Intake Form Design for Superintelligent Agents: Clean Fields
Which intake fields a small business should require for an AI agent, how to word them, and how to keep the data clean enough for later models to trust.

An intake form for an AI agent is the short set of fields a requester completes before any work begins. Its job is to turn a request into a record with a clear outcome, a named owner, a source of truth, explicit limits and a point where a person must approve, so the agent and any later model can act without guessing. For a small business, a few precise questions worded as facts a person can answer produce far more usable data than a long form full of open text.
Key takeaways
- An agent acts reliably only on what the form makes explicit, so each field should answer a question the agent would otherwise have to guess.
- Require the outcome, the accountable owner, the source of truth, the permitted actions, the approval point and the success check.
- Word each field as a fact a person can answer, choose or mark as not applicable, never as an open impression.
- Keep optional fields visibly optional, and let the agent stop work that is missing a required field.
- Review blank and rejected submissions on a schedule, because the form drifts as products, people and permissions change.
Why an agent needs a form instead of a chat message
Most requests arrive as a sentence in a chat window, an email or a voice note. A person reading that sentence fills the gaps from memory, from a phone call last week or from knowing who usually handles the account. An agent does not carry that background by default. It reads what is in front of it, plus whatever memory and tools you have granted, and when a request leaves something out it will often fill the gap with a plausible guess.
A form forces those gaps into the open at the moment of intake. It also creates a stable record. When a later model summarizes the work, reviews the outcome or takes over after a handoff, it reads the same fields instead of reconstructing intent from a long thread. That is what data a later model can trust looks like in practice: not perfect data, but data where every value has a known origin, a known meaning and a known owner.
Which fields a small business should require
Most small businesses need no more than eight fields for a first agent, and many need fewer. These are the ones that recur most often in useful intake designs:
- Requested outcome: one sentence describing the result that makes the request done.
- Requester and accountable owner: the person who asked, and the person who accepts the result.
- Source of truth: where the facts live, such as the order system, the customer record or a named shared folder.
- Subject reference: one identifier the agent can match against, not a name typed from memory.
- Deadline or trigger: the date or event after which the request no longer helps anyone.
- Permitted actions: what the agent may read, draft, update, send or cancel.
- Approval point: the step where a named person confirms before anything leaves the team.
- Success check: the observable condition that shows the work is correct.
A field earns its place when the agent would otherwise have to guess. A quote request may need the outcome, the subject reference, the deadline and the approval point, while a recurring report depends more on the source of truth and the success check. A task contract works alongside the form: the form captures this particular request, and the contract states what counts as done for every request of that kind.
How to word each field so the answer can be checked
Wording decides whether an answer is usable. "Describe the problem" invites a paragraph that no one can route. "Which product line and which region does this concern?" invites a value that a system can validate. Prefer questions with a closed set of answers, a stated format and an example in the placeholder text. Avoid relying on words such as urgent or important as the only signal. Ask for the date the customer needs an answer by, and ask what happens if that date passes.
"Deadline: the date the customer must have a reply by, in YYYY-MM-DD format, or none."
Use explicit answers such as none or not applicable, so a blank can never be confused with a forgotten field. Where the value exists in another system, have the form look it up and show the match to the requester for confirmation, instead of accepting free text. Keep one meaning per field. A field called owner should never sometimes mean the person who submitted the request and sometimes mean the person who will do the work.
Permitted actions deserve particular care. Use a short list of verbs, such as read, draft, update, send and cancel, and make sure each verb maps to a permission the agent actually holds. The OWASP authorization guidance is a useful reference here. Grant what the task needs, and check permission at the point of action, not only when the request is submitted.
Decide what stays optional and what the agent must refuse
Every field needs a rule for the case where it is missing. The decision is usually simple once it is written down:
- If the agent cannot complete the work without the value, mark the field required and block submission with a message that names the missing field.
- If the value improves the result but the work can proceed without it, mark the field optional and state what the agent will assume when it is empty.
- If the value can be found in a trusted source, use a lookup and ask the requester to confirm the match rather than letting the agent choose.
- If the value changes money, a commitment or a customer-facing message, require a named approver, and do not let the requester approve their own request.
- If the field holds personal data, collect only what the task needs and write the purpose beside the question.
The last rule matters more than it first appears. Personal data in an intake form creates obligations for the business whether or not an agent reads it. In Brazil, the national data protection authority publishes guidance on how personal data should be handled, and the same principle of collecting only what a purpose requires applies to any form, paper or digital. Routing of approvals is a related design question, and the approval routing guide covers how to send each request to the right person.
Illustrative scenario: a quote request that arrives half finished
Illustrative scenario: a small equipment rental company wants an agent to prepare first drafts of rental quotes. The first version of the form asked, "What do you need?" The agent received messages such as "the usual, for next week." It then had to guess the equipment, the dates and whether the customer already had a standing agreement, and some of those guesses were wrong.
The revised form asked for equipment from a fixed list, start and end dates in date fields, the customer from a lookup against the customer record, and a yes or no question on whether the standard pricing sheet in the shared folder applies. The deadline was required, and the approval point was set so that a coordinator reviews every draft before it is sent. Incomplete requests now return to the requester with the missing field named. The agent did not become more capable. The data it received became specific enough to check.
Be Human handles intake as part of the first engagement rather than as a form someone builds and forgets. The work starts with the single role the agent will perform, the fields that role needs, the wording agreed with the people who submit requests and the observable success check. Managed operations then review incomplete and rejected submissions, track which fields cause guessing, and route any change to the form through the same approval path used for changes to the agent's permissions.
Keep the form clean as the work changes
Forms drift. A new product line appears, a team member leaves, a permission is widened to save time. Review the blank and rejected submissions on a fixed schedule. A field that is marked not applicable almost every time may be optional in practice, and it should be labeled that way. A field that users edit after submission usually has unclear wording. Record the version of the form with each request, so anyone reading the record later knows which questions were actually asked.
Keep a short change log that records what changed, why, and who approved it. Treat the form as part of the agent's operating context, alongside its tools and permissions. Anthropic's guidance on writing tools for agents makes a related point: a model reads tool names and parameter descriptions, and ambiguous ones lead to misuse. Intake fields work the same way. Teams that want to bring a new person into this discipline can start with the onboarding guide for AI agents, which explains how the agent and the people around it share responsibility from day one.
If you are not yet sure which of these fields your first agent needs, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How many fields should an intake form for an AI agent have? Fewer than most owners expect. Start with the fields the agent cannot act without, then add only the fields that remove a specific guess the agent would otherwise make. Review the list after the first month of real submissions, because the fields that matter are usually visible only in the blanks.
Should the agent ask the requester for missing information or refuse the task? It depends on the field. Required fields should block submission with a clear message that names what is missing. Optional fields let the agent proceed, but the agent should record what it assumed so a reviewer can check that assumption before anything leaves the team.
Can an existing paper or web form be reused for an agent? Often in part. Older forms were written for people who fill gaps from context, so they tend to contain open text and vague labels. Keep the fields that map to real data, reword the rest with closed answers, and add the approval and permission fields that were previously left implicit.
Who should own the intake form after launch? One named person in the business, usually the operations lead for that workflow, should own the wording and the change log. Be Human's managed operations can handle monitoring, review of submissions and controlled changes, but the business still decides which fields reflect its own rules and who approves them.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


