A few days ago, I published a How To about prompting in outcomes rather than instructions:  give the system a finish line and what success looks like, not just instructions. I was also thinking about how the new /goal functions from the AI labs don’t quite fit into knowledge work. I kept circling back to it and realized that in knowledge work, we often use our expertise to map out a workflow for AI to follow, but what if the way a human adopts the workflow isn't necessarily the best way for AI to execute it? We still need a human in the loop, but the loop might be completely different. (Allie Miller wrote about the same concept today, too. Great minds…)

I run workflow audits. I start by walking through the process step by step as an expert would, label each task as needing live context, being repeatable enough to become a skill, or requiring human approval, then ask AI how to make it more efficient. It is reasonable, and it is, as Allie put it, a faster horse. A renovation of the existing house, rather than a question about whether you need the house.

The method that replaces this is simple and feels slightly disorienting the first time. Before describing any existing workflow, I write one sentence: what is this work meant to do, from the perspective of the person who receives it. Then I connect the model to the live sources where the work actually lives (my email, calendar, drives, and tools). Then I ask it to propose how the work should run, including where a human needs to be in the loop. I give it a destination and a set of live tools and ask what path it would take.

What comes back looks like what you would get if someone had never seen your existing process, had never met your team, and was starting from the outcome and the available tools, deciding from scratch. Sometimes the existing architecture survives, now with a clearer account of why it is shaped the way it is. More often, something shifts because the current shape was a workaround for a constraint that no longer exists.

Every organization has rooms that only exist because of a constraint that no longer applies. Take a quarterly report tracking what is happening in the market. Someone spends a week every quarter reading industry coverage, pulling relevant reports, and building a summary for the leadership team. It gets presented. Notes get taken. Three months pass, and the cycle runs again. If you audit that workflow, you label each step: gathering sources requires live context; synthesizing the findings creates context for what follows; formatting the final output is repeatable enough to hand off to AI. The audit speeds up formatting and may assist synthesis. The output is still a quarterly report, just produced faster. That is renovation.

When I ran the outcome-first pass on the same workflow, the question was what it was actually supposed to do for the people receiving it. The answer: they need to know when something in the market shifts before it becomes a problem. That outcome has no quarterly rhythm. It is continuous. Once a model has access to the same sources the human was reading, it can surface changes the day they happen. The quarterly report becomes the guest room nobody sleeps in, but everyone keeps making up with fresh sheets, while the ongoing feed becomes something the team queries whenever a question comes up.

My labeling system survived this, which I was relieved to find. Instead of pointing those criteria at the human workflow, I point them at the workflow the model proposes: which steps need live context or create it, and which are repeatable enough to become skills. These questions tell you which rooms in the new plan are load-bearing and which can be reconfigured. The human-in-the-loop label became the most important one. In the audit version, that was a flag for what could not be automated yet. In the outcome-first approach, it is a design choice. Those places deserve an actual decision.

Automating what exists produces real gains, and for most teams, it is still the right place to start. The outcome-first pass belongs at a different stage, once you have enough experience to ask whether the assets you have been making are still worth making. The audit puts you in the operations manager seat. This puts you back in the editor-in-chief seat, deciding what gets made and what gets cut.

If you want to have that conversation on a specific workflow, give me a buzz.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

I’ve also started publishing more frequently on LinkedIn, and you can follow me here

if you’d like to chat further about opportunities or interest in AI, or this newsletter, please feel free to reply.

banner images created with Midjourney.