Many AI projects are really information projects. The business already has the answers, but they are split across email, spreadsheets, cloud drives, a CRM, an online store, accounting software and the memory of the person everyone asks. An assistant built on top of that mess will reproduce the mess faster.

MapIdentify the systems, owners, sensitivity and source of truth.
ConnectGive the workflow only the access it needs and preserve provenance.
TestUse real questions to verify answers, abstentions and boundaries.

“We have the data” is not the same as “the system can use it”

Usable information is current, findable, permissioned and understandable in context. A policy without an owner or revision date may be available but still unsafe to use. A customer table with three versions of the same person is structured but not dependable. A folder full of historical contracts is searchable but may be far too broad for a customer-service helper.

NIST’s AI Risk Management Framework asks organizations to understand intended context, human oversight, data suitability, system limits and how a result will be evaluated. Its generative-AI profile also addresses plausible but incorrect output, privacy risks and information-security concerns. In plain language: know the job, know the material and test the edge cases.

1. Find the systems that actually run the business

Start with an information inventory. This is not a giant technical diagram. A working spreadsheet is often enough.

SystemEmail, file storage, CRM, commerce, accounting, support, calendar or spreadsheet.
Business purposeWhat work depends on it and what would break if it were wrong?
OwnerWho can explain the fields, approve access and resolve a conflict?
SensitivityPublic, internal, customer, employee, financial or otherwise restricted.
FreshnessHow often it changes and how someone knows a record is current.
RetentionWhat should be kept, archived or disposed of under the business’s real obligations.

The Federal Trade Commission’s small-business privacy guidance recommends knowing what personal information the business keeps, retaining only what it needs, protecting it and disposing of it safely. Those are sensible questions before any AI connection is enabled.

2. Name the source of truth

For each important subject, write down which system wins. The commerce platform may own product availability. Accounting may own recognized revenue. A policy folder may own the approved return language. The CRM may own the active sales stage.

If the team cannot agree on the source, that is the work. Do not ask the model to reconcile business rules that the business has not decided.

Reduce contradiction before consolidating

  • Archive superseded documents instead of leaving “final,” “final-2” and “use-this-one” together.
  • Add an owner, status and revision date to operating policies.
  • Standardize basic names and identifiers across exports.
  • Document exceptions that live only in someone’s memory.
  • Do not pour every historical file into one undifferentiated AI index.

3. Decide what the AI may see

Access should follow the workflow, not the curiosity of the team. A product-description helper may need approved product facts and brand guidelines. It does not need payroll. A meeting follow-up tool may need one project folder and a calendar. It does not need every shared drive.

Start read-only.

Let the first version find, summarize and draft. Add write access only after the answers, permissions, logging and review path are dependable.

For connected business systems, use the provider’s strongest reasonable authentication, preserve logs and regularly review who still needs access. CISA’s small- and medium-business resources emphasize multifactor authentication, logging, backups, encryption and secure cloud configuration.

4. Keep the source attached to the answer

A useful business assistant should show its work. Whenever possible, the result should link to the underlying document or record and expose the source date or owner. When two sources conflict, it should surface the conflict. When evidence is missing, it should say so.

This is more important than making every answer sound confident. A short answer with a current source is operationally useful. A polished answer with unknown provenance is a liability.

5. Build a small evaluation set

Before the assistant is broadly available, write a representative set of questions and expected outcomes. This is how the business turns “it seems pretty good” into something it can inspect.

  • Known-answer cases: questions with a current, approved answer.
  • Conflict cases: questions where an old and new document disagree.
  • Abstention cases: questions the assistant should refuse or route to a person.
  • Permission cases: requests a user should not be able to make.
  • Freshness cases: questions that depend on today’s inventory, status or policy.

Retest the same set after sources, permissions, prompts or models change. NIST’s AI RMF playbook treats measurement, testing and ongoing monitoring as part of the lifecycle rather than a one-time launch task.

A practical first deliverable

For many small companies, the first valuable “AI build” is a five-part operating foundation:

  1. An inventory of the systems and important information.
  2. A source-of-truth map for the core business subjects.
  3. An access matrix that says who and what may use each source.
  4. A small evaluation set built from real questions and failure cases.
  5. One bounded, read-only assistant that cites where each answer came from.

That foundation is not glamorous. It is what makes the next workflow easier to trust, easier to maintain and easier to measure.

Sources and further reading