Work · AI workflow product

Haylu — AI Business Operations Agent

My current portfolio work explores how an agent could investigate an order or payment enquiry across business systems, communicate verified information, and hand unresolved cases to a human. The first workflow and evaluation approach are being defined.

Stage
In definition
Evidence
Authored specification
Last updated
14 October 2026
For
Businesses whose staff answer recurring order and payment enquiries across separate systems.
My role
My own portfolio project. I own the product direction and the workflow definition.
Latest decision
Start with one enquiry: order status and payment confirmation.

The problem

A customer asks where their order is, or whether their payment went through. Answering properly can mean checking the order system, the payment provider and sometimes delivery records, then reconciling what they say. The first business question is which recurring enquiry consumes enough effort to justify a reliable workflow.

The stage

This is specification work, dated 14 October 2026. Nothing below has been tested in a working prototype yet.

The proposed workflow

The agent identifies the order, checks the payment provider's record and compares the information. It then either communicates a verified answer or hands the discrepancy to a human. The model interprets the request; the business systems supply the facts.

Designed, not yet tested · 14 October 2026

  1. Identify the order from the customer's enquiry.
  2. Check the payment provider's record.
  3. Compare the order and payment information.
  4. Communicate a verified answer, or hand the discrepancy to a person.
The proposed first workflow in the current Haylu definition.

What happens when it goes wrong

I'm defining the permissions and evaluation cases alongside the workflow: incomplete information, conflicting records, failed tools and uncertain outcomes. Quality, response time and operating cost belong in the same product decision.

What remains open

Whether a prototype can retrieve enough reliable evidence to answer the request, or make a useful handoff when it can't. That is the next thing to test.

What happened

Nothing has been tested or observed yet. This section will hold actual results, and the conditions they were obtained under, once there are some.

Next question

Can a prototype retrieve enough reliable evidence to answer the request or make a useful human handoff?

Artifacts

Look closer

Want to discuss the decisions behind this?