AI Assistants May Need to Remember Why We Work, Not Just What We Ask

The missing question: why?

Workplace AI systems increasingly draw on documents, conversations, and organizational resources. Yet access to information does not necessarily reveal what someone hopes to accomplish. Even a detailed instruction can leave the broader purpose unstated.

In Goals as a Representational Layer for Human–AI Collaboration, Gaia Molinaro, Lev Tankelevitch, Sean Rintel, and Scott Counts propose making those purposes explicit. Their idea is a persistent, structured collection of goals connecting immediate requests to the outcomes they are meant to serve.

The November 2026 preprint is a conceptual proposal, not a report of a newly tested system. It draws on research about human goals, workplace collaboration, and AI interaction to develop design requirements, practical recommendations, and unresolved challenges. Its central contribution is a framework, rather than evidence that a goal layer improves performance.

A changing map of intentions

The authors start with three properties of human goals. Goals are partly hidden: people may not articulate them clearly, and sometimes discover them while working. Goals change as circumstances and priorities evolve. They also operate at different levels, from concrete tasks to broad aspirations.

A useful representation would therefore need to be explicit, continuously revisable, and able to connect different levels of purpose. It could not simply be a permanent checklist attached to a user profile.

The proposed structure is a goal graph: a network in which goals connect to supporting activities and broader objectives. Unlike a simple family tree, one activity could support several goals, while several alternative approaches could lead to the same outcome. The framework also allows for tensions, where progress toward one objective obstructs another.

To describe each goal, the authors suggest seven questions: what, who, who else, when, where, why, and how. These cover the desired outcome, responsibility, collaborators or audiences, timing, relevant tools and resources, broader purpose, and supporting steps. Not every question needs equal detail for every goal.

Additional records would track completion, abandonment, revisions, and provenance—whether information came from a person or an AI proposal. An inferred goal should remain distinguishable from one its owner has explicitly endorsed.

How the layer could guide assistance

The authors recommend combining human articulation with AI inference. Conversations and clarification questions could help people express their aims. Evidence distributed across their activities could suggest additional candidate goals. People would shape and endorse the representations used to guide assistance, rather than having inferred intentions treated as authoritative.

For an individual request, the system would retrieve relevant goals and enough surrounding context to understand the purpose—not load the entire graph into every interaction. That information would then inform concrete choices: an audience could shape an explanation, a deadline could constrain a plan, and linked documents could direct retrieval.

The paper's report-summary example illustrates the distinction. Preparing for a meeting, drafting a proposal, and onboarding to a project may require different summaries of the same material. A goal layer could help carry that purpose across interactions without repeated explanation.

Consequential interpretations would need to remain visible and correctable. If goal context substantially changed a response or plan, the system should expose the connection and assumptions. New information could also trigger proposed revisions to the goals themselves.

Benefits are hypotheses, not results

The proposed benefits extend beyond more relevant responses. A shared representation could help collaborators compare priorities, notice dependencies, and reflect on whether current activities still serve their intended outcomes.

It could also change AI evaluation. Instead of focusing only on output quality, volume, or time saved, researchers could examine whether assistance supports outcomes people value, including learning and well-being. Simply counting completed goals would be inadequate: discovering that a different goal matters more could be a positive outcome even if completion falls.

These are research opportunities, not demonstrated effects of the proposed layer. The preprint supplies no new deployment trial or quantitative estimate of its benefits.

The risks of making purpose visible

Inferring intentions is difficult because records describe past behavior, while goals concern desired futures. Even people's own accounts can be incomplete or change through reflection. Maintaining an accurate representation could also burden users, and exploratory work may require deliberate ambiguity rather than precise targets.

Privacy presents another challenge. Gathering scattered information into one persistent representation could enable intrusive monitoring or evaluation. The authors argue that visibility must be bounded: a goal should not automatically become accessible to every manager, collaborator, or AI system.

Most importantly, recording a goal does not ensure an AI will follow it. The proposal offers possible infrastructure for alignment, not a solution to alignment itself. Its promise depends on whether systems can preserve human authority while making purpose useful.