I use AI for a lot of my ops work now. Policy docs, decision decks, macros, and digging through exports I don’t have time to pivot by hand. The difference between a meh answer and a genuinely useful one is almost never the AI. It’s the brief.

Most people use it like a search bar: one vague question, one vague answer, then four follow-ups to fix it. That’s slow, and it’s where a lot of “AI isn’t that useful” comes from.

Here’s what works for me instead. I use Claude, but most of this applies to any of the big AI tools.

Brief it like a new analyst on day one

Picture a really sharp new hire. Smart, fast, great with data, and knows absolutely nothing about your company. You wouldn’t hand them a CSV and say “thoughts?” You’d tell them what’s going on. Do the same here.

What’s worth frontloading:

• The goal and the audience. “A one-page summary for my VP” and “a macro for frontline agents” are completely different jobs.
• The raw data. Exports (CSV or Excel) beat screenshots of dashboards every time.
• Your policies and process docs. SOPs, runbooks, escalation paths, current macros, and the policy you’re trying to change.
• Your definitions. What counts as resolved? What’s the SLA? Is a reopened ticket a new contact? Every team answers these differently.
• Constraints and decisions already made. Region, date range, what’s out of scope, and what leadership already said no to.
• An example of good. Last quarter’s report, a deck that landed well, or a doc format your team already uses.
• The format you want back. A slide, a doc, a spreadsheet, a chart, or a table you can paste straight into Slack.

Ask it to interview you first

This is the one I wish I’d started sooner. Before it does anything, tell it to ask you questions.

It’ll catch the things you forgot to mention, like which timezone your timestamps are in or whether “escalated” includes reopens. Five minutes of answering questions up front saves an hour of “no, that’s not what I meant.”

Batch your asks

Instead of this:

• “Summarize this export.”
• “Okay, now break it down by category.”
• “Can you make that a chart?”
• “Which ones are trending up?”
• “Write that up for my manager.”

Send one message with a numbered list of everything you need. It can plan the whole job instead of redoing work every round, the pieces come out consistent with each other, and you use up your limits a lot slower. Long back-and-forths are expensive, because every new message means it rereads the whole conversation.

Getting numbers you can actually trust

If you want AI to quantify something (volume, trends, impact, cost), give it what an analyst would need:

• Row-level data, not totals. A ticket-level export can be sliced any way you want. A dashboard total is a dead end.
• What each column means, especially dates, statuses, and anything abbreviated.
• A time range and a baseline. “Up 30%” means nothing without “compared to what.”
• How you calculate the metric. Handle time, first contact resolution, backlog age. Give it your formula so it doesn’t invent its own.
• Cost inputs, if you want the impact in dollars: cost per contact, hourly rates, fees.

Then ask it to show its work: state its assumptions, run the actual calculations instead of estimating (tools like Claude can run the math in code, right on your file), and give you a couple of totals you can check against your source system. If the row count matches, you’re good. If it doesn’t, you just caught a problem before it landed in front of leadership.

Make it visual

AI is really good at turning a messy export into a chart people will actually look at. Tell it the story the chart should tell, not just “make a chart.”

A few other asks that work well:

• A one-page dashboard you can share with a link. Claude can build these as interactive pages.
• A spreadsheet with the charts built in, so you can drop in next month’s data and everything updates.
• A running log. Same chart, same format, every week. Ask it to write you a reusable prompt so next week is a two-minute job.

The prompt I actually use

Copy it, fill in the brackets, and attach your files.

Context: I’m a [role] at a [type of company]. My team handles [what].
Goal: I need to [decision / problem / deliverable].
Audience: [who will read it]. They care most about [what].
Attached: [data export], [policy or SOP], [example of a past report].
Definitions: [how we measure the metrics that matter here].
Constraints: [date range, regions, what’s out of scope, decisions already made].

What I want back:
1. [e.g. a summary of what the data shows]
2. [e.g. a chart of X over time]
3. [e.g. a recommendation with the tradeoffs]

Before you start, ask me any questions you need answered, all at once.
State your assumptions, show your math, and flag anything in the data that looks off.

Keep it real

• It can be confidently wrong. Double-check anything going in front of leadership.
• It doesn’t know your org’s history or politics unless you tell it.
• Your judgment is still the job. AI just gets you to the part where you use it, faster.

The ops brain is still yours. AI just gets you from raw export to “here’s what’s actually happening” before your coffee goes cold, or before school pickup, whichever comes first.

What’s the one report you rebuild by hand every single week? That’s where I’d start.