Module 2 — Conversation to Commitment · Lesson 2.1
The Capture Problem
Why most of the work that should happen never makes it onto a board
~11 min
What you'll learn
- Articulate the four main loss sites for commitments — meetings, chat, email, and hallway
- Diagnose your team's capture-loss rate from observable signals
- Explain why volunteer-based capture (note-takers, action items) systematically under-records
- Set the expectation Module 2 will then deliver against
Every team has two task lists. The one in the tool, and the longer one that exists only in the participants' heads as 'things I think I said I would do.' The gap between them is the capture problem, and it is by far the largest single source of dropped work in most organizations — bigger than missed estimates, bigger than reprioritization, bigger than turnover. Before we talk about how Kavanah closes the gap, we have to look at why it is so wide.
Where work is lost
Commitments arise in four surfaces. Each leaks differently.
Meetings. Synchronous conversation produces ambient commitments — 'I'll send you that doc,' 'we should look at the Q3 numbers,' 'let me set up a call with finance.' If there is a designated note-taker, perhaps half of these get captured. If there is not, perhaps a quarter do. The half that gets captured is biased toward the speakers most senior in the room.
Chat. Threaded chat tools are an enormous commitment surface and an even worse capture surface. The half-life of a chat message is hours; the commitment it contained gets lost when the thread scrolls off. The most reliable signal of a missed chat commitment is the next-day question 'wait, did anyone follow up on the thing from yesterday?'
Email. Email at least produces a written record. The problem is that the record is fragmented across inboxes and the commitment is usually phrased politely enough ('happy to take a look at this when I get a chance') that no system can recognize it as a commitment without help.
Hallway. The most dangerous category, because the commitment is real but no record exists at all. Hallway commitments are what makes the team's tracking system feel persistently incomplete, especially as the team scales beyond a handful of co-located people.
Why human capture systematically under-records
The standard organizational fix for capture is to assign someone to capture. A note-taker. An action-items column. A weekly review where everyone reports what they committed to. These work — partially. Their failure mode is structural.
Humans capture what their attention identifies as commitment-shaped. The shape of a commitment is a verb plus an actor plus a time horizon: 'I will send the deck by Friday.' Most real commitments in fast conversation are subtler than this: a question that implies follow-up, an objection that requires response, an implicit ownership change. The note-taker, focused on the next sentence, does not catch them.
Humans also under-capture commitments made by people other than themselves, because writing down 'Marcus said he would do X' feels presumptuous in the moment. The result is that the most senior person in the room has the lowest capture rate, because everyone is more reluctant to commit them on paper than to commit themselves.
Finally, humans get tired. By the third meeting of the day the note-taker's attention budget is exhausted; by the fifth, they have stopped capturing entirely. The capture rate falls through the day, every day, in a way that nobody notices because it is invisible. Whatever is missed never enters the system to be measured.
What shifts when the system listens by default
An AI agent with access to the team's conversation surfaces does not get tired, does not get embarrassed, does not have an attention budget that runs out by 3 p.m., and does not care about whether the speaker was senior or junior. Its capture rate is, to a first approximation, uniform across speakers and across the day.
This is not the same as saying the agent's judgment is perfect. It will surface ambiguous commitments as candidates for review. It will sometimes mistake rhetorical 'I will' for an actual one. The point is that the failure modes are different and inspectable. A human note-taker's misses are invisible. An agent's misses are auditable, because the conversation log still exists.
The shift, then, is from a system where capture is a person's job (and therefore patchy and biased) to one where capture is a default behavior of the team's communication surfaces (and therefore comprehensive, biased only by the model's priors, and inspectable). The manager's job changes accordingly. Instead of trying to extract commitments from the team in retrospect, the manager spends their time triaging a steady stream of agent-proposed tasks and deciding which become real.
What you are about to build
The next three lessons walk the three sides of this triangle. First, how Kavanah turns conversations into proposed tasks — what it reads, what it ignores, how it identifies a commitment. Second, how the proposed tasks get triaged into real ones, with the right owner, scope, and KVN. Third, how external conversations — public request pages, client portal asks — flow into the same pipeline so that work never has to be re-typed by a human.
By the end of the module the workspace should be in a state where the question 'did anyone follow up on the thing from yesterday' is no longer one anyone asks, because the answer is always 'yes, the agent surfaced it as a task; here is who has it.'
Measure your current capture loss
- 1
Pick a recent meeting and reconstruct its commitments
List, from memory, every commitment made in a meeting from this week. Then check the board. Note the gap.
- 2
Skim yesterday's main chat thread
Count the commitments that were made and which ones became tasks. The ratio is your chat-capture rate.
- 3
Look at your inbox's 'I owe you' debt
How many emails in your inbox represent a commitment you made but did not capture as a task? That count is your email-capture loss.
The capture-problem dashboard
- Capture coverage
- Fraction of commitments made in tracked conversation surfaces that become tracked tasks within 24 hours.
- Healthy signal: Above 90%. Below 70% is the 'did anyone follow up' zone.
- Capture latency
- Median time between a commitment being made and it appearing as a task in the tool.
- Healthy signal: Under 24 hours. Under 1 hour with an AI capture loop running.
- Capture-loss surface mix
- Of the missed commitments, what fraction are from meetings vs. chat vs. email vs. hallway?
- Healthy signal: Useful as a diagnostic — points you at which integration to connect next.
Key takeaways
- ·Most dropped work is dropped at capture, not at execution.
- ·Commitments leak from four surfaces — meetings, chat, email, hallway — each in a distinct way.
- ·Human capture is systematically biased: senior speakers under-recorded, late-day meetings under-attended.
- ·An AI agent with access to conversation surfaces shifts capture from a person's job to a default of the system.
- ·The shift makes the manager's job triage, not retrospective extraction.
Naming the problem is half of it. The next lesson goes deep on the actual data Kavanah reads — chat threads, email, AI agent conversations — and how it composes them into a continuous work-graph from which tasks can be proposed.