The ledger of design

In software and product development, the ledger of work is not a new concept. Software developers have used Source Code Management (SCM) tools for decades to track and memorialize decisions of work. There are also tools such as Jira that are a System of Record (SOR) that capture work. Despite various systems existing, Design hasn’t made progress on recording their work.

The artifacts for design have changed with AI. As a result, the ledger and proof of work has evolved. It’s no longer only documented artifacts such as mockups, diagrams, and decks to prototypes and living systems of code. One of the classic team challenges is visibility of work. Historically, it’s been difficult to find progress and proof of work because the immense effort it took. AI significantly solved this problem with search and retrieval, but there is a new problem as as result of AI. Now there is the abundance and slop problem. When teams can produce more work at higher volume, the process behind the work becomes easier to hide and harder to validate. Finished-looking outputs are cheaper now, which means credibility has to come from more than the finish.

Ledgers for various roles

This norm looks different across roles: individual contributors, managers, and even agents.

Contributors

AI can generate options, drafts, variations, and first passes, so the existence of an artifact no longer says as much as it used to. The stronger signal is how the contributor gathered context, shaped the problem, evaluated the options, and selected the final direction.

The quality of the artifact matters more, not less. When output is cheap, taste becomes more visible. A strong contributor can show the thinking that led to the work: the context they pulled in, the constraints that mattered, the prompts they used, the tools they created, and the directions they rejected along the way. The final deck, document, design, or code is only part of the proof. The rest is the judgment that explains why this version deserved to exist.

A contributor might create a prompt pattern, build a reusable skill, improve a workflow, or package a method so others can use it. That contribution is easy to miss if we only inspect the finished output. The best people will scale the systems along with their work.

Managers

The pressure is increasing for managers to have a ledger of their work. Much of the visible surface area of management is being compressed: meeting summaries, status updates, follow-up notes, planning documents, and context sharing. Those tasks still matter, but they no longer explain the value of the role. A good-looking plan can still be poorly reasoned.

The ledger has to move from activity to rationale. If everyone can ask an AI system for five plausible strategies, the manager’s contribution is not that they found an option. It is that they chose a direction and can explain why that direction is right for this team, this moment, this constraint. The decision trail becomes the proof. What was considered? What was rejected? What tradeoff was made? What risk was accepted? What would change the decision later? As iteration speeds up, teams produce more material than any one person can hold.

The manager’s job is to make the work legible: where the team is going, why it matters, what has changed, and what still needs human judgment. Management becomes less about being the source of every answer and more about creating the context that makes the right answer easier to recognize.

Agents

Ledgers are also important for agents, arguably more because agentic work requires human approval. A human has to be confident knowing what has been done. If agents operate across the toolchain of email, calendars, documents, browsers, and internal systems, users need to see what the agent was asked to do, what it accessed, what it read, what it changed. This is the most important trust layer for agentic work.

Instinct showed what happens when that layer is missing. During launch week, users found the agent had indexed and retained their email with no way to delete it. Access could be revoked; what had already been ingested stayed invisible. I’m confident the team will commit to improve the legibility of their agents work, but this is a lesson on how important visibility of work is that could errode trust.

GrokBot has a feature where a user can see the persistent cloud machine with a real browser, filesystem, and terminal, so the work happens somewhere you can point to. But every bot shares that machine, along with one credential pool and one set of browser sessions. xAI’s own documentation tells you not to treat separate bots as a security boundary. The product called its shared computer “a real blast radius.” Retention is undocumented, the audit view is promised but not shipped, and deleting a bot may not clear its files or sessions. You can see where the work happened. You cannot scope it, review it, or take it back.

That is the shape of agent credibility: visible execution, scoped access, reviewable history, reversible actions. The four are not interchangeable, and a product can nail one while failing the rest. If agents are going to be collaborators, they need what human work needs: a trail of context, action, and accountability.

Build the ledger

A useful ledger has to be easy to capture, or it becomes bureaucracy. The trace should come from tools people already use: documents, commits, prompts, agent runs, decisions, artifacts. If capture takes effort, it turns into after-the-fact performance.

It also has to be useful to review. A ledger that records everything clarifies nothing. Keep what someone would need to understand how the artifact came to be: what changed, why, who approved it, what evidence held. Where did this come from? What generated it? What did a human change? The wrong version of this is surveillance. The right version shows where judgment was applied and who stands behind the result.

Build the ledger around your work: the context, the decisions, the toolchain, the approvals, the artifact. Then hand someone the finished piece and ask them how it got made. If they can answer, you have one.