IMPIBot and What Mexico's First IP Chatbot Reveals About AI in the Mexican State
Mexico's first conversational assistant for IMPI reveals how artificial intelligence is being incorporated into the Mexican State, and where its limits remain, as Mexico discusses a National AI Law. Analysis updated to May 2026.
IMPIBot is the first conversational assistant of the Mexican Institute of Industrial Property (IMPI), Mexico's patent and trademark authority. It operates through WhatsApp, was delivered to the agency by the National Citizen Observatory (ONC) in November 2025, and allows users to consult three public databases in one place: IMPI tobacco trademark records, health alerts from the Federal Commission for the Protection against Sanitary Risks (COFEPRIS), Mexico's drug and health regulatory authority, and SAT Annex 11 with the brands authorized to import tobacco into Mexico. Its value does not lie in producing new information, but in bringing into one conversation what citizens previously had to search for across three different authorities.
On November 10, 2025, the Mexican Institute of Industrial Property (IMPI) and the National Citizen Observatory (ONC) formalized a collaboration agreement aimed, according to the official communication, at strengthening legality and contributing to the prevention of crimes and infringements in industrial property matters. The most tangible result of the agreement was the delivery to the Institute of a WhatsApp conversational agent called IMPIBot, accompanied by the documentation and technical inputs required for its operation.
The inaugural use case is the fight against tobacco piracy. According to IMPI's statement, one in five cigarettes sold in Mexico is illegal. During the event, Gaston Zambrano, a representative of British American Tobacco (BAT), estimated tax evasion from illegal tobacco at fifteen billion pesos per year.
At first glance, IMPIBot could look like one more note in the discourse of public-sector digitization: an authority adopts a technological tool, presents it as an operational advance, and connects it to the fight against illegality. But the case allows a less immediate reading. The relevant point is not simply that IMPI has a chatbot, but that a Mexican authority is beginning to use a conversational interface to organize public information, guide queries, and bring citizens closer to databases that were previously dispersed across different institutions.
This analysis does not start from the premise that artificial intelligence will solve the structural problems of the Mexican State. A friendlier interface does not, by itself, correct incomplete databases, systems that do not communicate with one another, or authorities that operate with limited resources. But it also makes little sense to ignore what these tools can do. When they are well designed, they can reduce friction, bring information closer, and make part of the public infrastructure that already exists more usable.
Behind that WhatsApp assistant lies a much broader question: if public information already exists, if the databases are already there, and if a significant part of the problem is that no one knows where to look, what other regulatory areas could change simply by redesigning the way citizens converse with the State?
I. How IMPIBot arrived and what it does
According to the information released by IMPI, IMPIBot operates through WhatsApp and allows users to consult three different sources: trademark records linked to tobacco products, health alerts issued by COFEPRIS, and SAT Annex 11, where the brands authorized to import tobacco into Mexico are identified.
The tool does not appear to be designed to produce new information. Its main function is simpler, though not irrelevant: to bring into a single conversation data that was previously dispersed across different public platforms. Before IMPIBot, anyone who wanted to carry out this type of verification had to know that part of the information was held by IMPI, another part by COFEPRIS, and another by SAT. The bot attempts to reduce that friction.
That is its first lesson. Very often the problem is not that the State lacks information, but that the information is fragmented, hidden behind different portals, or written for users who already know the administrative architecture. In that context, a conversational interface can produce a real improvement without yet transforming the underlying system.
But how did IMPIBot come to operate? IMPIBot was not presented as the result of an internal development financed by the agency itself, but as a tool delivered to IMPI together with the documentation and inputs needed to operate it. That fact allows the project to be read from another angle: not only as a technological innovation, but as a practical response to the budget constraints under which Mexican authorities often operate.
Recently, the reform to Mexico's Ley Federal de Protección a la Propiedad Industrial (Federal Law for the Protection of Industrial Property, LFPPI) expanded several IMPI powers, but implementation of those changes remained subject, to a large extent, to the resources already available. In other words, more was asked of the Institute without accompanying that demand with an equivalent budget expansion.
Nor should this be read only critically. Not every public improvement has to come from a new budget line. Sometimes it is enough to identify a concrete problem, use information that already exists, and build a simpler way to consult it. The fact that it is now possible to review, from WhatsApp, data that previously required moving across several official sources is not minor.
II. What problem does IMPIBot solve?
IMPIBot's value does not lie in discovering information that did not previously exist. It lies in something more modest and, perhaps for that reason, more interesting: it prevents the user from having to know in advance the internal map of public administration.
To verify a tobacco product, it is not enough to ask whether the mark exists. It may also be necessary to review whether there is a health alert, whether importation is authorized, or whether the information available from one authority matches the information held by another. In practical terms, that means moving between registries, notices, annexes, search engines, and official pages that were not designed to converse with one another.
That design shifts to citizens a burden that should not fall entirely on them. The user must know not only what they want to consult; they must also know where the answer is, which authority keeps each piece of data, and how to interpret the difference between trademark, health, and tax information. The State keeps the information, but the citizen must reconstruct the path to find it.
A conversational interface can be useful precisely there, because it makes access to correct information less hostile. If the system allows natural-language questions and returns organized information from public sources, the benefit does not lie in the sophistication of the bot, but in the reduction of intermediate steps for the user.
That reduction of friction has important value in markets where counterfeit products, illegal goods, or goods subject to health control circulate. Public information is useful only if it can be consulted at the right moment. A technically available database that is difficult to find or use performs only a limited function.
IMPIBot points exactly to that intermediate space: it does not replace inspection, it does not replace a complaint, it does not decide whether a product should be seized, and it does not turn the citizen into an authority. Its function comes earlier. It organizes a query that, without that interface, would require more time, more knowledge, and more familiarity with bureaucratic architecture.
For that reason, the tool should not be measured only by its technological novelty. The relevant question is whether it reduces a real barrier. In this case, the barrier seems clear: dispersed public information, authorities that concentrate different data, and users who do not necessarily know how to navigate among them.
That is what makes the case interesting. IMPIBot does not prove that artificial intelligence can modernize an authority by itself. It proves something more limited, but more verifiable: that part of the relationship between citizens and regulatory bodies can improve when the State stops requiring users to think like bureaucrats.
III. International context
To understand what kind of tool IMPIBot is, it is useful to look outward. In the main intellectual property offices, artificial intelligence is not entering only as a service desk. Its most relevant use appears inside institutional processes: classification of applications, prior-art search, image analysis, document review, consultation of criteria, and support for the technical work of examiners.
In those cases, technology is not limited to telling users where to search. It helps the authority process more effectively what it already receives every day. The point of support is not the citizen asking a question, but the official who must analyze, classify, or decide. Artificial intelligence appears not as a substitute for administrative decision-making, but as a tool to reduce workloads, organize information, and make examination more efficient.
IMPIBot operates from another place. It does not enter the heart of the trademark or patent procedure. Its point of contact is the citizen, merchant, or potential complainant who needs to verify dispersed information before making a decision or reporting a product. This does not make it less valuable, but it does place it in a more limited category. It is a tool for access, not yet a tool for institutional adjudication.
That contrast matters because it prevents an exaggerated reading. Not every use of artificial intelligence by a public authority implies a deep modernization process. Improving the conversation with citizens is one thing. Intervening in the processes that determine response times, examination quality, consistency of criteria, or coordination among agencies is another.
Seen in that light, IMPIBot occupies a preliminary but relevant level. It does not transform IMPI's substantive capacity, but it can change the way a person approaches information connected to industrial property, health, and taxation. That is not minor in a country where access to public information often depends less on whether the data exists than on whether the user knows how to locate it.
The international comparison also shows that the next question is not whether Mexican authorities should use artificial intelligence, but where they should place it. As a service layer? As internal support? As an enforcement tool? As a mechanism for detecting patterns? Each choice produces different legal, operational, and accountability consequences.
IV. The opportunity Mexico is not yet seeing
If IMPIBot serves for something beyond verifying tobacco marks, it is because it allows an uncomfortable question to be imagined: why is that same logic not being used in areas where information that fails to arrive on time can have much more serious consequences?
Pharmacovigilance and public health
Pharmacovigilance is one of the areas where a conversational interface could have the greatest impact. The early detection of an adverse drug reaction, a contaminated batch, or a medical-device failure depends first on someone reporting it. But that step, which seems simple on paper, is often the most fragile part of the system.
Today, official figures show a growing system. Between January and November 2025, COFEPRIS received 38,536 pharmacovigilance notifications and 38,582 technovigilance notifications through the 32 state centers. That is more than 77,000 reports in total. The same official communication states that Mexico ranked second in Latin America for adverse-reaction reports submitted to VigiBase, the World Health Organization's global platform.
Behind that apparent success, however, the documented problem of underreporting remains. Health professionals work under clinical overload, many do not clearly know which channel to use, others fear legal or administrative consequences, and almost everyone stops reporting when there is no feedback. On the patient side, the problem is even more basic: many people do not know what an adverse reaction is, which data are useful, or why reporting them matters.
If a conversational tool can reduce friction for checking whether a tobacco mark is registered, it could also reduce friction for reporting an adverse drug reaction. The point would not be to replace COFEPRIS or turn a chat into a health authority. It would be to let a person describe what happened in natural language, attach a photograph of the package, identify the batch or expiration date, and generate a structured report that can later be reviewed by the authority.
The same applies to suspected falsified medicines, stolen batches, expired devices, or products with public health alerts. A chatbot would not decide whether there is a health risk, but it could make reporting less intimidating, more complete, and more traceable.
A system of this type could operate with a logic similar to IMPIBot's: it would not invent information; it would organize it. It would not replace technical review; it would prepare the input so that review begins with better data.
- Natural-language interaction: Ask basic questions to the citizen or patient, such as what medicine was taken or what symptoms appeared, to route the event capture.
- Photograph capture: Allow the user to upload an image of the bottle, packaging, barcode, medicine, or device involved.
- Medicine and batch identification: Process the image or text through OCR to identify trade name, active ingredient, concentration, batch number, expiration date, or sanitary registration when visible.
- Automatic severity classification: Suggest whether the event appears mild, moderate, serious, or unexpected under standard criteria, so the authority can prioritize review.
- Structured report generation: Transform the dialogue into a pharmacovigilance or technovigilance report ready for formal review by COFEPRIS or the relevant state center.
- Pattern and signal analysis: Aggregate reports so the regulator can detect unusual patterns of symptoms, batches, or regions linked to a product.
- Health-alert consultation: Connect the reported medicine with COFEPRIS alert databases or public recall notices to inform the citizen whether the product has public regulatory actions.
- Authenticity verification: Integrate with sanitary-registration databases or falsification alerts, such as theft reports or WHO alerts, to validate whether the product may be illegal or expired.
- Guided public consultation: Answer common pharmacovigilance questions, such as what should be reported and how to notify it, through guided dialogue.
Consumer protection
Consumer protection exists because buyer and seller almost never have the same information. In Mexico, the Federal Consumer Protection Agency (PROFECO) has already begun to digitize service through platforms such as Concilianet, which allow complaints to be filed and conciliation to take place online. But the design remains reactive: the citizen complains after having been misled. Conversational AI makes it possible to imagine something else: intervention before the conflict, not only after it.
A conversational assistant operated by a consumer authority could function as preventive guidance. A user could paste the terms and conditions of a telecommunications service, a credit product, or a marketplace, and the system would flag abusive clauses, improper waivers, or undisclosed charges. In e-commerce, it could review price history to detect simulated discounts during sales seasons, one of the most common forms of misleading advertising in Mexico.
On the other side, a small business could paste its return policies or a marketing campaign and receive, before publishing it, a signal of what might violate the rules on verifiable advertising. For a small business operating without counsel, that preventive guidance may be the difference between a viable campaign and a sanction.
Conceptually, several uses could be explored:
- A PROFECO chatbot could answer real-time questions about adhesion contracts, such as whether a clause is abusive, statutory warranties, e-commerce, and related matters. It could guide consumers on their rights, including the Federal Consumer Protection Law and complaint procedures.
- It could briefly analyze advertising or contract texts, such as a photograph or text sent by the user, to identify suspicious clauses or require human review.
- In sectors such as finance or telecommunications, where frequent complaints include improper charges or cancellations, an assistant could indicate the steps the user should follow under current law.
- For online marketplaces, a regulatory chatbot could warn about return policies, prices, affiliations, or known fraud alerts before purchase.
V. What a chatbot cannot fix
None of the above means that the answer is to fill public authorities with conversational assistants. That would be too convenient a conclusion. An interface can reduce friction, but it can also hide it. It can make a procedure appear simpler without changing the underlying process. It can improve user experience and, at the same time, leave intact the problems that actually determine whether the authority works.
The MyCity case in New York is a warning. The chatbot was presented as a tool to help small business owners navigate city regulation. The problem was that it began giving incorrect answers about legal obligations. It failed because the trust generated by an official channel amplifies the harm when the answer is wrong.
That is a central difference between a private assistant and a public assistant. When a company deploys a defective chatbot, the problem may end in a consumer claim, a contractual dispute, or reputational loss. When an authority does it, the answer appears clothed with a different legitimacy. The user is not merely conversing with a machine; at least in appearance, the user is conversing with the State.
In this context, disclaimers have limited reach. Saying that the answer is not binding or that the official source must be consulted may serve legally up to a point, but it does not eliminate the practical effect of the guidance. If a citizen acts according to what an institutional channel tells them, the authority can hardly pretend that the interaction never existed.
In Moffatt v. Air Canada, the Canadian tribunal rejected the idea that the company could treat its chatbot as an entity separate from its own acts. The tool was part of the service offered to the user, and the company had to answer for the incorrect information it provided. Transposed to the public sector, that reasoning forces a question: if an authority misguides a citizen through a conversational assistant, who assumes the consequences?
In IMPIBot's case, the question still seems distant because its scope is narrow. But it should not be dismissed. What happens if the system wrongly guides a merchant? What happens if a consumer reports an authentic product as false? Who answers if the information consulted was outdated, incomplete, or misinterpreted by the tool? The authority that operates it, the party that developed it, the party that financed it, or no one?
Here the limit of conversation as a solution appears. Problems of data quality, source updating, administrative liability, personal-data protection, and coordination among authorities do not disappear because the user can write through WhatsApp.
The interface can make entry into the system easier, but it does not guarantee that the system is better built.
VI. Regulatory risks, legal certainty, and the false algorithmic promise
The idea that it is enough to add artificial intelligence for public administration to work better is seductive and, almost always, wrong. When the State transfers monitoring, enforcement, or public-service functions to an automated system, risks appear that administrative law has not finished resolving.
The arguments in favor are familiar. AI automates repetitive tasks, reduces delays in patent and trademark registration, provides service at any hour and in multiple languages, and lowers compliance costs for small businesses. In authorities such as PROFECO or COFEPRIS, it could allow a shift from acting after the problem to detecting patterns before they break open.
The counterarguments carry as much weight or more:
- Algorithmic capture and bias: Machine-learning models are not neutral entities; they learn by optimizing historical government data. If the data used to train the model contain biases, the predictive model will encode and amplify those directives.
- Hallucinations and legal certainty: Language models are probabilistic engines susceptible to generating fictitious but plausible information, known as hallucinations. If a PROFECO chatbot wrongly guides a supplier, or if an IMPI bot states that a mark has no impediments to registration and the citizen invests capital on the basis of that assertion, a legal-certainty conflict is created. Even if authorities impose liability exclusions, this leaves defenseless the citizen who acted under guidance from an institutional State channel.
- False anonymization and privacy: privacy filters used as an intermediate layer, for example the OpenAI Privacy Filter, create a sense of compliance that the system does not necessarily have. Those filters pseudonymize data, but AI models can reidentify a citizen from the context of the conversation without needing the person's name. Added to this is the risk of collecting dialogues without explicit consent, which in Mexico could collide with the protection of private communications.
- Digital exclusion: if service becomes primarily automated and human channels are not preserved, modernization ends up expelling those with less connectivity, digital literacy, or time. In Mexico, that often means Indigenous communities, older adults, and rural populations. An authority without a physical counter is, for that population, an absent authority.
VII. The difference between digitizing and modernizing
The underlying discussion is not whether an authority should or should not use conversational assistants. It probably will. In fact, it has already begun to do so. The relevant question is what part of the problem it is solving with them.
IMPIBot solves an access problem. It brings the user closer to public information that was previously dispersed across different sources. In a market where illegal products circulate, making consultation easier can help consumers, merchants, and authorities detect warning signs more quickly. But that improvement should not be confused with a complete transformation of the system.
An authority can have a better interface without having better data. It can answer faster without its databases being updated. It can receive more queries without having a clear route for processing what those queries reveal. It can appear closer to citizens without changing the way it coordinates information with other agencies.
That is the limit of many digital tools in the public sector. They make the entrance door more visible, but they do not necessarily change what happens behind it. In IMPIBot's case, the conversation becomes simpler; the user no longer has to travel separately through IMPI databases, COFEPRIS alerts, or SAT tax information. But the quality of the answer still depends on those sources being complete, updated, and correctly connected.
There the difference between digitizing and modernizing appears. Digitizing may mean enabling a WhatsApp query. Modernizing means ensuring that the information consulted is reliable, that authorities share data under clear rules, that update protocols exist, that someone measures errors, that results are published, and that there is responsibility when the tool fails.
An interface can be built relatively quickly. Institutional infrastructure cannot. It requires clean data, interoperable systems, maintenance criteria, personnel capable of auditing results, and a public decision about which part of the authority-citizen relationship can be automated and which must continue to involve human intervention.
For that reason, it is useful to resist the temptation to turn IMPIBot into an absolute symbol of modernization. Its existence is positive insofar as it reduces a real friction. But it also shows that the most difficult modernization occurs on another plane: less visible, less announceable, and probably less attractive for a press conference.
The conversation can change in a day. The administration cannot.
FAQ on IMPIBot and government chatbots in Mexico
- What is IMPIBot?
- IMPIBot is the first conversational assistant of the Mexican Institute of Industrial Property (IMPI), Mexico's patent and trademark authority. It works through WhatsApp and allows users to consult three official sources: trademark records linked to tobacco, COFEPRIS health alerts, and SAT Annex 11 with the brands authorized to import tobacco into Mexico.
- Who developed IMPIBot?
- IMPIBot was not developed by IMPI with its own budget. It was delivered to the Institute by the National Citizen Observatory (ONC) on November 10, 2025, together with the documentation and technical inputs required for its operation. Its inaugural use case is the fight against tobacco piracy.
- What does IMPIBot do?
- IMPIBot does not produce new information. Its function is to bring into one conversation data that was previously dispersed across IMPI, COFEPRIS, and SAT. It allows users to verify whether a tobacco mark is registered, whether it has health alerts, or whether it is authorized for importation.
- What other government chatbots exist in Mexico?
- The Ministry of Foreign Affairs operates the PTAT chatbot, developed with UNAM, for consular and passport procedures through WhatsApp. SAT uses statistical learning models to detect invoicing shell companies. Banco de Mexico has seven internal AI systems and a safe-use guide. Recent reviews indicate that at least 14 federal institutions use AI, mostly without specific guidelines.
- Does Mexico have an artificial intelligence law?
- As of May 2026, Mexico does not have a specific artificial intelligence statute. The Senate is discussing the National Law to Regulate the Use of Artificial Intelligence, introduced in February 2026, and the relevant Senate committee announced in May that an additional initiative would soon be filed. In January 2026, the non-binding declaration known as the Chapultepec Principles was published.
- Who is liable if a government chatbot misleads a citizen?
- Mexico still has no clear rule. The most frequently used comparative reference is Moffatt v. Air Canada, where the Canadian tribunal rejected the idea that the company could treat its chatbot as an entity separate from its own acts. Transposed to the public sector, the reasoning suggests that the authority operating the system is responsible for the information it provides, even if official disclaimers attempt to limit that responsibility.
Recommended reading