Before deploying AI agents, check the information they will use
12 November 2026
Check AI data readiness before deploying agents: document ownership, duplicates, permissions, retention and source accuracy for Australian SMBs.
Iron Mountain’s latest report points to information management problems as a barrier to AI deployment. For Australian SMBs, the practical lesson is to check AI data readiness before connecting an agent to shared folders, customer records or finance systems.
In CRN Australia’s coverage of Iron Mountain’s report, ANZ managing director Dominic Del Giudice describes difficulties finding and preparing suitable business information. The report’s findings are a warning, not a prediction of your own project’s outcome.
Start with a bounded workflow, approved sources and named reviewers. Agents should retain human approval and operate in a private Australian-hosted environment, but those safeguards do not make inaccurate or overexposed records safe to use.
Start AI data readiness with one workflow
Do not make cleaning every business record a prerequisite for a useful pilot. Define the information needed for one task, then assess that collection thoroughly.
For example, a Knowledge and SOP Agent might answer staff questions using approved procedures. Its initial scope could exclude customer files, draft policies and archived instructions. An AP/AR Automation Agent needs a different collection: authorised supplier records, purchase orders, invoices and matching rules.
Write down:
- The task the agent may perform and what is outside scope.
- The systems and collections it may read.
- The drafts or recommendations it may produce.
- The person who approves each consequential action.
- The conditions that require it to stop and request review.
This turns AI data readiness into a manageable operational task rather than an undefined clean-up project.
1. Give every approved collection an owner
A folder administrator is not necessarily qualified to confirm that its contents are correct. IT can manage storage and access; a business owner must decide which instructions, records and rules are authoritative.
Create a small source register. For each collection, record its location, business owner, technical custodian, sensitivity, review date and permitted use. Name a backup owner so staff changes do not leave the collection unmanaged.
For procedures, record who approves revisions and when each version becomes effective. For finance data, identify who can approve changes to supplier details and matching rules.
Deployment check: ask the owner to approve the exact collection the agent will use. If nobody can confirm whether a document is current, exclude it until someone resolves that uncertainty. A recent file modification date is not evidence of business approval.
2. Resolve duplicates without destroying records
Duplicate information creates more than storage overhead. Two customer entries can split account history. Multiple policy copies can produce contradictory answers. A scanned invoice and an emailed copy can be mistaken for separate transactions.
Separate exact copies from conflicting versions. Identical files may be straightforward to identify; documents with the same title but different instructions require a business decision.
Your checklist should include:
- Stable identifiers for customers, suppliers, orders and invoices.
- Rules for selecting the authoritative record.
- Human review of uncertain record matches.
- Links from superseded procedures to their replacements.
- An exception list for unresolved conflicts.
Do not automatically delete everything labelled duplicate. Some copies form part of an audit trail or must be retained. Where appropriate, exclude superseded material from the agent’s active search collection while preserving it in a controlled archive.
Deployment check: present a known duplicate and a conflicting version. Confirm the workflow identifies the issue rather than silently combining incompatible information.
3. Test permissions through the agent, not just the folder
An agent can make previously obscure information easier to discover. Private hosting does not correct overly broad access permissions.
Check the agent’s identity, connector permissions, search index and output destinations. Decide whether access follows the requesting user’s permissions or a tightly limited service identity. Either approach needs explicit boundaries and testing.
Review shared links, inherited permissions, former staff accounts and external sharing. Keep payroll, sensitive personnel records and unrelated customer information outside the pilot unless specifically authorised.
Test with separate accounts representing different roles. Can a sales user obtain finance-only information through an answer, citation or generated report? Can a cached response reveal material after access has been revoked?
Deployment check: verify both allowed and denied access, including after permission changes. Record the results. Our AI agent identity security checklist covers the account, permission and approval controls that support this work.
4. Set retention rules for originals and derived data
AI data readiness includes knowing when information should stop being used. Keeping everything indefinitely can increase exposure and leave outdated material available for retrieval.
Identify applicable legal, contractual and operational retention requirements. These vary by record type and business circumstances; obtain appropriate advice where obligations are unclear.
Map more than the original files. A deployment may also hold extracted text, search indexes, embeddings, temporary uploads, conversation logs and generated drafts. Each needs an agreed retention and deletion process.
Define what happens when a source expires, is superseded or becomes subject to a legal hold. Distinguish removing something from active retrieval from deleting the underlying business record. Check backup retention separately rather than assuming deletion instantly removes every copy.
Deployment check: withdraw a test document and confirm it stops appearing in new answers within the agreed update window. Verify how derived copies are handled and who receives alerts when removal fails.
5. Check source accuracy and traceability
A readable document is not necessarily a reliable source. Scanning errors, missing attachments, outdated price lists and ambiguous spreadsheet headings can all distort an agent’s output.
For each approved source, check:
- Whether it came from the expected system or authorised owner.
- Whether its effective date and status are clear.
- Whether extracted amounts, dates and identifiers match the original.
- Whether required fields and supporting attachments are present.
- Whether an answer can point back to the relevant record and version.
For external monitoring, approve the source list and record retrieval dates. Treat webpage content as evidence to assess, not instructions the agent should follow.
Deployment check: use examples with missing fields, poor scans and contradictory instructions. The agent should flag uncertainty for human review rather than invent missing values. Approval screens should show the supporting evidence, not just a polished recommendation.
Make the checklist a deployment gate
Before expanding access, assemble a short evidence pack: approved source register, duplicate-handling rules, permission test results, retention settings and accuracy checks. This makes AI data readiness visible to operations, finance and IT leaders.
Run representative and exception-heavy tasks in draft-only mode. Track incorrect answers, unresolved conflicts, reviewer corrections and preparation effort. Agree acceptable results before testing, and pause expansion if a critical access or accuracy check fails.
Human approval must be an actual workflow control. Sending messages, changing records or initiating financial actions should require the designated reviewer’s authorisation. Keep an audit trail of the source evidence, proposed action and approval decision. See our AI governance controls guide for ways to test that policy is enforced.
Recheck the collection when connectors, permissions, business rules or source systems change. AI data readiness is an ongoing responsibility, not a one-off migration task.
Prepare the information before expanding automation
A useful first deployment has a narrow purpose, trustworthy sources and a clear person accountable for approval.
Tech Engine can help assess your information and deploy managed AI agents that integrate with existing systems, retain human approval and run in a private Australian-hosted environment. Discuss your first workflow at aiaas.au, call 1300 088 324 or email sales@techengine.au.
Want this applied to your business?
Request an AI Blueprint and we will map the processes worth automating first.
Request an AI Blueprint