Personalized AI infrastructure
Superintelligence Company Checklist: Legal, Data and Access on Day One
A day-one checklist for legal terms, data ownership and access grants before any AI agent touches company records, with the steps to take in order.

A superintelligence company checklist is a short, ordered set of decisions a business settles before any AI agent reads, writes or acts on company records. On day one, that means confirming the contract terms for data use, retention and exit, deciding who owns the records and the agent's memory, and granting each permission to a named human owner with a narrow scope. Work through the list in order, because each step depends on the answers from the one before it.
Key takeaways
- Settle data use, retention and exit terms in writing before any connection is made.
- Name one accountable owner for each data source, each memory store and each agent permission.
- Grant access by task, start with read-only scopes, and log every grant and its reason.
- Write approval boundaries in plain language before the first live run, not after the first mistake.
- Define one observable success check so the first review produces evidence instead of opinions.
What belongs on the legal checklist before day one?
The legal work is not a formality to hand to procurement at the end. It determines what your company is allowed to connect, where information will live and how you leave if the relationship ends. Ask counsel to review the following points, and keep the answers in a file that operations can find later:
- Who is the contracting party, and which subprocessors or model providers receive your data?
- Are your inputs, prompts and outputs excluded from training shared models, and is that stated in the contract rather than a web page?
- Where is data stored and processed, and does that match the rules that already apply to your customers or your sector?
- How long are conversation logs, memory entries and tool call records kept, and who can delete them on request?
- What happens on exit: can you export memory, configuration and logs in a usable format, and is deletion confirmed in writing?
- What notice do you receive after a security incident, and within what window?
- Who may approve changes to the agent's model, tools or data sources without your sign-off?
This list does not replace legal advice, and the answers depend on your jurisdiction and industry. The point for an operations leader is simpler: if a term is unclear, the agent should not be connected yet. Frameworks such as the NIST AI Risk Management Framework are useful for structuring this review because they ask who is responsible for each risk, not only whether a risk exists.
Who owns the data, the memory and the outputs?
Most companies treat "the data" as one thing. For agents, it helps to separate at least four categories, because each one has a different owner and a different exit path:
- Source records: customer files, ledgers, contracts and inventory lists that already live in your systems.
- Agent memory: the notes, summaries and retrieved facts the agent keeps between sessions.
- Operating know-how: prompts, playbooks, decision rules and runbooks that describe how the work is done.
- Outputs: drafts, reports, messages and records the agent creates on your behalf.
For each category, write down the owner, the system of record, the retention period and the export format. Source records usually stay with the system that already holds them. Memory and operating know-how are the parts most likely to be lost or locked up if nobody decides on them early, so they deserve an explicit line in the contract and an explicit owner inside the company. If the agent learns how your business handles exceptions, that knowledge should be stored where you can read it, copy it and transfer it. The same logic applies to intellectual property, which is covered in more depth in Superintelligence and Intellectual Property: Protect What You Know.
Every data category needs one owner, one home and one exit path before the agent is connected.
Clean records also matter for ownership. An agent cannot be accountable for a figure that exists in three conflicting spreadsheets. If your sources are not ready, the work in Data Readiness Before Superintelligent Agents: A Six-Point Audit is the step to complete first.
How do you grant access without giving the agent the whole company?
Access grants are where most day-one risk sits. The safe default is least privilege: the agent receives only the scopes required for one defined task, and nothing else. In practice this means separate service accounts for the agent rather than borrowing a person's login, read-only scopes for every system at first, and write access limited to a dedicated folder, queue or draft area. OAuth 2.0 gives you a way to grant scoped, revocable tokens to a specific application, and the OWASP authorization guidance is a reasonable reference for checking that each permission is enforced where the data lives, not only in the agent's instructions.
Two details are easy to skip. First, tokens should expire or be rotated on a schedule you can name, so that a forgotten credential does not become a permanent door. Second, every tool the agent can call should be listed with what it can read, what it can change and who approved it. Tool descriptions also shape behavior, so the wording should be specific about the action and its limits. Treat any text the agent reads from outside sources, such as emails or web pages, as untrusted input, since instructions hidden in that text are a known risk that the OWASP guidance on prompt injection describes.
Step by step: the day-one sequence
Use this order on the first day of any agent engagement. Each step produces a document or a decision, and the agent should not receive a new permission until the previous step is signed off:
- Step one: confirm the contract terms on data use, retention, subprocessors and exit, and file the signed copies.
- Step two: list every data category the agent might touch, with an owner and a system of record for each.
- Step three: choose the single task the agent will perform first and write down what it must never do.
- Step four: create a dedicated service account and grant read-only scopes limited to that task.
- Step five: define the approval boundary, meaning which actions require a named human to say yes before they happen.
- Step six: set the logging rule, so every tool call, data read and write is recorded with a timestamp and a reason.
- Step seven: write one success check that someone outside the build team can verify in the first review.
- Step eight: schedule the first permission review and name the person who will run it.
If you cannot complete step two without a long search, that is useful information. It tells you the first project is a data project, and it protects you from granting access to systems you cannot yet describe. The governance structure that keeps these decisions owned is described in Superintelligence Governance: A Committee Charter for Small Companies.
What does a sound approval boundary look like in practice?
An approval boundary is a written rule that separates actions the agent may take alone from actions that wait for a human. Good boundaries are concrete. "Use good judgment on customer messages" is not a boundary. "Drafts may be created in the outbox folder, but nothing is sent until the account owner approves it" is. Start with a narrow set of autonomous actions, and move a boundary only after the logs show the action behaves as expected for a sustained period.
Write the boundary in the same document as the access grant, so that the permission and the restriction are reviewed together. When the boundary changes, the change should follow a recorded request and approval, which is the discipline described in How to Request and Approve Changes to a Managed AI Agent.
Illustrative scenario: a quote drafting agent in a services firm
The following scenario is illustrative and does not describe a real client. A small professional services firm wants an agent to draft quotes from its price list and past proposals. The owner's day-one review produces three decisions. The agent reads the price list and the proposal archive, both of which the firm owns and can export. It writes only to a drafts folder with a dedicated account. Any quote that leaves the drafts folder requires the account lead to approve it by name. The success check is that a draft quote matches the approved price list in every line on a sample reviewed each week. If the check fails, the agent's write access is removed before anyone debugs the prompt.
Notice what the scenario avoids. The agent does not receive the whole file server, the accounting system or the email inbox on day one. Each of those can be considered later, through the same checklist, when there is evidence that the first task works.
How does Be Human's managed approach handle day one?
Be Human designs and operates personalized AI infrastructure in which an agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion, so the checklist above becomes a written specification rather than a meeting topic. Managed operations then keep the work honest over time through monitoring, recurring evaluations, permission review and controlled changes, so that access granted on day one is still justified months later. The aim is not to make the agent autonomous as fast as possible but to make every permission explainable to the person who owns it.
If you are unsure which task should go first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
What is a superintelligence company checklist? It is an ordered set of day-one decisions covering legal terms, data ownership, memory, access grants and approval boundaries, completed before any agent works with company records. The list is useful because it forces owners to be named and rights to be written down while changes are still cheap. Operations leaders can use it as a gate, where each step must be signed off before the next permission is granted.
Which legal terms matter most before connecting an AI agent? The most important terms cover whether your inputs are excluded from training shared models, where data is stored, how long logs and memory are retained, and how data and configuration are exported and deleted at exit. You should also know which subprocessors receive your information and what notice you get after an incident. Ask counsel to confirm these points for your jurisdiction, since the requirements vary by country and sector.
How should access be granted to an AI agent on the first day? Use a dedicated service account, start with read-only scopes limited to one task, and add write access only to a narrow folder or queue that you control. Tokens should expire or rotate on a known schedule, and every tool call should be logged with a reason. Expanding access should require a recorded request and a named approver, not a verbal agreement made in a meeting.
How do you know the first agent deployment is working safely? Define one success check that someone outside the build team can verify, such as a weekly sample of outputs compared against an approved reference, and review the logs alongside it. Hold a permission review on a fixed date, remove any access that is no longer needed, and move approval boundaries only when the evidence supports it. A managed operating model keeps this review recurring rather than leaving it to a single launch week.
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.


