← Back to perspectives
/Intellicon FDE Team

Using AI to Compare Supplier Quotes: How Can Procurement Keep Placing Orders When It Gets Things Wrong or Goes Offline?

Australia’s critical infrastructure guidance recommends retaining manual processes so AI failures do not disrupt essential operations. Prepare the information needed for verification and alternative procedures first, then run separate drills for incorrect recommendations and tool unavailability.

AIEgentWrX人機協作導入
Using AI to Compare Supplier Quotes: How Can Procurement Keep Placing Orders When It Gets Things Wrong or Goes Offline?

Suppose a procurement team member is about to submit an urgent order, only to discover that the AI compared quotes using the wrong specifications, or that the platform has suddenly become unavailable. Can their manager catch the error in time and have them check the original quotes and obtain the necessary approvals? If the procedures and information they need are accessible only through AI, they may still be unable to continue working, even if the documentation says they can switch to manual processing.

At Intellicon Solutions, we believe acceptance testing for an AI Agent should cover two separate scenarios: incorrect recommendations and tool unavailability. When a recommendation is wrong, test whether staff can identify the problem and stop subsequent steps from proceeding. When the tool is unavailable, test whether they can complete predefined essential tasks without AI. Confirming that AI can produce answers under normal conditions is not enough to establish what to do when something goes wrong.

Before Planning a Manual Fallback, Define Which Tasks Cannot Stop

According to iThome, the responsible authority within Australia’s Department of Home Affairs has updated its guidance on critical infrastructure risk management programs to address the operational risks of AI. The authority noted that decision-support tools could disrupt critical services or expose sensitive information if they provide incorrect recommendations, malfunction, or are misused by attackers. It therefore recommends that staff verify recommendations before acting on them and retain manual ways of carrying out the work.

This guidance concerns Australia’s critical infrastructure. The report does not explain whether it applies to Taiwanese businesses, nor does it provide data on the effectiveness of the changes. We therefore recommend using its areas of focus to guide internal checks: business managers should first list the tasks that must still be completed during an outage, then ask IT to identify what is needed to make that possible. Foreign guidance should not be used on its own to determine whether a company meets its regulatory obligations.

In a hypothetical procurement quote-comparison scenario, a manager could start by setting clear requirements: urgent orders must still undergo specification checks, delivery-date confirmation, and approval, while analysis reports for non-urgent orders can wait. Next, identify which steps involve AI, who would be affected by errors, and which steps staff can complete using the information they already have, even without AI recommendations. Start with one critical operation and define what counts as completion during an outage and how long it can wait, so IT knows what to test.

Human Approval Must Specify What to Check, Not Just Require a Manager’s Sign-Off

If staff are expected to verify information before acting, the operating procedure must specify what they need to check. For example, before approving a quote comparison, a procurement manager should have access to the suppliers’ original quotes, applicable specifications, and delivery terms so they can check the AI’s summary item by item. If they can see only the summarized conclusion, they cannot verify it independently.

Approvers also need to know which step to return the work to when specifications do not match or information is missing, whom to ask for additional information, and which activities must remain on hold until that information is available. Assign approval responsibility to someone familiar with the operation, and designate a backup so other staff do not skip checks in a rush to finish when the primary person is absent.

EgentWrX workflows allow several tasks to run in sequence, with human approval required before proceeding to the next task. The workflow pauses and resumes only after approval. This gives staff an opportunity to check the content before execution, but human approval does not guarantee accuracy, nor does it replace the manual procedures a business needs when AI is unavailable. When deciding which steps require approval, also document which information must be checked, when the work must be sent back, and who is responsible. Then deliberately introduce incorrect recommendations to test whether staff can identify and stop the problem.

Can Staff Still Access the Information They Need When AI Is Disabled?

A manual fallback requires staff to retain access to operating procedures and essential information when AI is unavailable. Prepare documents and forms that can be opened without going through the AI platform, identify who maintains them and when they were updated, and ask the people responsible for the work to confirm that they can actually open them. If copies of data are used, document how to identify the correct version so staff do not continue working with outdated quotes.

IT should also map the connections between the AI platform, internal databases, external services, and existing operational systems. For each connection, record its purpose, which data it needs to read or write, who is responsible for it, and which tasks would be affected if it were interrupted. For example, after AI connections are disabled, can procurement staff still log in to the existing system to look up orders? If those queries still depend on the same disabled service, that dependency must be resolved before manual processing can continue.

With EgentWrX, for example, the platform can read company databases through internal network connections; its internal company database integration is read-only and cannot write data. Its outbound allowlist denies all external connections by default, requiring businesses to explicitly allow each domain they need to connect to. These settings can restrict access, but businesses should not assume that they have achieved network isolation or that critical systems can operate independently of AI. IT should still disable AI-related connections in an environment where the impact can be controlled and verify which tasks the existing systems and manual processes can actually complete.

Before Going Live, Run Separate Drills for Incorrect Recommendations and Tool Unavailability

Business managers and IT should plan the drills together, agreeing in advance on which data to use, which operations may be affected, and how to restore normal operation afterward, rather than directly disrupting production work. Acceptance test records should document what staff actually completed, not simply include a checked box confirming that fallback documentation exists.

The first drill can present a quote-comparison recommendation with incorrect specifications or omitted terms. Observe whether the approver checks the original information, identifies the errors, and stops further processing. If the approver cannot spot the problem, revise the checklist or provide the missing information needed for verification. Adding another signature is not enough.

In the second drill, have the staff responsible complete essential operations using the established procedures without access to AI. Confirm who declares the switch to manual processing, who takes over, and what records must be kept at each step. Once the system is restored, designate someone to reconcile the work completed during manual processing so orders are not submitted twice and outstanding items are not missed.

Bring the results of both drills to the next go-live meeting. Discuss which information is still accessible only through AI and who will address that, then decide which operations should remain outside AI’s scope for now and when to retest after corrections. Managers should determine the scope of the rollout based on the work actually completed during the drills, rather than approving it solely on the strength of a demonstration under normal conditions.

FAQ

Is human approval enough to manage the risks of AI errors or unavailability?

No. You also need the original information, a checklist, and a defined process for sending work back before you can meaningfully test whether staff can stop errors. If the platform is unavailable, the approval workflow itself may also be inaccessible, so businesses must separately rehearse manual procedures that do not depend on AI.

What should a manual fallback include first?

Start with operating procedures, essential information, and a list of people who can take over, all accessible without the AI platform. Designate who can declare the switch to manual processing. Then have the staff who actually perform the work follow the documentation to complete essential tasks. Confirm that the data versions are correct, access permissions are sufficient, and backup staff can take over, then resolve the problems found during the drill.

Must Taiwanese businesses comply with Australia’s updated guidance?

Its applicability cannot be assumed. The report discusses Australian critical infrastructure guidance and does not explain whether it applies to Taiwanese businesses. Companies can use manual verification and fallback drills to check whether their own procedures work. Their compliance staff should determine which regulatory requirements apply to the company.

References

30 minutes to map out which work to hand to AI first

Want every employee to have their own AI teammate?

A consultant will be in touch shortly.

ISO/IEC 27001:2022 Certified ISMS