AI workflow automation: rules, AI steps or agents?

AI workflow automation uses a model for part of a repeatable process. The surrounding workflow can still follow fixed rules. Traditional automation handles the steps you can write down as rules; the model handles the part that needs interpretation, such as natural language in a delivery note. An agent goes further: it chooses some of the next steps or tools while working toward an outcome. I would make that extra freedom a separate decision.

For a UAE operations team, a sensible first task might be preparing a weekly delivery-exception report from approved exports. Adding totals, interpreting a driver's note and deciding which system to investigate next are different jobs. They do not all need the same kind of automation.

Choose the smallest system that can finish the job

A fixed workflow can call several systems and handle exceptions explicitly. A language model inside that workflow does not automatically make it an agent. Anthropic's Building effective agents makes a similar distinction between predetermined paths and model-directed tool use. Its architectural distinction is the relevant part here; the article warns that its tooling discussion has changed since December 2024.

AI workflow automation example: a delivery-exception report

Assume the operations lead supplies an approved weekly CSV. Each row has a delivery ID, promised date, completion status, charge in AED and a free-text note. The output is a draft exception report with source-row references. A person checks and sends it.

1. Validate the export. Check required columns, date format and duplicate delivery IDs. Keep the source unchanged. Stop the report if the export period or schema is wrong.
2. Select and calculate with rules. Use the agreed cutoff and status rules to identify overdue deliveries. Add the relevant charges in code or a spreadsheet. An empty date stays missing.
3. Try one AI classification step. Ask the model to suggest a reason from a fixed list, attach the note it used and return “needs review” when the note does not support a category.
4. Validate the draft. Check the output format, permitted categories and source references. Route ambiguous or invalid answers to a person. A valid category can still be wrong, so review its meaning too.
5. Prepare the report. Combine the checked rows with the calculated totals. Keep sending and source-record changes outside this proof's permissions.

For example, a note reading “Customer asked us to return tomorrow” could suggest a customer-request category. It does not establish a new promised date. The system should preserve the source date until the person responsible confirms a change.

If the same information is already recorded in a reliable reason-code column, the classification step may add nothing. Use that column and measure whether any remaining manual work warrants AI.

When would an agent add something?

Suppose the CSV says delivered but an approved support export contains an unresolved complaint. A fixed workflow can flag the conflict. An agent might help if resolving it requires choosing among several approved searches based on each result.

Before allowing that investigation, name the sources it may query, the maximum attempts, the deadline and the action it must take when it cannot resolve the conflict. In this example the agent may assemble a case for review. It may not change delivery status or contact the customer.

Compare the investigation against a fixed lookup sequence and a human following the same instructions. If the agent cannot improve completion or reduce total handling time at an acceptable error rate, the simpler route remains a valid outcome.

Write the failure cases before the prompt

Freeze a test set before tuning prompts, then keep separate cases for the final check. Record failed drafts and reviewer corrections. The local-runner check shows one reason to test the whole receiving system: correct values inside Markdown failed a strict JSON parser.

Measure the work after the model replies

Record preparation, review, corrections and repeat runs separately. Add recurring maintenance and tool costs. Count cases completed correctly, cases escalated and cases incorrectly treated as complete. A faster draft can increase the total workload if someone has to reconstruct its sources.

For a UAE team, use the agreed staff cost in AED and include the languages the operators encounter. Keep the financial comparison in one place: the readiness guide's AED example shows how an extra hour of corrections can erase an apparent saving. It uses illustrative assumptions, not a promised return.

Use the workflow-choice worksheet

Download the workflow-choice worksheet. It includes a step map, the eight proposed test cases, blank result fields and a decision record. Start with the person who does the work; mark any assumption they cannot confirm.

Then use the agent permission matrix to name each allowed action and its enforcement point. The UAE workflow-readiness guide covers the wider assessment, while the production-agents worked example shows how the report's handoff can fail.

The decision should state which steps stay manual, which use rules and which AI step deserves a trial. Include the unresolved failures and the person who approves the next stage. The readiness guide has a table for the overall verdict and what must support it. If you want someone outside the team to run this check, my AI workflow readiness assessment ends with a written verdict: proceed, proceed with limits or do not automate.

Two more AI workflow automation examples

These are hypothetical designs, not client deployments or measured results. They show how a different task changes the permissions and the review.

Arabic and English enquiry triage

A shared inbox receives requests in both languages. Fixed rules validate the sender fields and remove attachments outside the agreed scope. A model suggests a category and a short summary with the original wording attached. A person selects the final destination. The proof cannot send a reply or add a contact to a marketing list.

Test a mixed-language enquiry, an ambiguous request and a message asking the agent to ignore its rules. If reliable categories already arrive from the enquiry form, start with that routing rule instead of adding a model. Keep customer identities out of synthetic tests.

Internal document retrieval

An operator asks a question of an approved policy collection. The workflow retrieves candidate passages, checks that they belong to the current approved version, and drafts an answer with page references. A missing source or conflicting policy should produce an unresolved answer for review.

The proof may search the approved collection but cannot change a policy or search a private folder outside it. Test a withdrawn document, an answer absent from the sources and a correct-looking answer with a wrong citation. If the question always asks for one fixed field, a direct lookup may be enough.

Compare the three workflows

Download the two additional case sheets. Pair them with the agent-testing case pack and record the full implementation and review cost before expanding the scope.

Suhail Mohebi · published 31 October 2026 · the cover is a generated illustration