Designing AI agent intake: clarify the request and carry the case forward
A proposed design workshop for turning an ambiguous request into an agreed deliverable, with explicit questions, case notes, corrections, and restart rules.
Define the deliverable before writing the dialogue
Our proposed approach to AI agent intake starts with a small design document: what the user wants produced, which decisions remain open, and what the agent is expected to do next. Treat the approach in this article as an editorial proposal for your own review. We do not present it as a validated operating standard or a description of a particular product's capabilities.
Imagine a fictional request: “Can you put something together for the planning meeting?” For this exercise, leave several interpretations open. The requested output might be an agenda, a briefing document, or a presentation outline. Our proposed first response is: “Would you like an agenda, a briefing document, or something else?” Include the open alternative rather than requiring the user to accept one of the suggested formats.
Write an acceptance statement for each deliverable. For an agenda, our example statement is: “A draft containing the meeting objective, proposed discussion items, and unresolved preparation questions.” For a briefing document, choose a different statement that specifies its audience and permitted reference material. These are design choices for the exercise, not claims about universal requirements.
Keep the permission to draft separate from permission to distribute. In this proposed workflow, “put something together” authorizes preparation within the agreed scope. Design any later distribution step as a separate, explicit decision. State that boundary in the case record rather than asking the user to interpret an internal task label.
Use the sources within their actual scope
The source bundle provides editorial guidance and a product registration page, rather than documentation of conversational intent recognition or case memory. Accordingly, the conversational mechanics below are our proposed conventions. The sources are used only for the narrower statements they support.
Google's guidance on generative content recommends human fact-checking and review before publication, including review of metadata. Google's generative content guidance. Its content self-assessment questions address original analysis, descriptive headings, and clear sourcing. Google's helpful content guidance. Use those statements as the editorial basis for checking a published guide, without treating them as evidence that a particular conversation design produces better service outcomes.
For your own design document, we propose a similar separation: label external evidence, local operating decisions, and fictional examples. When the team proposes a rule such as “ask about the audience before drafting,” describe it as a chosen rule for that workflow. Do not give it an invented research attribution. Apply the same convention to the sample responses, review criteria, and diagrams in this article.
Give every clarification question a decision
Our proposed question worksheet has three fields: missing information, the decision it affects, and the response to “I don't know.” Before adding a question to the dialogue, fill in all three fields. If the team cannot identify an affected decision, move the question to an optional section or remove it from this particular workflow.
For the fictional meeting document, choose audience and document type as required fields. Choose target length as optional. This is an example configuration, not a prescription for every writing task. A team drafting technical specifications might choose different required fields; write that alternative as its own configuration rather than silently expanding the meeting workflow.
Once the user selects a briefing document, our example next question is: “Who will read it, and what decision should the meeting address?” If that feels too broad for your intended interface, test a version that asks about the audience first. Record which version your team adopts and the criterion used to select it. Avoid declaring either wording universally superior.
Specify the unknown branch in advance. For “We haven't decided the audience,” offer a choice between pausing, preparing a clearly labeled outline, and asking the user to select a provisional audience. In our convention, an unanswered question remains unanswered. Do not replace it with a default unless the workflow explicitly permits a visible provisional value.
Also write the partial-answer branch. If the user answers “the operations team” but leaves the meeting decision unspecified, store the audience and ask only about the remaining decision. Our proposed response is: “I'll address the operations team. What should they be ready to decide?” Do not restart the entire intake sequence in this sample design.
Our proposed question worksheet
Specify the question and its branches
- 1
Pending decision
Write the choice that needs an answer.
- 2
Missing value
Identify the information required by your chosen rule.
- 3
Question
Draft wording that asks for that information.
- 4
Branches
Specify responses to unknown and partial answers.
Maintain a case note with explicit update rules
We propose a compact case note containing the requested output, intended audience, agreed scope, reference material, confirmed constraints, unresolved questions, and next action. Alongside each value, record where it came from. For the workshop exercise, that could mean a reference to a user message, a supplied document, or a labeled editorial assumption.
Adopt a visible distinction between stated information and proposed information. For example, record “operations team” as user supplied and “two pages” as an agent suggestion until the user accepts it. Choose your own status vocabulary, but specify which events change a status. In our convention, generating a suggestion does not count as receiving approval.
Define correction behavior with a concrete example. The fictional user first asks for a briefing and later says, “Actually, make it an agenda.” Our proposed update replaces the active deliverable, marks the previous deliverable as superseded, and opens a review of fields tied to the old format. Preserve the audience only if the user has not changed it and your chosen rules permit that carryover.
Handle contradictory details separately from explicit corrections. If one message says “Monday's meeting” and another says “the Wednesday session,” our proposed response is: “Which meeting should this document cover?” Leave the meeting field unresolved until the answer arrives. Do not let a team member's private assumption appear in the final case note as a user-confirmed fact.
For review purposes, retain a short account of meaningful changes: what changed, which message authorized it, and which open questions remain. Set retention and access arrangements through your own organizational process; this article proposes the contents of a design exercise, not a storage policy.
Our proposed update convention
Treat different changes explicitly
- Case note
- Store the supplied field; leave the other open.
- Next response
- Ask about the remaining gap.
- Case note
- Replace the active value; mark the old one superseded.
- Next response
- Confirm the change and review dependent fields.
- Case note
- Record the competing values as unresolved.
- Next response
- Ask which value to use.
Design a deliberate return to the conversation
For resuming a case, our proposed opening contains three elements: the agreed deliverable, the last confirmed constraint, and the next unresolved choice. In the fictional example, write: “We were preparing an agenda for the operations team. The meeting objective is still open. Would you like to define it or revise the document type?” Treat this as sample wording to review, not a claim that the system has persistent memory.
Design a way to correct that summary. If the user says “This is for the finance team,” update the audience and identify any draft sections that your workflow requires the team to reconsider. Do not interpret a correction to the audience as permission to change unrelated constraints.
Create a separate rule for a new request. If the user writes “Also help me draft a welcome note,” our proposed convention opens another case and asks which task to work on first if sequencing matters. Link the cases only when the user or your documented process provides a reason. Do not automatically copy the briefing's audience or meeting objective into the welcome note.
For a handoff, propose a short transfer note containing the active deliverable, confirmed details, open questions, and the last agreed next step. Include references to the relevant messages. In this convention, passing the case to another person does not itself approve a draft, resolve an ambiguity, or expand the task.
Review behavior with fictional transcripts
Our proposed review pack contains six short conversations: a complete request, an ambiguous request, a partial answer, a correction, a new topic, and a return after a pause. Write the expected case note beside each transcript. Evaluate whether the design follows its own rules; do not describe this exercise as proof of business performance.
For the ambiguous request, check whether the response asks about the deliverable. For the partial answer, check whether it preserves the supplied value and asks about the remaining gap. For the correction, check whether the superseded value is excluded from the active instructions. For the return, check whether the summary distinguishes confirmed information from open questions.
Add a refusal example: “I don't want to choose a length.” Under our sample configuration, length is optional, so the expected response should offer to continue without a fixed length. If your team makes length mandatory in another workflow, write a separate expected response that explains which step is waiting for that decision.
Finish the review by assigning an owner to each unresolved design choice. Approve the wording, field rules, and completion definition together. Our proposed first implementation scope is one inquiry type with a small set of reviewed examples; expand the design only after specifying the rules for each additional inquiry type.
Our proposed review sequence
From sample dialogue to agreed specification
Write examples
Cover ambiguity, partial answers, corrections, new topics, and returns.
Compare notes
Check each expected case note against the proposed rules.
Resolve choices
Assign an owner to unanswered design decisions.
Approve scope
Agree the wording, field behavior, and completion definition.
Take a concrete specification into your next step
As a practical exercise, write one page containing a deliverable definition, a clarification worksheet, a case-note template, and a restart response. Use the fictional meeting request above or create a scenario suited to your organization. Mark every sample assumption and every unresolved design decision explicitly.
The Max AI registration page describes the service as an “Operating system for AI agents.” Max AI registration. We use that page as the official product entry point; it does not substantiate the conversation mechanisms proposed here.
Choose “Create account” to visit registration. For your design work, keep the proposed specification and reviewed transcripts together, so the next reviewer has an explicit description of the behavior you intend.
Sources
- Max AI - Operating system for AI agents — maxaiagent.app; accessed November 17, 2026
- Google Search's Guidance on Generative AI Content on Your Website | Google Search Central | Documentation | Google for Developers — developers.google.com; published October 1, 2026
- Creating Helpful, Reliable, People-First Content | Google Search Central | Documentation | Google for Developers — developers.google.com; published October 5, 2026
Start working with your own agent
The agent knows this article context and still follows the assigned department checklist.
Related articles
How to evaluate AI agent economics: cost per case and operator effort
A proposed evaluation worksheet for comparing accepted outcomes, operator hours, and process costs before and after introducing an AI agent.
AI Agent Permissions: A Proposed Workflow for Data, Tools, and Action Approval
Define what an agent may read, prepare, change, and publish. Our proposed workflow includes a permission brief, a review packet, and a staged publishing example.
Monitoring AI Agents After Launch: A Practical Review and Response Workflow
Build an explicit review process for service quality, investigate failed conversations, and define the evidence required to change or restore an agent’s scope.
Designing an AI-to-human handoff: triggers, context and ownership
An editorial playbook for defining escalation triggers, preparing a handoff record and assigning responsibility through acceptance and follow-up.