An AI playbook for an operations team, on one page

Why "everyone just uses their own AI" doesn't work

Microsoft and LinkedIn's 2024 survey found 78% of AI users bring their own tools to work, and only 39% had any AI training from their company. On an operations team that means six ways of writing the same supplier email, customer data in personal accounts, and no one able to say which tasks AI actually made faster.

A worked example: the scan

Illustration: a made-up business. A 12-person operations team at a parts distributor lists its repeated tasks. Each is scored 1 to 5 on frequency, time per run, rule clarity and risk (5 is safest), and the four are multiplied:

The top three are the report, purchase orders and SOPs. Delivery-date replies take more hours, but a wrong date reaches a customer, so they wait until the first three work. The simplest approach for each: a shared report skill fed the stock export; a shared skill that turns a supplier email into order lines a person pastes and checks; and the SOP writer for walkthroughs. No agent yet.

Each runs for two weeks with one owner, a baseline (90 minutes a report, 6 minutes an order), a target of under half the time with the same quality, and a stop rule: if quality drops on more than 1 run in 10, stop and fix the instructions before going on.

The one-page playbook

1. The three tasks, with the approach, owner, baseline and stop rule for each.
2. Company accounts only: ChatGPT Business or Claude Team, which don't train on business data by default.
3. A person checks anything that reaches a customer, supplier or the books before it goes.
4. One owner per task keeps its instructions current and is the one people ask.
5. A review date: after the two weeks, keep, fix or stop each task, then score the next one.

Next step: the skills behind the playbook

Each task on the page becomes a skill the whole team's AI uses. Claude skills for operations covers the five to set up first, and the AI pilot plan covers running an experiment that ends in a decision.

Questions

How do I roll out AI to an operations team?

Scan the team's repeated tasks and score each on how often it happens, how long it takes, how clear the rule is and how safe a mistake is. Take the top three, give each the simplest approach that works, usually shared instructions or a skill, and run each as a two-week experiment with a baseline and a stop rule. Write the rules and owners on one page.

Which operations tasks should AI do first?

Frequent, time-consuming tasks with clear rules where a mistake is caught in review: weekly reports, order entry from emails, SOP drafts. Leave decisions about people and anything legally binding out of the first round, and keep a person checking anything that reaches a customer, supplier or the books.

Does an operations team need AI agents?

Not to start. Anthropic's guidance on agents is to find the simplest solution possible and only add complexity when needed. Most first wins are a shared skill or instructions a person runs and checks; an agent that acts on its own fits only where its output can be checked automatically and errors are cheap.