Agents should notice ambiguity before they guess: designing for vague requests
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Agents now take requests directly from marketers, often through chat, and those requests arrive the way requests between people do: short, partial, and full of shared context the agent does not have. How an agent behaves in that gap decides whether it saves time or creates rework.
"Can we do something for the webinar?" "Make the email more exciting." "Target the enterprise segment." Each leaves important questions open. People resolve this through conversation and shared history. Agents, by default, fill the gaps with plausible assumptions and proceed, and the result is polished work that misses the point.
Why confident guessing is the default
Language models are trained to produce a complete, helpful-looking response to whatever they are given. A request with a gap still gets a full answer, because filling gaps fluently is what the model does well. Instructions that reward completion, such as "always deliver a finished draft," reinforce the habit.
The cost lands later. A requester who asked for something for the webinar receives a three-email sequence when they wanted a social post. They now spend time explaining what they meant and reviewing work they did not want. Measured across the whole workflow, the agent was slower than a single clarifying question would have been.
Four kinds of ambiguity
Ambiguity comes in distinct kinds, and each calls for a different response.
Missing information. The request lacks something essential, such as the audience, deadline, channel, or format. The work cannot be done well without it.
Vague goals. The request names a direction but not a destination. "More exciting" could mean a stronger subject line, a shorter body, a new offer, or a different tone.
Multiple interpretations. The request makes sense in two or more ways that lead to different work. "Enterprise segment" may mean company size, contract value, or a specific named list in the CRM.
Conflicts. The request contradicts a policy, the brief, or another instruction. "Send it to everyone" may collide with a consent rule or a frequency cap.
Four ways to respond
The responses map loosely onto the kinds, though a single request can call for more than one.
Ask one good question. When essential information is missing, ask a single specific question rather than a questionnaire. A question like "Is this for existing customers or new prospects?" moves the work forward, while a list of ten hands the thinking back to the requester. If more than one thing is missing, ask about the one that most changes the work and state assumptions for the rest.
State assumptions. When the ambiguity is minor or the work is cheap to redo, proceed with a sensible assumption. Put it at the top of the output: "I assumed this is for existing customers. Tell me if not." The assumption becomes a visible, one-line correction instead of a hidden error.
Offer options. For vague goals, produce two or three short alternatives that interpret the request differently, each labeled by its interpretation. The requester chooses, and the choice becomes context for the next round.
Escalate conflicts. When the request conflicts with policy, do not resolve it alone. Explain the conflict, name the rule, and ask who should decide. See escalation paths.
Choosing the right strategy
A simple rule covers most cases: the more consequential and less reversible the work, the more the agent should ask before acting. An internal draft can proceed on stated assumptions. Anything customer-facing, costly, or hard to undo should wait for an answer. This mirrors the reasoning behind the autonomy ladder, applied to a single request instead of a whole workflow.
Here is how we would encode that rule as a decision sequence in the agent's instructions.
- Check the request against policy, and if it conflicts, escalate and stop.
- Identify what is missing or unclear, and whether it would change the output materially.
- If nothing material is unclear, proceed.
- If the work is customer-facing or hard to reverse, ask one question and wait.
- If the goal is vague but the work is cheap, offer two or three options.
- Otherwise, proceed on a stated assumption.
Consider a hypothetical request to target the enterprise segment with a new launch email. Step two finds that "enterprise" has three plausible meanings, and step four applies because the email goes to customers. The agent asks whether enterprise should mean companies above a size threshold, accounts above a contract value, or the named enterprise list in the CRM. One answer settles the work and saves a rewrite of the whole send.
Teaching agents to notice ambiguity
Noticing comes before handling, and it is the harder skill to instill. Include examples of ambiguous requests in the agent's instructions, each paired with the preferred response and a sentence on why. Contrast them with clear requests that should simply proceed, so the agent does not learn to ask about everything.
Over-asking is a failure mode too. An agent that asks a question on each request trains people to stop using it. Track both sides: how often the agent proceeds on a request a reviewer judges ambiguous, and how often it asks about something already clear from the request or the brief.
Add ambiguous cases to your golden dataset and grade the type of response, not just the quality of the writing. A well-written email built on the wrong interpretation should fail the test.
Grade with a small, fixed label set so results stay comparable over time: asked, assumed and stated, offered options, escalated, or proceeded. Have a reviewer assign the expected label to each case before the agent runs, then compare. Disagreements between reviewers on the expected label are useful too, since they show where your own team finds a request ambiguous.
Reduce ambiguity upstream
The most effective fix is often better input. Structured request forms that ask for audience, goal, channel, and deadline close the gaps we would expect to see most often before they reach the agent. Clear briefs and shared campaign plans supply context that would otherwise have to be asked for.
Even with good inputs, some ambiguity will remain, because some of it is real: the requester has not decided yet. The open question we are studying is how agents can help people decide, by surfacing the trade-off behind a vague request, without drifting into making the decision for them.
How to apply this: add five deliberately ambiguous requests and five clear ones to your agent's test set. Grade whether it asks, assumes, offers options, or escalates when it should, and whether it simply proceeds when the request is clear.
Related research
Small models, big jobs: matching model size to marketing tasks
Not every marketing task needs the most capable model available. Matching model size to the task can improve speed and cost without a loss in quality.
Tool errors are marketing errors: building resilient agent actions
When a tool call fails halfway, the result can be a half-sent campaign. Resilient agent actions keep technical errors from becoming marketing mistakes.
Agent personas are not characters: defining roles that hold up
Giving an agent a name and a personality is easy. Giving it a role with clear responsibilities, boundaries, and standards is what makes it dependable.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
