synthesis across the three sources is a test any knowledge layer can be held to, whether it comes from a learning vendor, a suite vendor, or a company brain. Call it the Four Records Test. A layer earns the word trusted only if each record it holds can answer the same four questions: what is the source, what is the date, which version did it supersede, and who approved it joining the record.
First, meaning. Semantic knowledge holds the company's own definitions, so the same term does not mean one thing in sales and another in finance. The MIT survey found semantics is exactly where production leaders separate from the rest. A definition without a date and an owner is a rumour with good formatting.
Second, history. Episodic knowledge holds what happened: decisions, incidents, and outcomes, in order. An agent that cannot tell last quarter's policy from this quarter's will answer from whichever text is closest, which is the failure mode examined in our post on evolving workplace records.
Third, method. Procedural knowledge holds how work is done now, not how it was done when the wiki page was written. Procedures change when a system, a supplier, or a regulation changes. The record needs the change, not just the current text.
Fourth, capability. Skills knowledge holds who knows what, what they are learning, and where the gaps are. This is Docebo's addition, and it is the record with the highest correction burden, because it describes people. It needs a stated source for every claim, a review route the person described can use, and an expiry when the evidence ages out. Without those, a skills layer is an inference engine about employees, and employees will treat it as one.
Across all four records, one condition governs: a correction must propagate. When a definition, a decision, a procedure, or a skills claim is corrected, the correction has to reach the answers and the queued agent actions built on the old version. Sharing a layer shares whatever it holds, a point set out in our post on shared AI context. A layer that spreads a correction slowly spreads the error first.
The counter argument: a wider layer widens what must be verified
The case against building the layer wider is not theoretical, and it arrived two days after the Docebo launch in reported form. SAPinsider covered Pathlock's AI Governance Gap Report on October 28. Pathlock surveyed 286 IT, compliance, and security decision makers in May 2026 at organizations that run or plan to run AI agents. In that survey, 38% allow agents to create or modify records such as vendors, customers, or materials, and 28% allow them to approve transactions or requests. Yet 52% cannot fully verify every action agents execute across business systems. Only 19% report complete, real time visibility into agent actions, and 48% cannot trace an agent's activity end to end across systems. Some 23% had an agent incident in the past 12 months that required investigation, remediation, or escalation.
Susan Stapleton, a governance, risk, and compliance expert at Pathlock, framed the shift in that report: "For decades, governance focused on controlling who could access a system. AI agents introduce a different challenge: understanding what actually happened after access was granted." Her colleague Chris Radkowski said that once AI becomes part of how transactions are executed, governance "stops being just an IT concern and becomes a matter of business risk, financial control, and compliance."
Apply that to a knowledge layer that now includes skills. The audit question is no longer only which document an agent read. It is which picture of a person the agent used, where that picture came from, and whether the person had ever seen it. Docebo's published controls answer the first half: permissions inherited from source systems, admin approval for defined actions, and a logged audit trail. The published release does not state how a skills claim is sourced, how a person contests one, or when one expires. Those are not reasons to reject the direction. They are the questions a buyer should ask before the layer, rather than after an agent has acted on it. The same discipline applies to any acting agent, as our post on operating knowledge set out this week.
There is also a limit in the evidence itself. The MIT figures are sponsored survey associations from 300 executives. The edX figures are a vendor commissioned survey of learning leaders. The Docebo pilot figures are the vendor's own, with one headline figure stated as an expectation. None of the three proves that adding skills data to a knowledge layer causes more projects to reach production. What they establish together is narrower and more useful: practitioners keep naming knowledge as the constraint, the knowledge they name is widening from documents to people, and the verification systems around agents are not yet keeping pace with either change.
What this means for the company record
The direction is settled even if the proof is not. Knowledge layers began as search over documents. They now hold definitions, decisions, procedures, and, from this week, claims about what each employee knows. Each step makes the layer more useful to an agent and more consequential when it is wrong.
That is why the record, not the retrieval, is the work. A company brain that employees can open and read, where each entry carries its source, its date, the version it replaced, and the person who approved it, gives an agent something it can be checked against and gives an employee something they can correct. How that record is built is set out in How Jilu Works, and the context discipline behind it is set out in our pillar on context engineering. The vendors racing to widen the layer have made the case for it. The next race is narrower: whose layer can show, for any fact an agent used, where it came from and who stood behind it. That is the layer worth trusting.