AI Tools Field Guide

Open edition · no CAC needed

Beginner — Use AI tools at work

Where it earns its keep

ChatGPT, Gemini, Grok and Claude are language tools for staff work. The jobs they are best at are the ones that already have a document at the start and a document at the end.

Jobs by role

If you are… Reach for it when A first job to try
Watch team or ops A long instruction needs to become the twelve things your team has to do "List every action the watch team must take, with the paragraph number"
Admin, N1, personnel The same package comes round every quarter This quarter's award citations from the bios, in last year's approved format
Training petty officer A qualification standard needs a plan, slides and a quiz A 12-week plan as a table, with the standard's paragraph beside each week
Department head or XO Eleven SOPs from four departments need one shape An agent that rewrites any SOP as numbered steps, one action per step
METOC / weather Products go out all day and the CO wants the numbers A statistics block from today's product log, totals shown for you to recompute
PAO, protocol, events A ceremony needs a flyer, a program and a seating plan that agree Flyer and program text from one details sheet; every name and rank matched

Where it fails

This is the full list. Every row has a fix in Every limit has a workaround, and none of them means "not for me".

Failure What it looks like What to do
It doesn't know your command Confident answers about your instruction, roster or last week's message when nothing was attached Attach the source; it is a language processor, not a knowledge store
It invents references An instruction number or citation in the right format that does not exist or does not say that Make it quote the sentence and paragraph; a second chat checks each citation against the file
It doesn't do arithmetic Watchbill counts, per-diem, days between dates that look right and were never computed Make it write the formula or use a tool that computes (Grok's skills, Excel, PowerShell); the model plans the sum, the tool does it
Its knowledge has a cut-off date It does not know this year's policy change unless you give it Attach the current version, or keep it in a project or helper so every chat has it
Long files are read unevenly A fact on page 60 of 100 gets missed Attach the section, not the binder, or ask about that section by name
It drifts in long chats It keeps returning to a version you rejected Start a fresh chat with only the source and the request that still matter
A document can steer it Text inside an uploaded file that reads like an instruction gets followed Treat outside documents as untrusted; a person reviews before anything acts on the output
It takes only what is allowed A public tool should never see controlled or classified material, PII or PHI; a work account may allow more, per your organization's rules If a file would be a problem on the share drive, it is a problem here

The three-question test

Run this on any task before you open the tool.

  1. Do I have the source? (the instruction, the old package, the notes)
  2. Can I check the output? (against that source, or with a calculator)
  3. Is the data allowed? (nothing controlled, classified, or personal in a public tool; within your organization's rules for a work account)

Three yeses and it is a job for AI. One no and it is not, yet: find the source, plan the check, or take the prohibited columns out first.