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

Personalized AI infrastructure

Superintelligence Data Audit: Find Records Agents Cannot Reach

A ranked checklist for finding company records a future superintelligent agent would need but cannot reach, with owners, access tests and approval limits.

Be Human typographic cover with the title: Superintelligence Data Audit: Find Records Agents Cannot Reach
Conceptual illustration created with AI · Be Human

A superintelligence data audit is a structured review of the company records that a future, far more capable agent would need to do useful work, compared against what agents can reach today. The result is a ranked list in which records with the highest business value and the widest access gap come first, each with a named owner, a source system and a permission decision. Run it before choosing infrastructure, because the gaps it reveals decide what a first agent can responsibly do.

Key takeaways

  • Start from the decisions the business makes, then trace the records behind each one.
  • Rank records by the revenue, cost, customer impact or risk they touch, not by how tidy they look.
  • Test reachability for each record: does it exist in structured form, is there an approved connection, and is its scope no wider than the role needs.
  • Assign an owner and a source of truth to every record before any access is granted.
  • Repeat the audit on a schedule, because new tools, new staff and new channels change what exists and who can see it.

What the audit measures

A record is not reachable just because it exists somewhere in the company. A useful audit tests four separate properties: whether the record exists in a durable form, whether it is structured enough for software to interpret, whether an approved connection exists that an agent could use, and whether the permission on that connection matches the job the agent will do. Most gaps sit between these tests. A price list may live in a current, well maintained spreadsheet, yet no connection is approved for it, so the agent cannot read it at all. A complaint history may sit fully inside a helpdesk, but its resolution notes are free text that the agent would have to interpret with no way to confirm accuracy.

Draw the stack before you list records, so every record has a place on the map. A one-page view, such as the one described in Draw Your AI Stack: A One-Page Architecture Map for Leaders, shows which systems hold which records and where the connections would have to run.

Rank records by business value before you touch access

Ranking prevents the common mistake of connecting the easiest system first. Score each candidate record on three questions: how much money, time or customer trust depends on decisions that use it; how often those decisions happen; and how much risk attaches to it, including personal data, contractual obligations and anything that would create liability if an agent misread it. Use a simple scale with three points. Precision matters less than agreement among the people who own the decisions.

Treat risk as a separate flag rather than a reason to drop a record down the list. A high-risk record keeps its place in the value order, but it moves into a stricter approval path. For the wider preparation work, Data Readiness Before Superintelligent Agents: A Six-Point Audit covers the cleanup steps that usually follow the ranking.

The checklist: records to find, in rough order of value

The list below is a starting point for most service and product businesses. Adjust it to your operations, and mark each item as present, partial or missing before you rank it.

  • Customer history across channels: orders, quotes, renewals, complaints and how each was resolved.
  • Pricing rules, discount approvals and quote logic, including the exceptions people actually grant.
  • Contracts and commitments: renewal dates, service levels, notice periods and supplier terms.
  • Decision records: why a request was refused, why a price changed and who approved the exception.
  • Operating procedures as they are practiced, including the steps experienced staff skip or adapt.
  • Exception and incident logs: refunds, rework, missed deliveries and escalations.
  • Support tickets with resolution notes, not only ticket counts or status labels.
  • Role and authority records: who may approve what, and up to which amount or risk level.
  • Financial close data and margin figures by product, client or site.
  • Shared inboxes and channel threads where promises and commitments are made.
  • Approved product claims, brand language and forbidden phrasing.
  • Documents that exist only in personal drives or personal messages.

Several items on this list are not systems at all. Decision records and the reasons behind exceptions often live in messages, and they are among the most useful records a future agent could learn from, which is why they belong near the top.

How to test whether an agent can reach a record

Run this sequence for each candidate record, starting at the top of your ranking. Once owners are identified, it takes little time per record, and it produces the evidence a later permission review will need.

  • Name the decision the record supports, in one sentence.
  • Identify the system of record and the person who maintains it.
  • Check whether the record is structured enough to interpret, or exists only as free text in a document.
  • Check whether a connection exists, and under which identity an agent would connect.
  • Compare the scope of that connection with the role's needs, granting only the narrowest access that lets the task finish.
  • Label the gap as missing, unstructured, unreachable or reachable but over-permitted.
  • Name the fix and the approval owner before any access is granted.

Be strict about the identity an agent uses. A connection that runs under a shared administrator login is over-permitted by default, even when the agent needs only one table. The OWASP authorization guidance favors deny-by-default and least privilege, and the NIST AI Risk Management Framework asks organizations to document who is accountable for how an AI system behaves.

Illustrative scenario: a rental company's first audit

The scenario below is illustrative and does not describe a real client. Imagine a ten-person equipment rental business whose owner wants an agent to draft quotes for standard items. The audit finds three things. The price table lives in one spreadsheet maintained by the operations lead. It is current, so it ranks high and becomes reachable after a small structural cleanup. Damage fee decisions are made in a shared messaging thread, which makes them high value but unreachable and high risk. Return condition notes exist only as free text in the booking system, which is partly reachable but over-permitted, because the only available connection uses an administrator login.

The owner decides that the agent's first scope is quote drafts for standard items, using the price table only. Pricing exceptions and damage fees are routed to the operations lead for approval. Damage fee rules move onto the checklist as a capture task rather than an access task, because no connection can fix a decision that was never written down.

Four quick rules help when the audit turns up something unexpected:

  • If a record exists only in one person's messages, treat it as missing and schedule a capture step before any connection work.
  • If two systems disagree on a price or status, name one source of truth before connecting either.
  • If the only available connection uses a shared login, treat the record as over-permitted until a narrower identity exists.
  • If a record holds personal data about customers or staff, involve the person responsible for privacy before access, and check the rules that apply where you operate.

Where Be Human fits, and the first step

Be Human treats the audit as part of the first engagement rather than a separate project. The engagement defines one useful role, the memory and tools that role needs, its approval boundaries and an observable success criterion, and the audit determines which records can responsibly feed that role on day one. Records that fail the reachability or permission tests go onto a scheduled backlog instead of being forced into the first release. During managed operations, monitoring, recurring evaluations and periodic permission reviews keep the access map accurate as systems and people change. The early months of that work are described in 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

How long should a superintelligence data audit take? The length depends on how many business decisions you want an agent to support, but a useful first pass covers only the top few records by value, not every file the company holds. Keep it inside a defined scope, review the findings with the record owners, and widen it once the first role is running and its boundaries are clear.

Who should own the audit in a small company? A business owner or operations leader should own the priorities, because the ranking reflects which decisions matter most. Each record needs a named person who confirms what it means and whether it is current, and someone technical should verify connections and scopes. A short governance structure, such as the one in Superintelligence Governance: A Committee Charter for Small Companies, keeps those roles written down.

Do we need to clean every record before an agent can start? No. Clean the records the first role needs and leave the rest on the backlog. Cleaning a record that no decision depends on usually costs time without changing what the agent can do.

What if a valuable record contains customer or employee personal data? Treat it as higher risk from the start, limit the fields an agent can see, and involve the person responsible for privacy before any connection is created. Check the rules that apply in your country, and in Brazil that means consulting the guidance of the national data protection authority, the ANPD.

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 →