In May 2025 I published an article built around a simple question: "Still struggling to talk to AI?" It became one of the most discussed posts on my profile. People asked me for templates and for "the perfect prompt" for their area.
A year and a half later, the question I hear in meetings with heads of Operations, Logistics, Finance, Procurement and IT, across very different industries, has changed. It's no longer "how do I talk to AI?" but "if we're already talking to it well, why aren't we seeing results in operations?"
My short answer: because we're almost always asking AI to fix a process that doesn't yet exist as a system.
1. The symptom: teams that master the chat, operations that don't change
Many teams now use AI assistants every day. They draft emails, summarize contracts and prepare presentations.
But look at the end-to-end process (onboarding a supplier, confirming a delivery, authorizing a return, bringing in a new hire) and almost nothing has changed. Requests still arrive by email or WhatsApp. Documents still sit in shared folders. Approvals still depend on someone answering a message. And at month-end nobody can say for sure how many cases ran late, or why.
AI sped up individual tasks. The process stayed the same.
2. AI doesn't fail for lack of words, but for lack of context
A good prompt helps. But in any operation, whether a distribution center, a store network, Finance or a bank, the quality of what AI does depends on four things that better writing won't fix:
- The process. What are the steps, in what order and under which rules? If that only lives in two people's heads, AI will improvise.
- The data. Where does the information come from, and how complete is it? A form that allows empty fields produces incomplete cases, with or without AI.
- The documents. Where is the case file, and which version is the right one? If the correct file is in someone's inbox, no assistant will find it.
- The owner. Who validates what AI did, when, and by what criteria? Without that answer, nobody will trust the result, and rightly so.
When these four pieces are in place, "talking to AI" becomes much easier, because you no longer have to explain everything every time. The context lives in the system.
3. Four questions before your next prompt
Before looking for another prompt template, try this exercise with your team on a single process:
- What specific task do we want AI to move forward? Not "improve supplier onboarding" but "check that the file contains the required documents and flag what's missing."
- What data and documents will it work with? And does that information arrive complete and in one place today?
- Who decides after the AI? Which person reviews, approves or corrects, and how much time do they have?
- How will we know it worked? With indicators the process should already have: cycle time, returned cases, SLA compliance, incidents or audit findings.
If any of the four has no answer, the next step isn't a better prompt. It's putting the process in order.
4. An example: onboarding a supplier
Take a process that exists in almost every company, in logistics, retail, manufacturing or services: onboarding a new supplier or carrier.
Without a process, someone receives an email with attachments (articles of incorporation, tax certificate, bank details, contract), saves them in a folder, checks by hand whether everything is there, requests what's missing in another email and, weeks later, nobody remembers who approved what or why the first payment was delayed.
With a process, the request comes in through a form that won't move forward without the minimum data. Documents are linked to the case. AI can do the first review: detect gaps, compare data across documents and prepare a summary for the approver. That person decides. Every step is logged with date and owner. And the dashboard shows where cases get stuck.
The same logic applies to proof of delivery in logistics, a return in a store, an authorization in a hospital, a support ticket at a tech company, onboarding an employee in HR or, in banking and insurance, a customer file, where compliance rules also apply.
The difference isn't how well we talk to AI. It's that AI now works inside a flow with rules, and the team keeps the decision.
5. Where control comes in
This is what worries leadership most, and rightly so. The challenge isn't more capable AI agents, but agents reliable enough to earn more autonomy, operating within clear limits of accountability. When that isn't defined from the start, the pattern repeats: runaway costs, business value that's hard to explain, and pilots that impress in the demo but never reach production.
In Mexico, the framework is also shifting. The new Federal Law on the Protection of Personal Data Held by Private Parties, enacted in 2025, treats any manual or automated operation on personal data as processing. That applies to any company handling data on customers, patients, employees or suppliers, not just regulated sectors. And Congress is discussing bills to regulate artificial intelligence. You don't need to guess how regulation will land to act sensibly: if a process can explain what AI did, with which data and who validated it, it will be better prepared for whatever comes.
That's why at FieldWy we start from a simple idea: AI executes; your team decides. It's the design principle behind AURA AI OS, our AI agent layer with executive governance, which runs on processes modeled in FieldWy FLOW. I'm not mentioning it to sell you anything here, but because it's how I believe AI should enter any operation: inside the process, with traceability and with a person accountable for every decision.
6. How to start this week
You don't need a big project. You need a well-chosen process:
- Pick a process that depends on email, Excel or WhatsApp today, repeats many times a month and has a clear owner.
- Sketch it on one page: steps, owners, documents and timing. If your team can't agree on the sketch, you've found the first problem.
- Mark where AI could help without making the final call: classifying, checking for gaps, summarizing, comparing.
- Define two or three indicators before changing anything, so you can compare later with your own data.
- Start small, with people validating every result, and expand autonomy only when reliability justifies it.
Learning to talk to AI is still worth it. But the real advantage won't go to whoever writes the best prompt; it will go to whoever has the clearest processes for AI to work within, and people with the judgment to decide.
A question for you: which process in your area would you put in order first, before adding AI?
If you'd like to review a specific process, write to contacto@fieldwy.com or request a demo.
Want to see how this would apply to your operation? Request a demo or write to us at contacto@fieldwy.com.