Skip to main content
Boermans Digital
Main menu

Best practiceIntelligent Automation

AI workflows and agents with human review

A message is classified as an address change. The summary sounds convincing, but the sender actually wants one delivery sent to a temporary address. If that distinction disappears, a polished summary can trigger the wrong action.

Example workflow with human review.

Key terms explained

A workflow is a sequence of agreed steps. AI might summarise or categorise a message within it. A confidence score expresses how certain a system estimates an output to be; it is not proof that the output is correct.

Define what the system may prepare

Start with suggesting a category for an incoming message, for example. A suggestion must not quietly become permission to amend customer records. Identify actions outside the task and who has authority to perform them.

If structured fields or clear rules already solve the problem, they may be simpler. AI becomes a candidate where meaning or variation in the input matters. Permitted actions and access rights still need to be defined in advance.

In this article, an AI agent means software that uses an AI model to choose a next step and call available software functions. The label does not tell you what it may do in your business. Separate three decisions: reading information, preparing a proposal and making a change. Suggesting a new delivery address does not authorise changing it.

A fixed route or freedom to choose the next step?

In a defined AI workflow, the steps are set in advance: read a message, prepare a proposal and show it to a reviewer. An AI agent can choose a next step or tool within its permitted capabilities, such as looking up further information. That freedom is useful only when the task needs it.

Consider a request to change a delivery address. If the information and approval needed are known in advance, a fixed route may be enough. If the order must first be identified, investigate whether additional lookup capabilities help. Do not automatically grant permission to change the order as well. Compare the extra steps and controls with the simpler route.

Discuss existing software, AI and workflow options and check which business data you may use.

Separate suggestion, review and execution

On a narrow screen, scroll the table sideways. With a keyboard, focus the table and use the arrow keys.

Fictional example: a message about a delivery address.
StageAssistanceResponsibility
PrepareSuggest a category and relevant passageKeep the original and uncertainty visible.
ReviewShow the suggestion beside source informationAn authorised person checks intent and permission.
ExecuteCarry out only the approved changeAccess controls and process conditions still apply.
VerifyCompare the outcome with the instructionFollow up missing or incorrect processing.

Enforce permissions in the connected systems, rather than relying only on an instruction to the AI. A task that only needs to look up information does not need permission to amend orders. The receiving application should check whether the requested action is authorised. Available permissions and approval steps depend on the product and configuration. OWASP identifies excessive functions and permissions as a risk in AI applications. OWASP: Excessive Agency

Make human review practical

Give the reviewer the original message, relevant records and a clear explanation of why the case needs attention. An approve button is a weak control when the evidence is missing. Allow time for review and make rejecting or returning a case as practical as approving it.

Decide what happens during absence. A queued case should not become automatically approved merely because a timer expires. Keep it visibly pending or follow an agreed escalation route.

Test the mistakes that matter most

Build a test set with ordinary messages, ambiguous wording and requests outside the intended task. Define the expected handling in advance. Look beyond average scores to incorrect suggestions accepted without being noticed.

Choose any thresholds according to the task and consequences of failure. Reassess outcomes when the model, instructions or document stream changes. A threshold used in a demonstration does not establish that your process is adequately controlled.

Distinguish a score returned by a recognition service from a confidence percentage written by a language model. Do not treat a self-reported percentage as measured reliability. Use your own test cases to establish what a score means for the task; not every service provides a useful confidence score.

Make these decisions explicit

  • Which evidence can a reviewer use?
  • Which data may be sent to the selected service?
  • Who may authorise the subsequent action?
  • How is an error traced and corrected?
  • When is automated assistance disabled?

Messages and documents remain input, not authority to change workflow rules.

If confirmation is missing after an approved change, first establish whether the change has already taken effect. Trying again may repeat the action. Agree who investigates the status, can pause further actions and authorises a correction. A correction is not always a simple reversal. Microsoft: Retry pattern

Example: a trial plan for product questions

Use the fictional source TEST-H1, version A, for an initial exercise. The aim is not to send replies automatically, but to assess whether a draft helps the person doing the work. This is a proposed trial, not a completed customer case.

Scope

Draft a reply to a product question about the PX-10 label holder only. Do not promise prices, discounts, stock or delivery dates. No customer data or connection to a sending system.

Preparation

A process owner selects the task and the permitted test environment. A team member records in advance which information the source does and does not support for each question. Use the fictional source, not real customer data.

Test cases

Ask about colours: black and white appear in the source. Ask about red: it is not listed. Ask for delivery tomorrow with a discount: that information is missing. Include the other cases in the exercise logbook too.

Review

The reviewer sees the question, source version and draft together. Reject a draft that makes an unsupported commitment. Record missing information, corrections and time spent searching and reviewing.

Decision

Discuss the errors found, the correction effort and whether a template or existing feature is enough. Agree in advance which errors block further use. A small exercise does not authorise autonomous sending.

Download the fictional TEST-H1 source and instructions

Use the reading route and editable trial logbook

A useful trial does not have to change records

During an initial trial, run suggestions alongside the existing process without letting them modify records independently. Review differences with the process owner. This reveals where assistance helps and where clearer instructions or a simpler approach would be preferable. Expand authority only after an explicit decision.

Fictional example: a wholesaler receives a message with a new delivery address. The assistant finds the relevant customer record and prepares a change. An employee checks the original message and the delivery. Only the approved change may proceed. During the trial, also explore a message requesting an additional action and a missing confirmation. With consistent, reliable input, a conventional import or workflow may be simpler. This is a design example, not a customer result.

Questions

Is an AI agent the same as RPA?
No. RPA performs actions through software, often following defined steps. An AI model can help interpret unstructured information or choose a next step. The two can be combined; the label does not determine which solution you need.
Can a good instruction replace access permissions?
No. Define permitted actions and enforce those boundaries in the systems. Instructions describe intended behaviour; they do not replace authorisation.
Does every action need human approval?
Not necessarily. Decide according to the action, consequences of error and ability to check the outcome. When uncertain, start with suggestions that cannot change records and evaluate them before expanding permissions.

Which actions would you like help with?

Discuss the step you want to improve and who retains the decision. All exploratory conversations are free of charge and without obligation. The first takes about 15 minutes. Further investigation and implementation follow only after the approach and costs have been agreed.

Want to turn this assessment into a plan for your work? We examine the task, compare suitable options and work with technical colleagues to define a focused trial. Further advice and implementation are agreed separately through Korper.

See how AI consulting can help with a decision or trial plan

Discuss AI in your workflow

Explore one process with a process map

Boermans Digital / privacy

Your cookie preferences

You decide whether Google Analytics may load. The website, scans and calculator also work without Analytics.

Completed fields, scan answers and calculator results are not sent to Analytics.

Which data and retention periods?

We remember your choice on this device for up to 180 days, including a refusal. The _ga and _ga_* Analytics cookies are configured for up to 180 days. Retention at Google is configured separately in the Analytics account.

Storage for website functionsFunctional

The website remembers your privacy choice. Only when you choose to switch language can scans, the guide and calculator temporarily carry answers within the same tab. Contact details are not stored in that transfer.

Contact form protectionAt request review

Google reCAPTCHA loads when you open the review of your completed request. Google then processes technical data and may use security cookies. This is separate from your Analytics choice. Contact by email remains available.

Analytics is off.