Personalized AI infrastructure
Data Readiness Before Superintelligent Agents: A Six-Point Audit
Before an AI agent reads your company data, audit six things: source systems, record ownership, freshness, and access. A practical guide for business leaders.

Data readiness for AI agents means every dataset an agent will read has a known source, a named owner, a stated freshness, and access rules that someone approved on purpose. Before any agent reads company records, a business should complete a six-point audit covering source systems, systems of record, ownership, freshness, access and traceability. The audit also shows which first task is safe to hand over.
Key takeaways
- An agent is only as reliable as the records it reads, so audit those records before you choose tools or models.
- Every dataset needs one named person who can approve access and answer questions about what the data means.
- Freshness should be written down for each source, as a stated interval and a visible timestamp, never assumed.
- Agent access should follow the same role boundaries a careful employee would have, and it should be reviewed on a schedule.
- If you cannot trace an answer back to a source and a date, the data behind that answer is not ready for an agent.
What data readiness means for an agent
Having data and being ready to use it are different things. Most companies hold plenty of records across spreadsheets, a customer system, a ticketing tool, accounting exports and shared folders. A person reading those records can ask a colleague what a column means, notice that a price list looks old, or skip a folder that seems abandoned. An agent cannot make those judgments reliably unless the surrounding system supplies the context.
This matters most in retrieval-based designs, where the application fetches relevant documents at the moment of a question and the model writes an answer from them. Google Cloud's reference architecture for retrieval-augmented generation describes this pattern in detail. The practical consequence is simple: the answer can be no better than what the retrieval step can reach, how current that material is, and whether it is the right version. Readiness is the work of making those three things known before the agent starts.
Readiness is also not a one-time cleanup. Data changes every day, people change roles, and integrations break quietly. The audit gives you a baseline, and the operating rhythm keeps it true.
Points one and two: source systems and systems of record
The first point is an inventory of every place where a fact about the business lives. Include the obvious systems and the informal ones, such as a shared spreadsheet that the operations team updates each Friday or a mailbox where supplier terms arrive as attachments. Informal sources are often the ones an agent would read most confidently and understand least.
For each source, record the following:
- The name of the system or file location, and whether it is a live system or an export.
- What kind of facts it holds, such as customers, orders, contracts, schedules or policies.
- Whether it contains personal data, financial data or confidential material.
- How records enter it, whether by people typing, by an integration, or by an upload.
- Whether it is the only place that fact exists.
The second point is the system of record. When two sources disagree about a customer's billing address, one must win, and the business should decide which one before the agent is connected. Write the decision down as a single sentence, for example: the billing system is the record for payment terms, and the customer relationship tool is the record for contact names. Without this sentence, an agent will pick whichever source it retrieves first, and two correct-looking answers will contradict each other.
If you are still building the foundation of clean records, the guidance in Data Foundations for Superintelligence: Clean Records Before Agents covers how to tidy the sources you have chosen.
Point three: a named owner for every dataset
Every dataset the agent can read needs one owner who is a business person, not only a technical contact. The owner decides who may read the data, approves changes to what the agent can see, and answers questions such as whether a status field is updated when an order ships or only when it is invoiced. IT or the software vendor owns the system. The business owns the meaning of the content.
The ownership check fails in predictable places. Contract folders created by a departed employee, customer lists that three teams edit with different conventions, and reports generated by a script that nobody has reviewed in a year are all common. Ask, for each source, who would notice first if the content were wrong, and who has the authority to fix it. If the answer is nobody, that source is not ready, and the agent should not read it yet.
Ownership also protects you when the agent gives an unexpected answer. The owner is the person who can check the source, correct it, and decide whether the agent's access should change.
Point four: freshness you can state
Freshness is how old a record can be before an answer built on it becomes misleading. The audit should state this for each source in plain terms. A delivery status that refreshes every few hours serves a different purpose from a supplier catalog that changes twice a year. Neither is wrong, but each needs a rule the agent can follow and a reason the owner accepts.
Write three things for each source: how often it updates, how old it may be before an answer must carry a warning, and what the agent should say when the data is older than that limit. A good answer names the source and the date of the record it used. A poor answer presents a stale figure as current and gives no way to check it.
Freshness also affects the choice of first task. Tasks that depend on information that changes by the minute need live integrations and careful monitoring. Tasks that depend on policies or reference material that changes slowly can start sooner, with a simpler setup.
Points five and six: access and traceability
The fifth point is access. An agent should have the minimum permissions needed for its role, read access before write access, and separate permissions for each source. Authorization should be enforced at the point where data is retrieved, not assumed because the agent sits inside a trusted network. The OWASP authorization guidance describes this principle in general terms, and it applies directly to agents that call tools on behalf of users. Write down which roles may ask which questions and which sources each role can reach. Then compare that map with what the agent can actually access, because the two often differ.
Also treat retrieved content as untrusted input. A document, email or web page can contain text written to instruct the model, and an agent that obeys it has been manipulated. The OWASP prompt injection guidance explains why instructions inside data must be separated from the agent's own instructions. For readiness, the practical step is to decide which sources are trusted for instructions, which are only trusted for facts, and which are excluded from the agent entirely. For a deeper method, see How to Run a Permission Review for Your AI Agents, Step by Step.
The sixth point is traceability. Every answer should record which source was used, which version or timestamp it carried, and which permissions applied. Observability concepts from OpenTelemetry, such as traces and logs that connect a request to the operations behind it, are a useful model here. If a manager asks why the agent quoted a particular delivery date, the business should be able to answer within minutes. If it cannot, the data is not ready, regardless of how good the answers look in a demonstration.
An illustrative scenario and a sequence to follow
Illustrative scenario: a regional equipment distributor wants an agent to answer account managers' questions about open orders and delivery dates. The audit finds three sources. Order status lives in the warehouse system and refreshes several times a day. Contract terms sit in a shared folder with three versions of the same agreement and no owner. Customer contacts live in a sales tool where duplicate accounts are common. The team decides that the warehouse system is the record for order status, with a stated refresh interval and a timestamp on every answer. Contract questions are excluded until a named owner cleans the folder and approves access. Customer contacts are excluded until duplicates are resolved. The first task is narrow, the risk is bounded, and the next task is already defined.
A practical sequence for the audit looks like this:
- List every source that holds a fact the agent might need, including informal files and exports.
- Name the system of record for each fact that more than one source holds.
- Assign one business owner to each dataset and confirm that the owner accepts the role.
- State the freshness rule and the warning the agent must give when data is older than the limit.
- Map which roles may ask which questions, and compare that map with the proposed permissions.
- Mark untrusted content and decide which sources are excluded from instructions.
- Choose the first task that uses only sources that passed every previous step.
- Set the traceability fields each answer must carry, and assign someone to review a sample of answers.
The last step matters because readiness is a judgment that the owner and the operator keep making. The NIST AI Risk Management Framework is a useful reference for organizing that ongoing risk work around governance, mapping, measurement and management, without tying the company to one product.
How Be Human handles data readiness
Be Human designs and operates personalized AI agents that keep their identity, memory, channels, tools, permissions and operating context under human control. In a first engagement, the readiness audit is part of defining the one useful role the agent will perform. The work identifies the sources the role needs, the system of record for each fact, the owner who approves access, the freshness rule for each source, and the traceability fields each answer must carry. Permissions and approval boundaries are set from that map, not from what is technically reachable.
Managed operations then keep the audit current. Monitoring tracks whether sources are updating on schedule, recurring evaluations check whether answers still match their records, integration maintenance handles changes in connected systems, and permission reviews confirm that access still matches the roles. Changes to sources or scope go through a documented request and approval path, so readiness is rechecked when the data changes instead of being assumed once at launch. For what the early period of such a setup involves, see What the First Months With a Managed AI Agent Look Like.
If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Do we need a data warehouse before an agent can read our data? No. Many useful first tasks read from a single operational system or a small set of well-owned files. A warehouse can help when many sources must be combined, but the audit matters more than the storage choice, because an agent reads whatever it can reach.
Can we start with spreadsheets? Yes, if the spreadsheet has a named owner, a clear update routine, a stated freshness rule, and a single role in the system of record decisions. Spreadsheets fail when they are edited by many people with no owner, so the audit will usually reveal whether that condition holds.
How often should we repeat the audit? Repeat the full six points before adding any new source or changing the agent's scope, and review ownership, freshness and access on a regular schedule, such as quarterly for most companies. Monitoring should flag stale data between reviews, so the calendar is a backstop rather than the only check.
What if our data includes personal information? Treat personal data as a separate decision inside the audit. Confirm the legal basis for processing, limit access to the roles that need it, decide whether the agent should see identifiers at all, and record that decision with the owner. Guidance from your national data protection authority, such as Brazil's ANPD for Brazilian companies, should inform these rules alongside your internal policy.
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.


