Editorial
Research Report
The State of the Digital Workplace
If your AI agent projects are failing the problem often isn’t the AI; it's the content you are feeding it. Digital workplace teams have the skills to fix this, but they need to reframe familiar disciplines as risk management, not housekeeping.
Two Kinds of Questions
A seemingly simple distinction has significant consequences once you take it seriously.
Some employee questions have a specific, documentable answer. "How do I book international travel?" has a policy, a booking tool and an approval chain. The answer exists. The task is anchored against something knowable. Whether AI can retrieve that answer reliably depends almost entirely on whether the content is any good.
Other questions do not have a documented answer and may require human judgement being applied to the specific situation. "Who is accountable for data governance in our Greek operations?" is not a question a content library can resolve. We can't document every aspect of our businesses before going bust. In any event, the answer might only become apparent in the moment someone sits down and works it out. The task is unanchored: there is no fixed target to retrieve.
AI cannot tell the difference. It approaches both with equal confidence and produces a fluent, well-structured response in either case. That is potentially useful for anchored tasks. But for unanchored ones, it is a hallucination-as-a-service and employees reading the response have no way to know which kind of answer they have just received.
Unanchored tasks are far more common than organizations expect. First you need to identify them, and then you need to tell the agent not to answer that sort of question. Identifying your anchored tasks and your unanchored tasks is one of the more useful things you can do before expanding AI retrieval.
What Makes Content Dangerous Under AI
Anchored tasks can still go badly wrong. The content they depend on is frequently unfit for the purpose at hand.
Outdated content is the obvious failure mode, but not the most dangerous one. When rival documents cover the same process with no clear indication of which is authoritative, AI synthesizes a response that is neither clearly right nor clearly wrong — but is presented confidently. Geographic and business unit variation that has never been explicitly disambiguated produces the same problem at scale: a global organization with 14 different expense processes gets an answer that is correct for one business unit and wrong for everyone else, with nothing to signal the difference.
Formatting is easy to overlook, with PDFs of PowerPoint slides being the worst offender. The ingestion process that feeds these systems flattens content into plain text, destroying the spatial relationships that give slides their meaning. Boxes and arrows become a sequence of nonsensical fragments. The content was never designed to be read this way, and AI cannot reconstruct what it cannot see.
The One Lever You Control
Outside of agent-level prompts, organizations have limited control over the model, the pipeline or how answers are constructed. Those decisions sit with the tech giants, not the digital workplace team.
What does sit with that team is the content: what goes into the system, in what form, with what structure, from what authoritative source.
The boundary between "posted," "saved" and "published" is a fundamental change of purpose — content crafted deliberately, for a specific audience, to a defined standard.
Content governance, content design, clear ownership, publishing standards and publisher training and coaching have long been treated as nice-to-haves. AI reframes these disciplines into the primary controls to make AI retrieval trustworthy. The digital workplace field has been advocating for them for years. The AI deployment timeline is rarely calibrated to the content improvement timeline, but the lever is the same.
What Good Looks Like
The goal is not a catalogue of perfect information. It is an information estate that is trustworthy in the areas that matter most, with a clear and maintained boundary between what is in scope for AI retrieval and what is not.
In practice, this means identifying the tasks where a wrong answer has real operational, legal or safety consequences. For most organizations, this points to HR processes, compliance and regulatory guidance, and anything that varies across geographies or business units. Understanding this risk surface and placing it under management is an essential tool of control.
Good in this case looks like: a single authoritative source for each answer; clear ownership with a named review cycle; language that is explicit rather than hedged; structure that survives ingestion. It also means routing employees to a human when the question has no reliable documented answer. Escalation to a human doesn’t mean the AI implementation failed. It’s a sign that it has been designed around the ambiguity and complexity of how organizations really work.
What Comes Next
In the third post: I make the case for Provenance as a named discipline beyond what has been seen as "AI governance" on a per system basis, why no existing function can own this problem on its own, and what a coordinated approach to the epistemic layer of the digital workplace looks like.
Conference
Apr
22
HR Tech Europe Amsterdam 2026
Register
Conference
Apr
27
Gartner Digital Workplace Summit London 2026
Register
Conference
Jun
2
Gartner Application Innovation & Business Solutions Summit Las Vegas 2026
Register
Webinar
On demand
Do Learning Programs Really Work? How to Turn Education Into Engagement In Healthcare
See how leaders are using learning programs to build trust with healthcare professionals and create measurable engagement.
Watch Now
Editor's Note: Catch up on the first installment of this three-part series and other takes on readying your digital estate for AI:
- AI Is Only as Good as Your Knowledge – And That's a People Problem — The organizations getting the most from AI aren't just buying better tools. They're doing the unglamorous work of fixing their knowledge first.
- Connected Intelligence Can't Fix Your Broken Knowledge Management — AI-powered knowledge management is the latest tech promising to fix what thirty years of tools couldn't. But what if the technology was never the problem?
- Context Is the New AI Infrastructure — AI agents can access data, but not your decision-making context. Context graphs capture how changes happen. Here's how to get started building your own.
Learn how you can join our contributor community.