Your Team Doesn't Need More Copilots. It Needs Shared Context.
AI can make every person on a team faster and still leave the team with five versions of reality.
One of the most useful lessons I learned from building multi-agent workflows came from a failure.
An upstream AI would finish its work and hand the next AI a reference to what it had done.
Technically, the handoff existed. Operationally, it was useless.
The next agent would come back asking some version of: What's my list? Where is the source? Which records am I supposed to use?
So the rule changed.
If another worker needs the list, the list travels with the assignment. If it needs the criteria, those travel too. If something is unresolved, the uncertainty travels with it.
Signal 013 reduced that lesson to one sentence:
A handoff is not a pointer. A handoff is a complete input package.
Previously: Signal 013
I initially thought that was an agent lesson. I increasingly think it is a team lesson.
Five productive people can still create five versions of reality
Most workplace AI has been introduced through a personal productivity model.
Give me an AI. Give you an AI. Give marketing an AI. Give finance an AI. Give the manager an AI.
Everyone gets faster.
There is an obvious upside. There is also a failure mode that becomes more serious as the assistants become more capable.
My AI knows what I discussed. Yours knows what you discussed. Mine has Tuesday's customer context. Yours has Monday's pricing assumption. Another assistant has the finished presentation but not the reasoning that changed the recommendation.
The CRM has a precise-looking field that is already stale.
Everyone can be highly productive while the organization becomes increasingly incoherent.
I have encountered the smaller version of this directly. A CRM record can be useful without being current. The most important relationship context may live in a note. A deal stage can look precise while the account remains ambiguous.
That led to another working rule:
The CRM is evidence, not truth.
The broader problem is the same. Important state has to survive movement between people and systems.
AI can bridge expertise
A field experiment at Procter & Gamble gives us a useful look at what happens when AI enters real collaborative work.
Researchers randomly assigned 791 professionals working on product innovation problems to work individually or in two-person teams, with or without generative AI.
Individuals using AI matched the performance of teams working without it.
But another finding is more interesting for this argument.
Without AI, R&D professionals tended to produce more technical solutions, while commercial professionals leaned toward commercially oriented solutions. With AI, the outputs became more balanced across those dimensions regardless of the participant's professional background.
Source: Dell'Acqua et al., Organization Science
AI was doing more than making someone faster.
It appeared to help individuals reach across an expertise boundary that normally separated functions.
That is a potentially important form of organizational value. But it does not mean that adding more AI automatically improves teamwork.
Same AI. Different process. Opposite result.
Microsoft researchers ran a 2026 field experiment with 388 employees at Gap Inc.
Everyone had access to the same AI.
What changed was the way people were instructed to work with it.
One intervention required pairs to follow a structured protocol for using AI together. Instead of improving collaboration, that condition was associated with worse document quality and substantially less document production than more naturalistic use.
Most strikingly, 26.8% of treatment pairs failed to produce a document at all, compared with 4.1% of controls.
Source: Scaffolding Human-AI Collaboration
The experiment has important limitations, including a time-of-day confound, differential attrition and sensitivity to AI-based grading. The authors are explicit about those limitations.
There was also a useful counterweight.
A separate intervention trained individuals to approach AI more like a thought partner than a search engine. Participants in that condition were more likely to receive a perfect document score, 77.0% versus 61.8%, although that result was sensitive to the study's attrition assumptions.
The lesson is not that structure is bad.
Process design matters enough to hurt you when you get it wrong.
“Our team has AI” tells me almost nothing about whether the team has learned how to work with it.
What shared AI starts to look like
A study published in September in Computer Supported Cooperative Work offers a more grounded picture.
Researchers studied an approximately 200-person IT-services company using a customized LLM integrated into Slack and connected to systems including Salesforce, Jira, Confluence, NetSuite and Tableau. The system also recorded, transcribed and summarized meetings.
This was a qualitative study in one organization, not evidence that every company should copy the design.
What makes it useful is what employees started doing around the system.
Meetings became persistent artifacts that could be searched later. People used prior conversations to reconstruct customer context and prepare for subsequent work. Some participants said they could pay more attention to the client instead of frantically taking notes.
Workers also began changing how they spoke because they knew their words would later become machine-readable organizational memory. They would make context more explicit, say ticket numbers aloud or clarify points for the recording.
At that point, the AI is no longer just an assistant. It is participating in the team's memory system.
That is a much bigger design decision.
Shared context can also spread shared errors
There is a danger here.
The same study documents cases where AI-generated sentiment or summaries could flatten the nuance of what happened.
If a private assistant gets something wrong, one person may carry the error.
If the shared organizational memory gets something wrong, the error can propagate.
That is why shared context cannot mean unquestioned context.
Teams need to preserve the difference between a verified fact, a working interpretation and an AI-generated summary.
Academic work on human-AI teaming arrives at a similar conclusion. A 2026 PNAS Nexus framework argues that complementarity between people and AI depends on factors such as shared mental models, role design, coordinated attention, training and knowledge infrastructures.
Source: Gonzalez et al., PNAS Nexus
Simply placing a human and an AI beside one another does not guarantee a better team.
In ordinary language, I think the questions are simpler:
- Does the next person know what we know?
- Do they know what remains uncertain?
- Can they continue the work without reconstructing the whole project?
The next bottleneck may be synchronization
The first era of workplace AI made intelligence personal.
My chat. My assistant. My prompts. My context.
That was probably necessary. It made experimentation cheap.
But companies do not create value as collections of isolated intelligent individuals.
Work moves. Decisions move. Customer context moves. Responsibility moves.
And every handoff creates an opportunity for meaning to disappear.
The small failure that taught me the handoff rule was just one AI passing work to another. Once the inputs traveled with the assignment, the workflow worked again.
Teams face the same problem at larger scale.
The next competitive advantage may not come from giving every employee the best copilot. It may come from getting unusually good at keeping important context intact as work moves between people and machines.
Five productive people with five versions of reality are not a high-performing team.
They are five fast people with a synchronization problem.
Signal follows the shifts that actually change how work gets done.
Subscribe to Signal