AI in practice.

These notes describe workplace work without naming the organization or sharing internal materials. The example below is fictional. Outcomes are observations and user feedback; I did not record formal accuracy rates or timed before-and-after measurements.

Finding the right response.

I took the initiative to configure a Copilot agent for a familiar support problem: someone knows what the customer is asking but isn’t sure which response template to use.

I uploaded response documents to SharePoint and pointed the agent at them. The employee pasted a customer email into the agent. It returned the template name and full text, or a suggested response with an explanation when no template fully fit. The employee reviewed and edited the result before sending it.

For example, a fictional email about a delayed delivery could lead to the relevant delivery-response template. An email combining a delivery problem and an account problem needs closer review: one template may not answer both.

Making categories easier to choose.

A second Copilot agent used an Excel file in SharePoint containing the available ServiceNow categories. The employee pasted the email into the agent, received a category suggestion, and manually selected the category in ServiceNow.

This helped when someone was unsure which category fit. The agent suggested a choice; it did not update ServiceNow or categorize tickets automatically.

Test. Adjust. Try the same cases again.

I revised the agents’ instructions when answers missed the issue or included unsupported details. I removed redundant wording, added specific “do not” constraints, and used the same customer emails again to compare results.

The checks were straightforward: did the answer address the complaint more directly, and did it avoid making up facts? Poor grammar and emails with several issues were harder cases. Many responses needed no changes in my use; others needed manual correction.

Teach the workflow and the check.

I showed coworkers how to find the agents, explained what they did, walked through examples, and let people try them. I taught them to consult the relevant knowledge-base articles or standard operating procedures when they were unsure about a suggestion.

That gave the tool a role in learning the procedures, too. A few coworkers tried the agents and reported saving time and effort. They continued using them regularly. Those are qualitative reports, not a measured team-wide productivity result.

Keep the source material current.

When new response templates became available, I added them to SharePoint so the response agent could use them. The source materials were accessible to my coworkers. Keeping that library current was part of maintaining the agent.

Instructions people could follow.

I also developed two AI-assisted instructional PDFs for employees. I used AI to research learning resources and draft the material, directed the language and visual style, and reviewed and edited the results. I wanted approachable explanations without internal company information.

After approval, I distributed the guides to the team. Feedback was positive: colleagues said the material answered questions they had. My contribution was choosing the audience and approach, directing the draft, checking it, and making it clearer.

What I take into the next workflow.

Start with a task people already need to do. Give the agent relevant sources. Test difficult examples as well as easy ones. Show people how to use the result and when to check a procedure. Keep listening after the first demonstration.

My public projects give me another place to explore those questions: TinyDesk for workflow design and evaluation with fictional tickets, TinyTalk for memory and answer provenance, and Cricket Mode for repeatable instructions. TinyDesk’s twelve-ticket results belong to that demo and are separate from the workplace observations here.