More context can produce a less relevant answer
In a six-month randomized field experiment with 6,000 workers, active Copilot users spent three fewer hours — or 25 percent less time — on email each week, yet meeting time did not significantly change. The tool moved work one person could change independently; it did not move the coordinated system around that person.
Consider “Can we deploy this?” The relevant answer changes with the reader’s role, jurisdiction, data classification, approval authority, time horizon, existing architecture, and tolerance for residual risk. Retrieval may find the policy and product documentation. It cannot infer which decision is on the table unless the product models that human context.
Four contexts hide inside one prompt
Factual context says what is true about the world and the product. Task context says what the person is trying to accomplish. Social context says who else is involved, which roles and norms matter, and what must be communicated across a group. Decision context says what alternatives exist, what evidence is sufficient, and what action follows.
These layers fail differently. Missing factual context creates hallucination or omission. Missing task context creates generic completeness. Missing social context produces an answer that cannot travel through the organization. Missing decision context produces information without consequence. A useful product needs to diagnose which layer is absent instead of treating every weakness as a retrieval problem.
Users need leverage, not prompt homework
Dynamic Prompt Middleware explores controls that help users refine the contextual framing of an explanation. The design principle is important: the product should make consequential context expressible without requiring people to become prompt engineers. A labeled choice such as “compare for procurement” can be more reliable than hoping a user writes the right paragraph.
The system should also reveal the context it inferred. “I am answering for an Israeli hospital procurement team evaluating a pilot” is correctable. Invisible personalization is not. Editable assumptions turn context into a shared object between customer and model, reducing the risk that a fluent answer rests on a private fiction.
Some context belongs to groups
Social-RAG investigates generation grounded in prior group interactions. This matters because organizations contain local vocabulary, unresolved disagreements, historical commitments, and tacit criteria not captured in a knowledge base. An answer that ignores them may be factually correct and politically impossible.
There are obvious privacy and governance risks. A person should not receive confidential group history merely because it would improve relevance. Context needs access rules, provenance, expiry, and a reason for use. Human-centered personalization is selective: it uses what improves the customer’s task while protecting boundaries the model is capable of crossing.
Find the boundary of useful context
Choose a recurring question whose good answer varies by role or stage. Build a minimum context schema with only variables that domain experts believe could change recommendation, evidence, format, or escalation. Compare generic retrieval with task-context controls and with editable inferred context. Keep the underlying model and knowledge base constant.
Measure task success, correction effort, perceived relevance, inappropriate disclosure, and downstream action. Examine whether each context field actually changes outcomes. Remove fields that merely make the answer feel personal. Keep only the smallest, transparent set of context that produces a materially better decision.