Beliefs
AI Moves Faster. Your Team Still Has to Agree.
The agent can build the feature overnight. Your team still has to agree on what the feature should do.
Anthropic's October 2026 workforce study reports rising meeting time alongside extensive AI adoption. People are still getting together to work things out.
Our read at Earmark: as execution gets faster, the quality of the decisions directing it matters more. Teams need a dependable way to carry those decisions into the work that follows.
What Anthropic found
20% more meeting time
The median Anthropic employee spends 20% more time in meetings than a year earlier, based on calendar data.
16% more for technical individual contributors
The increase extends to technical ICs, whose meeting time rose 16% over the past year.
Source: Anthropic, October 16, 2026.
Company growth and changing roles explain part of the increase. This single-company study does not establish that AI caused it. We should question the assumption that adopting AI automatically clears the calendar.
Faster execution makes misunderstood decisions more expensive
Consider a product team discussing a customer export feature. The conversation settles several details: administrators get access first, sensitive fields stay excluded, and a broader rollout depends on a security review.
A ticket that says “build CSV export” misses most of the decision. A coding agent could implement that ticket perfectly and still deliver something the team cannot release.
This is a hypothetical, but the failure is familiar. The feature name survives the handoff. The constraints disappear. Someone discovers the mismatch during review, and the team pays for another round of explanation and rework.
As implementation accelerates, a vague instruction can become a substantial amount of finished work before anyone notices. The practical response is to improve how teams preserve scope, tradeoffs, and unresolved questions at the moment they discuss them.
The faster your agents execute, the more expensive it becomes to lose what your team actually decided.
The meeting is where instructions change
A planning discussion can change a deadline, narrow a requirement, reject an approach, or reveal a dependency. Those changes may alter the work already underway.
Anthropic describes agents missing context when an in-person decision never reaches a written record. The report puts the underlying requirement plainly:
“Keeping human work visible across the company ensures autonomous agents reason from fresh context.”
— Anthropic, How employment is changing at Anthropic
For product and engineering teams, this suggests a useful way to think about meetings: they are one of the places where the instructions for execution get updated.
That makes the handoff consequential. An earlier requirement may still look authoritative after the team has changed its mind. An exploratory idea may look like approved scope once it appears in a polished summary. A proposed owner may look like someone who accepted the assignment.
A useful decision record preserves those distinctions. It gives the next person—or agent—enough context to proceed and enough clarity to know when to ask.
Give the next step the context it needs
For a recurring planning meeting, we would start with five questions:
Decision: What did the team explicitly agree to?
Reasoning: Which constraints and tradeoffs shaped that choice?
Change: Which earlier requirement or assumption does this replace?
Ownership: Who accepted responsibility for the next step?
Uncertainty: What remains open, and what must be resolved before execution?
Then turn the answers into the work the team needs: a revised requirement, a ticket with acceptance criteria, a decision log, or a brief for a coding agent. Review it while the people who made the decision can still correct it.
For the export example, the handoff would specify administrator access, excluded fields, and the security review dependency. The wider rollout would remain a conditional next step. That is a much more useful starting point than a feature name followed by a transcript.
Protect attention as well as context
Anthropic also estimates that reading all the interactive agent output generated for non-technical groups would consume more than half a workweek. That is a modeled reading burden, rather than measured time spent reading.
Meeting tools should take that attention constraint seriously. The useful output is the smallest clear account that lets someone act: the decision, the change, the owner, and the question that still needs an answer. Keep supporting evidence available for review.
Access matters, too. A conversation may contain information that belongs within a specific team or project. Making its decisions useful requires deliberate choices about what gets shared and with whom. A need for context does not justify unrestricted access to every conversation.
What this means for Earmark
We build Earmark to bring conversation and usable work closer together. Live drafts help teams examine what they are agreeing to. Project memory helps them revisit the reasoning across related discussions. A reviewed specification or agent brief helps carry that reasoning into execution.
Our ambition is to make that handoff dependable: preserve the team's judgment, expose what is unresolved, and reduce the manual reconstruction between a conversation and the next step.
The test is practical. How long does it take to turn a discussion into reviewed work? How many clarification loops follow? How much rework comes from an outdated requirement or a missing constraint? Those are better signals of value than the number of summaries generated.
Start with one meeting that regularly produces requirements or decisions. Review the output together, carry it into the actual workflow, and see whether the next person has what they need.
Your agents can move faster than ever. Give them the decisions your team actually made.
Put your next conversation to work.
Use Earmark to prepare the spec, ticket drafts, or follow-up while the conversation is fresh. Review it, then carry it into your team's workflow.