AI and Automation

What you can and cannot send to a model

In this article

Nori.mx11 min read

Crosses freelyManuals, standards, totalsCrosses with a contractYour costs, your pricesDoes not crossCustomer drawings, IDsGet it in writingfrom the model provider1No training on your data2How long it keeps it3Where, and with whom4Breach notice · art. 19

The question is almost never asked by IT. It comes from quality, or from the trade compliance group, the moment somebody runs the demo: “and where does that drawing go?”

It is the right question and it deserves a better answer than “it’s in the cloud.” Because there are two separate things at stake, and the one that usually worries people is not the one that can cost you the customer.

The first is personal data law, which in Mexico was replaced outright in 2025. The second is the agreement you signed with your customer, which is almost always stricter than the law and which nobody reads before connecting anything.

First: the word “processor” changes everything

Mexico’s Federal Law on Protection of Personal Data Held by Private Parties in force today dates from March 2025. It replaced the 2010 law and changed things at the foundation.

The definition that matters here is in Article 2. A persona encargada — a processor — is the “persona física o jurídica que sola o conjuntamente con otras trate datos personales por cuenta del responsable”: a natural or legal person who processes personal data on the controller’s behalf, following its instructions, without deciding what is done with the data.

Now the part almost nobody has read. The same law defines a transfer as:

Toda comunicación de datos personales dentro o fuera del territorio mexicano, realizada a persona distinta de la titular, del responsable o de la persona encargada del tratamiento.

Any communication of personal data, inside or outside Mexican territory, to a person other than the data subject, the controller, or the processor.

Read the ending. The definition expressly excludes the processor. Sending data to someone who processes it on your behalf is not a transfer, and therefore does not fall under the consent regime the law requires for transferring data to third parties.

That is the good news, and it matters because the most common fear — “so I need a signed consent from every employee before an email can hit an API” — generally does not apply when the provider acts as a processor.

Now the bad news, and it is the part that governs: the responsibility does not move. Article 13 requires the controller to ensure compliance with the law’s principles and to see that the privacy notice is respected at all times, including by the third parties with which it has a legal relationship. The provider being a processor takes nothing off your plate. What it does is leave you as the one who answers for whatever happens over there.

Second: the authority you answer to is no longer INAI

This is the practical change that most compliance documents still have wrong.

INAI no longer exists. The authority that supervises and verifies compliance with this law is now the Secretaría Anticorrupción y Buen Gobierno, named as such in Article 2 and given those powers in Article 39.

A concrete, free conclusion: if your privacy notice mentions INAI, it is out of date. And if you published it before March 2025, there is more to review: the new law requires identifying sensitive data and separating the purposes that need consent from those that do not.

The fines, so the conversation has a size

Not to frighten anyone, but because without numbers this discussion stays at the level of opinion.

Article 59 sets fines of 100 to 160,000 UMA for one group of conducts and 200 to 320,000 UMA for another, plus an additional fine for repeat offenses. Where sensitive data is involved, the amounts may be increased up to double.

With the 2026 UMA at 117.31 pesos per day, that is:

  • 160,000 UMA = 18,769,600 pesos
  • 320,000 UMA = 37,539,200 pesos
  • Doubled, where sensitive data is involved = up to 75,078,400 pesos

And there are criminal penalties, not only fines: Articles 62 through 64 provide for imprisonment, with the terms doubled where sensitive data is involved.

The point is not that you will be fined for piloting an agent. The point is that the size of the exposure justifies half an hour of review before connecting anything, and that half hour is almost never taken.

Third, and this is what can actually cost you the customer

Everything above is law. Here is what is not law and weighs more.

If you supply a maquiladora, you signed a non-disclosure agreement. Read it. Almost all of them say some version of the same thing: you will not disclose the information to third parties without written authorization.

Sending your customer’s drawing to an AI provider’s API is, literally, disclosing information to a third party. It does not matter that it is for one second, that it is automatic, or that the provider promises not to train on it. If your contract requires written authorization and you do not have it, you are in breach.

This does not end in a fine. It ends with you off the approved supplier list, and there is no coming back from that with an apology.

The good news is that it is usually solved by asking. Many corporate customers already have an AI usage policy and a list of approved providers. Asking costs you one email. Finding out afterward costs you the account.

The same applies to your trade compliance data. Inventory control records under an IMMEX program have to be defensible before an authority, and every movement needs an accountable owner. An agent can read that data and raise a flag; whether it writes to it is a different conversation, and not a technical one.

The list: three piles

In practice, everything you might send to a model falls into one of three piles. This is the classification we use.

Crosses without a second thought

Text that identifies nobody and reveals nothing about a customer.

  • Your own procedures, manuals and internal documentation.
  • Generic descriptions of a technical problem.
  • Aggregate data with no names: “how many pieces were rejected in September.”
  • Public catalogs, standards, statutory text.

Crosses with a contract and with judgment

Operational data of your own that is not public but is not somebody else’s either.

  • Your internal part numbers, your costs, your prices.
  • Operational email with names replaced by identifiers.
  • Your quoting history, if the customer contract does not cover it.

The rule here is simple: it crosses if there is a written agreement with the provider and if what is not needed can be stripped out.

Does not cross

  • Sensitive personal data. Health, ethnic origin, beliefs, preferences. And a near case: time-clock biometrics, where the law leaves more doubt than people assume.
  • Full names with national ID or tax ID, in bulk. An employee roster has no business passing through a model so that somebody can classify an email.
  • Your customer’s drawings, specifications and prices, absent written authorization.
  • Passwords, API keys, connection strings. This looks obvious and it is the most common leak, because they ride along inside a config file somebody pasted whole.
  • A worker’s name together with their schedule, location or performance. That combination identifies a person and exposes them, and the law also forbids you from making decisions about people on that basis.

The four things to get from the provider in writing

A marketing page saying so is not enough. It goes in the contract or in the terms you accept, and you keep a copy.

1. That they do not train on your data. Serious providers state this in their commercial terms for API use, and it is usually different from what applies to their consumer product. Confirm which of the two you are using.

2. How long they retain it and how it is deleted. There is retention for abuse and safety reasons, with defined periods. Get the period in writing.

3. Where it is processed, and with whom. Region, and which sub-processors sit behind it. If your customer has data residency requirements, this is the first thing they will ask for.

4. How and how fast they notify you of a breach. This is the one that gets forgotten and the one that can hurt most. Article 19 of the law requires you to inform the data subject immediately when a breach significantly affects their rights, and you cannot report what nobody told you. If the provider’s contract sets no notification deadline, your legal obligation is hanging on somebody else’s goodwill.

The security rule almost nobody has read

Article 18 requires establishing and maintaining administrative, technical and physical security measures. That sounds generic. What is not generic is the sentence that follows:

Los responsables no adoptarán medidas de seguridad menores a aquellas que mantengan para el manejo de su información.

Controllers shall not adopt security measures lower than those they maintain for handling their own information.

It is a bar you set for yourself. If your payroll system sits behind multi-factor authentication with logged access, and the model integration runs on a key pasted into a file three people can read, you are below your own standard and the law measures it that way.

Article 20 adds something that gets overlooked: the confidentiality obligation of anyone involved in the processing survives the end of their relationship with you. That includes the contractor who built the integration and left.

What you can do without sending anything out

Here is the practical part, and it is what turns a “no” into a project that can actually be done.

Strip what is not needed before the boundary. The model does not need a person’s name to classify an email. Replace names with identifiers before the call and put them back when the answer returns. That change, a few hours of development, moves most cases out of the forbidden pile.

Send the dimension, not the file. You almost never need the drawing: you need the dimension, the material and the quantity. Extracting those on your side and sending only that is a design decision with legal consequences.

Keep the data at home and send the question. The agent can query your system and bring back a result without the full contents of the database leaving your network.

And for what truly cannot leave, there are models that run inside. They are less capable and cost more to operate, but for classification and extraction they are enough. That comparison is worth making with numbers rather than on principle: how to cost an agent works the same way for both options.

Five questions before turning on the first agent

  1. Exactly which data will leave my network? Write it out, field by field. If nobody can answer this, the project stops here.
  2. Does any of those fields identify a person, or belong to a customer? If so, can it be stripped without the task breaking?
  3. What does my NDA with that customer say? And if it is not clear, who do I ask in writing?
  4. What do the provider’s terms say about training, retention, region and breach notice? All four, filed.
  5. Is my privacy notice current? If it mentions INAI, the answer is no.

None of the five needs a vendor or a budget. All five can be answered this week, and all five will be asked again the day your customer audits you.

Where we come in

We do this scoping before writing code, because reworking an integration already in production costs more than designing it properly. And if your case turns out to be one where no data needs to leave at all, we will tell you that too.

If you want the list of what leaves your operation today written down, Chira asks where the data lives and who touches it, and gives you a document in five minutes. It is free and you do not end up on a sales call.

And if what worries you is not the model but who maintains what is already connected, that is a data risk too: a system nobody can review cannot be audited either.

References

The quotations, articles and amounts come from the sources below. We explain what they mean for a system; if you are making a compliance decision, read them yourself and consult your counsel.

Related service

AI and Automation

Automate the repetitive work and see what is coming.

See the service

Articles on this service

WhatsApp · 9 p.m.Do you machine6061 aluminum?Could you sendthe drawing?Complete requestDrawing and rev.Material, finishVolume and lotTimingQualityShip-to, freightWho is askingNever saysPricesDatesCapacityA person confirmsLaura, by 5

AI and Automation

Your sales chatbot should not quote

What a sales assistant does well at an industrial company: it answers what is public, completes the quote request and hands it to a person. Never prices or dates.

9 min read

Out of every 100 companiesEvaluated60Reached a pilot20In production5MIT NANDA, July 2025Why they stall1Picked for show2Lives on another screen3Does not learn4Nothing measured first5No owner in operations

AI and Automation

Why your AI pilot never leaves the pilot

An MIT study says 95% of companies get nothing out of generative AI. What it really measures, why pilots stall and how to pick your first project so it ships.

7 min read

DRAFTQuoteagent preparedPriceA person signsWhen to let goWhat can be undonefirstWhat gets seenlaterPrice, date, authorityneverAcceptable correctionsInternal1 in 10Customer1 in 50

AI and Automation

The agent prepares, the person signs

How to test an AI agent before trusting it with work: what to measure, when to loosen the leash, what it never signs alone and what happens when it slips.

10 min read