Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

AI Scope Statement Template: One Page to Fill In Before Any Vendor Call

A one-page AI scope statement template covering role, memory sources, tools, approval points and success criteria, to complete before any vendor call.

Be Human typographic cover with the title: AI Scope Statement Template: One Page to Fill In Before Any Vendor Call
Conceptual illustration created with AI · Be Human

An AI scope statement is a one-page document that names one agent role, the memory sources it may read, the tools it may call, the points where a person must approve an action, and the observable result that shows the role is working. Write it before any vendor call, because a scope lets you compare proposals against the same requirements instead of against the most polished demo. The template below can be filled in by the leader who owns the outcome, then shared with whoever will build and operate the agent.

Key takeaways

  • Scope one role with one owner, one outcome and one channel before naming any technology.
  • List every memory source and tool by its specific name, with its access mode and who can change it.
  • Write approval points as conditions a person can test, such as a recipient outside the customer record or a change to an order.
  • Define success as something you can observe in records you already keep, measured over a fixed review window.
  • Bring the completed page to every vendor call so each answer is checked against the same written scope.

Why the scope comes before the vendor call

Vendors demonstrate what their platform does well. A demo shows capability, not fit. Without a written scope, every proposal sounds reasonable, and the comparison slowly becomes a judgment about presentation skill. A scope turns the call into a set of concrete questions: can this platform read the memory source we named, can it restrict a tool to draft mode, can an approval step pause the workflow before a customer receives a message.

The scope also protects ownership. Be Human describes an AI agent as one that keeps its identity, memory, channels, tools, permissions and operating context under human control. Those six things are exactly what a scope page must describe. If a proposal cannot tell you where each one lives and who can change it, the proposal is incomplete, whatever its feature list says. The NIST AI Risk Management Framework organizes AI risk work around governing, mapping context, measuring and managing risk. A one-page scope is the practical starting point for mapping context for a single role. If your company needs a group to approve the page, the governance committee charter describes a lightweight way to set that up.

The one-page template

Copy the fields below into one document and complete each one in plain sentences. Leave a field marked as open if you have not decided it yet. An open field is useful information, and it becomes a question for the vendor.

  • Role: one sentence that starts with a verb, for example, answer inbound questions about order status and route exceptions to the operations lead.
  • Business owner: the named person who accepts the outcome and can stop the agent.
  • Users and channels: who talks to the agent and where, such as the internal chat, the support inbox or a customer portal.
  • Memory sources: each source by system or folder name, the fields or documents it contains, the date range it covers, and whether the agent may write to it.
  • Tools: each action the agent may take, the system it touches, and its mode: read, draft or execute.
  • Service identity: the account the agent uses, with the narrowest access that lets the role work.
  • Approval points: the exact conditions that pause work until a named person approves.
  • Out of scope: what the agent must never do, written as plainly as the in-scope items.
  • Success criterion: one observable measure, the record it comes from, the baseline period, and the date of the first review.
  • Failure handling: who is alerted when the agent errs, what it does on error, and how it is switched off.
  • Open questions: decisions not yet made, each with an owner and a date.

Keep the page to one side of paper. If a field needs more than a few lines, the role is probably too broad, and the fix is to split the role, not to shrink the font.

Name every memory source and tool, not categories

The phrase all company documents is not a memory source. A usable entry names a system, a folder or table, the fields included, the date range, and a flag that says whether the agent may write. Narrow naming also makes the vendor's answer checkable. Ask the vendor which retrieval path reaches each named source, and whether that path can be turned off without disturbing the rest of the agent.

Separate reading from writing. An agent that drafts replies may need to read a customer history, while the right to update that history stays with staff. Documents that the agent retrieves can also contain text written by someone outside the company, and that text can attempt to instruct the agent. The OWASP Top 10 for LLM Applications treats this class of risk seriously. The practical rule for the scope page is that retrieved content can inform an answer but can never grant a new permission or change a tool's mode.

Tools follow the same discipline. Write each one as a verb, a system and a mode. Look up an order in the order system, read only. Draft a reply in the support inbox, draft only. Issue a refund in the payments system, execute, which should normally sit under an approval point. OWASP's authorization guidance points toward least privilege and enforcement of access decisions outside the model itself, so the scope should name the service identity and avoid vague words like admin or full access.

Illustrative scenario: a hypothetical regional parts distributor wants an agent to answer order status questions. Its scope lists two memory sources, the order table with read access to status and shipping date fields, and the approved returns policy document. Its tools are a read-only order lookup and a draft reply in the support inbox. Refunds and address edits appear under out of scope, and the vendor must confirm in writing that those actions are blocked at the permission level, not only by instruction.

A tool listed without a mode is a tool that nobody has decided about yet.

Approval points and success criteria you can observe

An approval point is a condition, not a general sense of caution. Compare two versions. The first says the agent should be careful with customers. The second says the agent pauses and requests approval before sending any message to an address not already in the customer record, before any change to an order, and before any quote above the limit the sales lead sets. Only the second can be tested, and only the second tells a vendor what to build.

Start approval thresholds tight and loosen them with evidence. A narrow first boundary lets the team see what the agent actually proposes, which decisions reviewers accept without edits, and where the agent misreads the situation. The approval thresholds guide explains how to raise limits step by step rather than all at once.

A success criterion has four parts: the measure, the record it comes from, the baseline period and the review date. The record matters most. If the only evidence of success is a feeling that the agent helps, the criterion is not yet written. Good candidates are records you already keep, such as response times in the ticketing system, the share of drafts that staff edit before sending, or the number of exceptions routed to a human. Set targets from your own baseline rather than from a published benchmark. Anthropic's guidance on evaluating agents recommends a fixed set of representative tasks that are rerun after each change, which turns the criterion into a repeatable test instead of a one-time demo.

How a managed engagement keeps the scope current

Be Human's first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. The scope page is the reference point for all of them. When a tool, memory source or threshold changes, the change is recorded against the page, and permissions are reviewed against it on a regular schedule. The change request process shows how that approval works in practice, and the quarterly permission review shows what evidence to keep.

The page is not frozen. The business changes, products change, and people leave. The discipline is to update the page first and the agent second, so that the written scope always matches what the agent can actually do.

A checklist before the first vendor call

Run this list against your draft page. If any answer is no, resolve it or mark it as an open question before the call.

  • The role has one owner, one outcome and one channel.
  • Every memory source is named, with its fields, date range and read or write flag.
  • Every tool is listed with its mode: read, draft or execute.
  • The service identity and its access are written down, not described as admin.
  • Each approval point is a condition that a person can test.
  • Out-of-scope actions are written as plainly as the in-scope ones.
  • The success criterion uses data you already collect.
  • A named person can switch the agent off, and you know how.

Send the completed page to each vendor and ask for written answers against it, line by line. Proposals that answer in general terms have not yet shown fit. The questions on choosing an implementation partner can help you judge who will own the result after the build. If you are not yet sure which task on your list deserves to be the first one handed to an agent, take the free Be Human AI diagnostic to find it.

Frequently asked questions

How long should an AI scope statement be? It should fit on one page, because a page forces a decision about each field. If a field needs several paragraphs, the role is probably too broad, and splitting it into two narrower roles is usually the better path. Keep the language plain so that a business owner and a technical lead read the same thing.

Who should sign off on the scope before a vendor sees it? The business owner who accepts the outcome should approve the role, the success criterion and the out-of-scope list. A technical lead should confirm that the memory sources, tools and service identity are real and reachable. Both signatures matter, because a scope that only the technical side understands will drift toward whatever the platform makes easy.

Can I use the same template with any AI platform? Yes, the template describes your requirements rather than any product's features. Ask each vendor to map every line to a specific capability, and treat any line that cannot be met as a gap to weigh in the decision. A gap that is stated early is far cheaper to handle than one discovered after the agent is live.

What if we do not have a success metric yet? Use the records you already keep and set a baseline period before setting any target. The first weeks of an agent are often better spent measuring how the work is done today than defining a number in advance. Once the baseline exists, the first review date can compare the agent's results against it.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →