How Lucy learns what works

Five steps, kept separate

“Lucy learns what works” is a precise claim, not a slogan. Our product design keeps five steps separate, because blurring them is how AI products overclaim.

1. Observe

   With permission, Lucy notes what actually happened: the apps, steps and timing of real work.

   Example: Someone is researching a company.
2. Interpret

   She forms a best guess about the goal, notes how sure she is and what else it could be, and stays quiet when evidence is weak.

   Example: Meeting prep? A supplier check? Just learning?
3. Confirm

   She asks. The person confirms or corrects her, and a correction always wins for that person.

   Example: “It’s a supplier review on Friday.”
4. Measure

   Completion, adoption, capability and value are measured separately. A finished task is not proof of skill.

   Example: The review prep they built is used again next week.
5. Reuse, carefully

   What worked is vetted before it helps anyone else, and reaches other people only after privacy checks, as protected aggregates or solutions their builders choose to share, never your business processes or trade secrets.

   Example: A supplier-review workflow, offered to the next person who needs it.

Why the distinctions matter

What Lucy saw is not the same as what she thinks it means. What she thinks is not the same as what the person tells her. A finished task is not the same as a new skill, and a new skill is not the same as business value. And one person’s success has to be qualified before it informs anyone else. Each of those confusions is a common way for AI products to claim more than they know.

An example

Someone researching a company might be preparing for a meeting, assessing a supplier or simply learning. The activity alone does not settle the goal. Lucy forms a best guess with alternatives, stays quiet if the evidence is weak, and asks. If the person says it is for a supplier review on Friday, that correction wins. It applies to that person’s intent and does not rewrite the raw evidence.

Plans then work backwards from the goal through its prerequisites: the person’s current skill, the tools they are allowed to use and the feasible next step. A novice and an expert get different paths to the same outcome, and the first useful plan can start from a goal the person states without waiting for weeks of observation.

Measuring what actually changed

Lucy measures completion, adoption, capability and value separately, against a baseline. Success is capability, not usage: needing Lucy less for a task can be the win, because the person can now do it alone. Throughout, the person chooses the tools, takes the actions and validates the solution; nothing in the design lets Lucy click, type, install or deploy on their behalf.

Simple on the surface, science underneath

We plan to hire across six disciplines, each owning a different question: workflow research (what people really do), applied science (what someone is trying to do, and when to ask), behavioral science (which coaching changes habits), knowledge engineering (what a goal requires), AI evaluation (is the inference right, against independent examples) and product data science (did capability and the work improve, and why).

What compounds

The asset that grows is not a pile of screen data. It is structured knowledge of what helps, whom it helps and when it applies. That is why every win makes Lucy a better coach: she learns how to help people build better solutions, adapt to different preferences and understand different industries, and better coaching helps the next person. See How Lucy learns for the visual version.

Next: Seeing work got cheap. Trust and coaching didn’t.