Useful work per dollar
OpenAI published a playbook this week called How to manage AI investments in the agentic era. It's written for the people who pay the bill: five steps covering usage dashboards, spend controls, governance, portfolio funding, and capacity tiers. The framing is good. Token price alone doesn't show whether AI is creating value, so measure useful work per dollar, and for priority workflows track the cost per accepted outcome.
Almost every recommendation in the piece comes with a product attached. Visibility is the Admin Console. Governance is ChatGPT Work. Capacity is Guaranteed Capacity and Scale Tier. That's fair enough, it's a vendor's playbook. But one sentence in step two has no product attached to it at all:
Clear instructions, focused tools, reusable context, and explicit stopping conditions can reduce loops and wasted spend.
One sentence, in passing, and it's the whole ballgame. Reusable context is the input no vendor can sell you, because it's yours.
What actually moves cost per accepted outcome
Take the metric seriously for a moment. An agent that resolves the task on the first attempt costs a fraction of one that fails, retries, and produces work someone has to correct. The difference is rarely the model. It's whether the model knew the constraint before attempt one instead of discovering it after attempt four.
And where did the constraint get stated? Out loud, in the meeting where the work was decided. The customer named the edge case on a call. The team ruled out an approach in the kickoff and explained why. That explanation is the reusable context OpenAI is pointing at, and in most companies it evaporates within the hour, leaving a two-line action item behind. The agent gets the two lines. Then it loops, and the loop shows up on the dashboard as spend.
Companies already pay to produce this context. It's called meeting hours. They just don't keep the output.
Reusable context is a folder of notes
Boswell records a meeting on your Mac, transcribes it, and writes the summary, the decisions, and the action items with on-device AI, filed as plain Markdown. That last part is what makes the context reusable in practice. Markdown needs no connector, no integration, and no line in the budget. Every model and agent you use this year already reads it. When you hand a task to an agent, you paste the three paragraphs from the meeting where the task was actually specified, and attempt one starts from what the room already knew.
The playbook's fourth step says the strongest strategic bets are built around proprietary company context. We'd only add that the most proprietary context a company has isn't sitting in a database. It's what your people say to each other all day, and it's the part nobody keeps.
The trust question, answered earlier
The playbook handles sensitive workflows the way an enterprise vendor has to: access controls, retention posture, compliance visibility, and Zero Data Retention options on their infrastructure. For a company running agents against production systems, that work is real and necessary.
For the meeting layer, we think there's an earlier answer. Boswell's recording and transcript never leave the Mac that captured them, so there is no retention posture to configure and no access policy to audit. The question "who can see the recording" has a short answer: you. When context does go to a model, it's the paragraph you chose to hand over, not the recording, and not by default.
What we don't claim
Boswell is not an admin console. It won't track your spend, run your evals, or govern your agents, and the playbook is right that somebody at an enterprise has to do those jobs. We do the narrow thing at the front of the pipeline: we keep the context your meetings already produce, in a format your tools already read.
Useful work per dollar is mostly decided before the first token is spent, by what the model is told. You already paid for that telling once. Keep it.