The short version
- An AI agent should answer only questions it can ground in something a human declared, and refuse the rest.
- The dividing line is declared vs undeclared, not easy vs hard.
- Facts, status, and recorded decisions are fair game. Speculation, intent, and undecided questions are not.
- A refusal is a valid, useful answer, because it tells the asker to escalate to a person.
An AI agent should answer only questions it can ground in something a person has explicitly declared, and it should refuse everything else. The boundary is not difficulty or sensitivity. It is provenance. If the answer traces back to a decision, a status, or a fact that a human stated, the agent can answer. If answering requires inventing intent, predicting what someone would decide, or resolving a question no one has resolved, the agent should decline and point to a person.
This framing solves the problem most teams struggle with. They try to enumerate safe topics and end up with brittle allowlists. The durable rule is simpler: an agent may relay what is known and declared, and must not manufacture what is not.
The line: declared vs undeclared
The right dividing line is declared versus undeclared. A declared answer exists because a human put it on the record: "we chose Postgres," "the launch is set for the 14th," "this decision is reversible." An undeclared answer does not exist yet; producing one means guessing. Agents are dangerous precisely because they are good at making guesses sound like declarations.
Note that this line ignores difficulty. A hard question with a declared answer is answerable. An easy-sounding question with no declared answer is not. The agent's job is to know which side of that line a question falls on, which requires it to draw from declared knowledge rather than an index it has to interpret. That distinction is explored in declared vs indexed knowledge.
What it should answer
- Recorded decisions: What was decided, by whom, when, and under what authority. These have a clear source and a clear owner.
- Declared status: Whether something is done, in progress, or blocked, when a person has stated it. Status the agent can point to is safe to relay.
- Stable facts: Ownership, dependencies, and context that a human documented. Answering here is retrieval, not invention.
- Where to go next: Who owns a question and how to reach them. Routing is always a safe answer, even when the substance is undeclared.
Everything in this list shares one trait: the agent is repeating something a person is accountable for, not originating a claim. That is the safe zone. For a fuller treatment of what an agent needs before it can answer anything, see what context AI agents need.
What it should refuse
| Question type | Why refuse |
|---|---|
| Undecided questions | No declaration exists; answering means guessing |
| Intent and motive | Inferring why someone decided invents a mind-state |
| Predictions | What someone will decide is not yet a fact |
| Contested topics | Discussion is not resolution; no authority to pick a side |
| Stale declarations | A superseded decision should be flagged, not relayed |
The common thread: each of these requires the agent to originate a claim rather than relay one. The moment an agent starts answering these, it is speaking on authority it does not have, and its confident tone makes the fabrication harder to catch than a human guess. This is the core case for teaching AI to decline, which we made in should AI refuse to answer.
Why refusal is an answer
A refusal is not the agent failing. It is the agent giving you real information: this question has not been decided, so go ask a person. That signal is often more valuable than a synthesized answer, because it prevents someone from acting on a decision that was never made. Treating refusal as a first-class response, rather than an error, is what makes an agent trustworthy. We unpack this in refusal as information and in silence over speculation.
The practical payoff: teams learn to trust what the agent does say. When an assistant never bluffs, its answers carry weight, and its refusals become a map of what your team still needs to decide.
How to enforce the boundary
Enforcing this boundary is an architecture choice, not a prompt. An agent backed by a search index will always be tempted to answer, because it can always retrieve something. An agent backed by a declared decision record can only answer when a real declaration exists, so the boundary is enforced by what the system can and cannot see.
This is StandIn's model. Its representative answers only from what your team declared, refuses when there is no declared answer, and routes the asker to the right person instead of guessing. That is how you get an AI that helps without overstepping the line between relaying and inventing.
Common Questions
Should an AI agent answer questions it is unsure about?
No. If it cannot ground an answer in something a human declared, it should refuse and route the question to a person. Uncertainty is a signal to stop, not to produce a best guess, because a confident-sounding guess is more dangerous than an honest refusal.
How do you decide what an AI is allowed to answer?
Use provenance, not topic. If the answer traces to a declared decision, status, or documented fact, it is answerable. If answering requires inventing intent, prediction, or resolution of an open question, it is out of bounds. This rule holds up better than any list of approved subjects.
Is refusing to answer bad for user experience?
Only if refusals are dead ends. A good refusal names why it declined and who to ask instead, which is genuinely helpful. Over time, users trust an agent that refuses honestly more than one that always answers, because they stop having to double-check it.
Can an AI agent answer questions about decisions that changed?
It should answer with the current declared decision and flag that a prior one was superseded, if that history is recorded. What it must not do is relay a stale decision as if it still stands, or reconcile conflicting records on its own authority.
Get async handoff insights in your inbox
One email per week. No spam. Unsubscribe anytime.
Ready to retire your daily standup?
Distributed teams use StandIn to start every shift with full context, no standup required. Engineers post a 60-second wrap. The next shift wakes up knowing exactly what to work on.