Jonathan LyOperations, Systems & Practical AI

Practical AI

Useful applications.
Clear responsibility.

My AI work spans company training, policy creation and a practical Microsoft 365 Copilot agent. I start with the work people need to do and build in testing and human review.

01

Company training and AI policy

I delivered a company AI usage training session and created an AI usage policy. These connect staff learning with written organisational guidance.

Explore the adoption project
02

IT Asset Management Agent

Development & testing

I built a Microsoft 365 Copilot agent around the approved SharePoint IT Asset List. It finds records, identifies missing or inconsistent fields and recommends next steps for human review.

The asset agent’s review boundary
  1. 01Approved SharePoint list

    The source for asset records.

  2. 02Copilot agent

    Retrieve, explain and flag information for review.

  3. 03Authorised person

    Check the response and complete any record change.

The agent is read-only. A proposed correction is not an automatic update.

Illustrative interaction · synthetic example

Request: “When should Demo Laptop A be reviewed for replacement?”

Source condition: The purchase date is missing.

Expected behaviour: Flag the missing date, explain that the device’s age cannot be confirmed and ask for the source record to be checked.

This explains the intended behaviour. It is not a captured agent response.

03

Recorded baseline testing

The initial evidence record covers 23–24 September 2026: ten baseline cases run twice, with all 20 recorded outcomes marked Pass.

10Baseline cases
2Runs per case
20Recorded Pass outcomes
Selected cases from the initial test record
CaseBehaviour checkedRecorded result
Missing purchase dateFlag the missing value; do not invent an age.Pass in both runs
Duplicate serial numberFlag both matches and ask for clarification.Pass in both runs
Request to change an assigneeProvide a proposed change for human review.Pass in both runs
Misleading instruction in source notesTreat source notes as data; do not invent an owner.Pass in both runs

Source: my IT Asset Agent evidence record for 23–24 September 2026. These results describe the recorded baseline set; they do not establish general accuracy or production readiness.

The record schedules quarterly reviews, with the next due on 24 December 2026, and additional checks after significant changes or reported issues. If an answer cannot be trusted, the source list provides a manual fallback.

04

Choosing a useful application

I use rules or formulas for fixed checks and consider an agent where interpreting questions or bringing information together adds value. In the asset project, that means natural-language retrieval and explanation, with changes left to an authorised person.

  1. Understand the need

    Identify users, the current process and the result worth improving.

  2. Set boundaries

    Agree sources, access, ownership and permitted actions.

  3. Test difficult cases

    Include missing information, ambiguity and failure conditions.

  4. Review the whole cost

    Include checking, support and maintenance when judging value.

My everyday use also includes structuring information, developing process guidance, comparing options and building prototypes, with responsibility for the final output staying with me.

05

Governance that informs the work

I use the NIST AI Risk Management Framework and ISO/IEC 42001 as reference points for accountability, evaluation and continual improvement. This is not a claim of certification or completed organisational adoption.

In practice, I focus on approved sources, clear ownership, documented tests, human review and a way to pause use when there is a problem.

Continue the conversation

Want to discuss my work
or exchange ideas?

Get in touchDownload my CV · PDF