profile

People-first AI

How to make sharing context with AI safe


Welcome back. Nvidia and Anthropic stole the AI headlines this week, but a passing conversation with a colleague grabbed my attention more. It went something like this: "We should find ways to agentify what we do." "But aren't we training our replacement?" It's less exciting than spies and IPOs, but it's a conversation that many workers are having.

In this issue, we talk about the psychological resistance to giving AI the context it needs to perform at its best, and how we can help manage that resistance. Enjoy!

--Neil

AI @ WORK BRIEFING

I have a confession: I’ve started a lot of AI training, but I've only finished a few.

To help, Workplace Labs built AI Foundations Daily Experiments. It sends you one short experiment each morning, written for your job, that you can try with your real work. Ten minutes and no log-ins.

It’s out now and free for individuals.

BIG PICTURE

How to make sharing context with AI safe

What’s one of the main things AI-native companies do differently? According to Nathaniel Whittemore of the AI Daily Brief, it’s “giving agents the whole picture.”

He gives several examples of how the companies pushing agentic work furthest are building shared environments where context can accumulate.

  • ​Uber created an AI Context Graph connecting more than 30 internal systems. In one comparison, an agent using the graph answered a question correctly in 38 seconds; without it, the agent spent 20 minutes searching and reached the wrong conclusion.
  • ​Shopify stores skills, runbooks, intent documents, and conventions directly inside its monorepo. Because its River agent operates in public Slack channels, when one person solves a problem, that solution becomes a starting point for the next person.
  • ​Ramp’s Glass scan apps and stores personal memory, team guidance, and company context that update the company’s shared context.

Together, these systems turn personal memory into organizational infrastructure. They help agents find the right sources, apply the right exceptions, and avoid producing outdated guidance more efficiently.

But every one of these systems runs on a psychological bargain. Agents need employees’ context, and employees may have rational reasons to keep it.

Most leaders will file that problem under culture. They’ll promote knowledge sharing, buy better documentation tools, celebrate contributors, and expect the context to flow.

More than a decade of research on knowledge hiding says otherwise.

Hiding is not the absence of sharing

Back in 2012, Canadian researchers Catherine Connelly and her colleagues defined knowledge hiding as an intentional attempt to withhold or conceal knowledge that someone else has requested, and identified three forms:

  • Evasive hiding: providing incomplete or misleading information, perhaps while promising to share more later.
  • Rationalized hiding: offering a justification for not sharing or attributing the restriction to someone else.
  • Playing dumb: pretending not to possess the requested knowledge.

These behaviors already have recognizable AI-era equivalents.

Evasive hiding sounds like, “Just pull what you need from the deck.” Rationalized hiding becomes, “Legal won’t let us put that into the system.” Playing dumb becomes, “I wouldn’t even know how to write that down.”

Sometimes those answers are true, but knowledge hiding is intentionally withholding information. With agents, it could be downplaying what you know about using AI or withholding process and tacit knowledge from agents themselves.

Poor context may reflect a skills problem, a policy constraint, or a deliberate attempt to protect something valuable. But when it’s intentional, a 2023 meta-analysis covering 131 studies and nearly 50,000 participants found that knowledge hiding and knowledge sharing are different things with different causes. Hiding was associated with job insecurity, distrust, and territoriality. Psychological safety and supportive leadership were associated with less of it.

In other words, hiding is not simply what remains when an organization fails to promote enough sharing. The same action that increases sharing may do little to reduce hiding.

A new study makes this distinction especially relevant to AI adoption. In a survey of 428 knowledge workers in China, Liu, Gao, and Yu found that generative AI use predicted more knowledge sharing through increased self-efficacy and more knowledge hiding through heightened job insecurity. A competitive climate strengthened both pathways.

One technology produced two opposing behaviors.

Consider how this could play out in a sales organization. AI-assisted dashboards and automated outreach may help sellers identify and pursue leads, increasing their confidence and their willingness to share useful techniques. The same tools may give employees the impression that the company will need fewer people to produce the same pipeline. Add an internal climate in which sellers compete for recognition, accounts, and advancement, and the organization has the right conditions for both knowledge sharing and knowledge hiding.

More training might help someone explain a sales process to an agent. It won’t necessarily make them want to.

Context is part of the employment deal

The psychological contract helps explain why.

In 1989, Denise Rousseau described psychological contracts as employees’ beliefs about the two-way obligations between themselves and their organizations. These obligations are often unwritten. They reflect what people believe they owe the organization and what the organization owes them in return.

Asking an employee to document their expertise for an agent is not a neutral request within that relationship. The organization is asking the employee to give up something that may have made them valuable: an understanding of which customer signals matter, which exceptions require escalation, which numbers can be trusted, or whom to call when the official process breaks.

The implied exchange might be relational: Help us build this capability, and we will use it to make your work better, invest in your growth, and give you a meaningful role in what comes next.

Or employees may hear a different offer: Teach the system what you know so we can determine whether we still need you.

If they expect the second deal, withholding context is a predictable form of self-protection. For employees, the possibility of breaking the psychological contract creates feelings of mistrust.

Paying people for each document they contribute may make things worse. In a field study of 154 managers from 27 Korean organizations, Bock and colleagues found that the more employees anticipated tangible rewards (pay or bonuses) for sharing knowledge, the less positive employees' attitudes were toward knowledge sharing.

Once context has a price, employees have reason to weigh what they’ll provide for that price against expertise they believe is worth more.

A better approach is to treat increasing knowledge sharing and reducing knowledge hiding as separate jobs.

To reduce hiding, leaders can say plainly which tasks will remain human-led and how productivity gains will be used. They don’t need to eliminate internal competition, but they can change what people compete on. Sellers can still compete on outcomes while shared playbooks, prompts, and techniques are treated as team assets rather than personal edges. Leaders can also build confidence in senior leadership’s intentions, not only trust in an employee’s direct manager.

To increase sharing, organizations can build employees’ confidence through progressively harder AI-assisted work. They can preserve credit and authorship by attaching contributors’ names to the skills or playbooks they create and showing how those assets are used. Most importantly, they can give contributors an ongoing role in maintaining the knowledge they shared.

That last step changes the deal. The employee is no longer being asked to pour their expertise into a static repository and step away. They become the steward of a living organizational capability.

What this means for HR leaders

1. Diagnose before you intervene. When an agent receives thin context, determine whether the employee can’t provide more or won’t. The first problem may require workflow analysis, training, or better tools. The second requires attention to job insecurity, trust, competition, and the employment deal. A culture campaign aimed at both will blur the distinction and probably solve neither.

2. Design the deal. Don’t pay for it. Make the psychological contract surrounding agentic work as explicit as possible. Which tasks will remain human-led? How will additional capacity be reinvested? What opportunities will contributors receive? How will they retain credit for what they create? Documentation bounties are easier to launch than a credible answer to those questions, but a credible answer reduces the threat.

3. Staff the maintenance. Context upkeep and exception handling need to appear in roles, workloads, and performance expectations. Otherwise, it becomes invisible work: feeding the agent, supervising its output, correcting errors, and watching knowledge decay as the business changes. A shared context environment is temporary without accountable maintainers.

4. Make the cost of stale context concrete. An agent working from incomplete or outdated guidance can produce confident, incorrect work faster. The tokens, failed runs, review time, rework, and downstream errors are measurable. So is the difference between an agent grounded in the organization’s actual knowledge and one left to search for clues.

The people most likely to withhold context may also be the people best equipped to keep it alive. The question is whether the organization gives them a deal worth accepting.

6065 Roswell Road, #450, Atlanta, GA 30328-4011

Say hi on LinkedIn!
​Unsubscribe · Preferences​

People-first AI

In under 10 minutes, you’ll have the context and language you need to lead smarter AI conversations at work—and stand out as the go-to guide on your team.

Share this page