Custom GPTs are the AI tool most field leaders actually use, and the one operators most often build badly. Done right, a Custom GPT replaces the blank page for a repeatable writing task and gives an area director back five to ten hours a week. Done wrong, it becomes a novelty nobody opens after week two.
Here is what I have built, what has held, and what I would build first if I walked into a new operation tomorrow.
What a Custom GPT actually is
Skip this section if you have built one. Read it if you have not.
A Custom GPT is a version of ChatGPT with three things layered on top: a set of instructions that tells it its role, a knowledge base of documents you upload for reference, and optionally a set of actions that let it call other tools. For a field leader you almost never need the actions. You need clear instructions and good reference documents.
You build a Custom GPT inside ChatGPT Team or ChatGPT Enterprise. Team is $30 per seat per month. Enterprise is negotiated. Both keep your data out of training. The free tier does not. Do not build a GPT that touches operational data on the free tier.
The design principle: one task, not one assistant
Every mistake I have made building GPTs has come from trying to build a general operations assistant. What worked is the opposite. Every GPT does one thing.
A corrective action drafter does corrective actions. Nothing else. A coaching note assistant writes coaching notes. Nothing else. A general manager who wants to use one of these hits it, hands it the situation, and gets back a first draft in 30 seconds.
Compare that to a general operations assistant, where the general manager has to explain what they want, and then explain again, and then correct the tone, and by the third turn they wish they had just typed it themselves. General assistants are demos. Task-specific GPTs are tools.
The AI tools that get used in the field are the ones with a single, obvious purpose. If a manager has to think about how to phrase the request, the tool is too general.
Five Custom GPTs that earned their keep
1. Corrective Action Drafter
Input: The manager pastes the situation. Who, what happened, what SOP was broken, previous history.
Output: A first-draft corrective action document in the company's standard format. Legally reviewable tone. References the SOP by name. Notes the required followup timeline.
Training: The knowledge base holds five real, redacted corrective action examples the operator wrote themselves. Instructions specify tone: neutral, specific, no adjectives, no emotional language. Also specifies never to make legal claims and to always flag when the situation may need HR review.
Time saved per use: 15 to 20 minutes. General managers used to defer corrective actions for days because writing them was painful. Now the draft exists inside a shift and the manager edits and delivers.
2. Coaching Note Assistant
Input: Name of the employee, the observation, what went well, what could improve, the specific example from shift.
Output: A three-paragraph coaching note in the manager's voice, structured as observation, impact, and next step. Tone is direct and warm. Length capped at 150 words.
Training: Uploaded 20 anonymized coaching notes the operator wrote themselves as example outputs. Instructions tell the GPT to match the operator's cadence, avoid corporate jargon, and never criticize character.
Time saved per use: 10 minutes. More importantly, coaching notes actually get written. Coaching that gets documented is coaching that sticks.
3. Inspection Prep GPT
Input: The unit, the type of inspection (health, fire, franchisor audit), the date, and any recent findings.
Output: A targeted 15-item prep checklist for that specific unit and inspection type, prioritized by common findings in the region, referencing the specific SOPs and log locations.
Training: Knowledge base has the local health code sections most often cited, the company's own SOP index, and the last 24 months of inspection findings across the region (anonymized).
Time saved per use: 45 minutes of prep work becomes 90 seconds. Also raises floor quality on repeat inspections because the checklist references specific past findings.
Fig. 1 · Examples do most of the work.
4. Catering RFP Responder
Input: A pasted RFP from a corporate catering prospect. Sometimes an email, sometimes a full document.
Output: A structured first-draft response in the group's proposal format: capabilities summary, menu recommendations from the current catering menu, headcount pricing at three tiers, logistics plan, references to relevant past clients (Stanford, Google, Apple, Meta, LinkedIn, Salesforce, Cisco, Adobe, Nvidia where relevant).
Training: Knowledge base has the current catering menu, the standard proposal template, ten example prior proposals, and a bank of language for common corporate concerns (dietary, dietary compliance documentation, hot-hold logistics, invoicing terms).
Time saved per use: 60 to 90 minutes. Catering manager still edits and reviews, but the blank page is gone. This alone paid for the whole GPT stack in a quarter.
5. Weekly Narrative Editor
Input: The area director drops in the raw weekly numbers and their voice-memo-transcribed notes on what happened.
Output: A cleaned-up weekly narrative for the operating review, structured in the standard format, edited for clarity but preserving the operator's voice.
Training: The knowledge base has 15 of the operator's own past weekly narratives as voice examples. Instructions specify to keep the writer's word choices, tighten sentences, add structure, and never introduce facts not in the input.
Time saved per use: 30 minutes per weekly review. Also raises the quality of the review because the narrative is consistently structured.
How to train a Custom GPT that actually works
The instructions field is where most operators over-invest. They spend an hour writing a page of rules and wonder why the output is generic. The truth is that instructions matter less than examples.
Here is the ratio that works, in my experience:
- 10 percent of quality comes from the instructions.
- 70 percent of quality comes from the examples of good inputs and outputs.
- 20 percent of quality comes from the reference documents in the knowledge base.
What that means in practice. Do not write more instructions. Write fewer, sharper instructions and paste more real examples. Ten example input/output pairs in the operator's voice will do more than three pages of prose about tone.
The instructions template
The instructions I use, in order:
- One sentence describing the exact task.
- Voice guide: two to three sentences on tone with a "sounds like" and "does not sound like" pair.
- Format specification: what the output looks like, section by section, with word counts.
- Refusals: what the GPT should decline to do and how to handle it.
- Reference behavior: how to use the knowledge base and when to say the answer is not in the reference.
Total length under 400 words. If it goes longer, the GPT is doing too many things and should be split into two.
Where I got it wrong
Three failures worth naming.
The general operations assistant
First GPT I built. Named it "Ops Assistant." Loaded it with everything: SOPs, menu, org chart, scorecard, past incidents. Instructions said "help with anything the operator needs." Nobody used it past week three. Too general to know what to do with a prompt like "what should I do about the labor variance in unit 4."
The GPT that made things up
Built a GPT that answered questions about the company's policies. Uploaded the policy manual. Instructions told it to answer questions "based on the manual." It confidently answered questions the manual did not cover, and it took two weeks and one HR near-miss to catch. Fix: added an explicit refusal rule ("if the answer is not in the reference document, say 'I do not have that information, ask HR'") and tested with 20 out-of-scope questions before releasing again.
The GPT nobody owned
Built a health inspection prep GPT with the exiting operations manager. Six months later the reference documents were out of date, the health code had updated, and the GPT was giving stale advice. Nobody had been assigned to maintain it. Now every Custom GPT has a named owner and a quarterly review, same as an SOP.
Adoption is a management problem
Building a good Custom GPT is technical. Getting field leaders to actually use it is management. What works:
- Introduce it in a one-on-one, not in an all-hands. The area director watches the general manager use it once, live, together. Adoption jumps 5x versus emailed instructions.
- Ask about it in the weekly meeting. "Did you run the coaching note through the GPT? Show me." Same accountability trick as with the dashboard.
- Bookmark it on their phone. If the GPT takes more than two taps to open, it will not get used. Custom GPTs can be pinned as home-screen shortcuts.
- Show the time savings, honestly. Do not oversell. Say "this used to take you 20 minutes, this makes it three." Field leaders can smell overselling from a mile away.
The point
Custom GPTs are the highest-leverage AI tool for field leaders because they are cheap, fast to build, easy to use, and easy to kill if they do not work. Start with the corrective action drafter and the coaching note assistant. Build them with real examples, not clever instructions. Assign an owner. Review quarterly.
Done that way, a stack of five Custom GPTs gives every area director an extra day a week. That is not a small number in a multi-unit operation.