Is your business data ready for an AI application?
Data readiness depends on the task. Identify the sources an application needs, who may access them, how current they must be and how you will detect a wrong answer or action.
Follow one workflow through its sources
Choose a concrete request: explain an opportunity’s status, prepare an operational report or interpret a leave request. List each source required to complete it. Record the source owner, identifiers, update frequency, access rules and integration method.
A useful inventory distinguishes the authoritative record from supporting context. A meeting transcript may explain a commitment; the CRM records the current opportunity state. The application needs a way to handle a disagreement rather than silently choosing whichever text is easiest to retrieve.
A practical source inventory
- Purpose: which part of the task needs this information?
- Ownership: who can confirm its meaning and correct an error?
- Access: which users and systems may read or change it?
- Freshness: when was it updated, and how stale can it be for this task?
- Integration: how is it retrieved, linked, validated and monitored?
- Failure behaviour: what should happen if it is unavailable or inconsistent?
Separate integration problems from model problems
Missing identifiers, inconsistent definitions and delayed source updates affect the application before the model produces an answer. Test these boundaries directly. Confirm that the right records join together, that missing values are visible and that the application can explain which source supports an answer.
The CloudHero utility data platform case study describes data integration, BigQuery and Dataform in an operational reporting context. Its reported morning coordination window involved several people; that does not by itself establish a number of labour hours saved. The useful lesson for an AI project is the engineering sequence: connect and validate the information before depending on an automated output.
Data work can also produce value independently of AI. A reliable report or reconciled dataset may be the appropriate first deliverable. Preserve that decision when planning the roadmap.
Make access part of retrieval and execution
Define permissions at the source and enforce them in the application’s search and tool layers. Test with users who have different access. Check whether cached results, attachments and derived summaries follow the same restrictions.
Otto in Octora combines structured and full-text retrieval over business context, with source readers and user-scoped tools. Its implementation shows why access and citation handling matter even without claiming a vector-search architecture. Choose retrieval methods for the information and task you actually have.
Reading data and writing a record are separate capabilities. Specify which changes may be proposed, who approves them and which server-side checks apply when the action runs. Include the case where the source record changes after the proposal was prepared.
Use realistic examples before expanding the scope
Build a small set of representative tasks with expected sources and outcomes. Include missing documents, duplicate records, conflicting dates, unsupported requests and restricted information. Evaluate extraction and interpretation separately from the business transaction they trigger.
In an HR workflow, document intake connects extracted information with HR records, permissions and approval rules. Evaluate the workflow against the company’s own documents and source systems.
Turn findings into an implementation plan
For each gap, record the affected task, the owner and the evidence needed to close it. Sequence dependencies: source access before retrieval checks, reliable identifiers before record linking, and permissions before controlled actions.
A readiness outcome can be “proceed with this use case,” “repair these data flows first” or “run a limited experiment.” Our data engineering services cover platforms, integration, reporting and forecasting. An AI readiness assessment connects those findings to an implementation decision.