AI & Data9 min
More and more governments are putting AI assistants in charge of answering questions about procedures, requirements and deadlines. The risk is not that the assistant talks, but that it confidently answers something no regulation says. This article explains why a public answer needs a verifiable source.
AI & Data9 min
More and more institutions use AI to suggest classifications, draft facts or compare records. The risk is not only in the model, but in not knowing what the person did with the suggestion. This article explains why human oversight only protects when it leaves a verifiable trail.
José Campillo
AI & Data8 min
Pressure to adopt AI in public institutions is rising, but most failures start before the model. When detentions, case files, citizen reports, evidence, and closures remain fragmented, AI simply accelerates low-quality decisions. This article explains why operational traceability, interoperability, and data governance are prerequisites for AI that creates real institutional value.
José Campillo
AI & Data9 min
Every three years, many municipalities hand over their systems as an external hard drive and a list of passwords in a sealed envelope. The law asks for more: case files that can be read, explained and audited later. This article explains what separates a backup from an archive and why it matters on handover day.
José Campillo
AI & Data8 min
Digital government platforms fail when access is treated as an IT checkbox. In justice, public safety, and citizen service, role-based permissions, audit trails, and visible responsibility are what let institutions share data without losing control.
José Campillo
AI & Data9 min
Two departments report the same indicator and get different numbers, and neither is lying. The problem is the definition, not the data. This article explains why every public indicator needs a sheet that states what it counts, from when and by which rule, and what happens when that rule changes.
José Campillo
AI & Data8 min
When detentions, citizen reports, evidence, locations, and service requests enter with weak fields, free text, and inconsistent categories, institutions lose decision quality before analytics begins. This article explains why structured intake is the foundation for traceability, interoperability, and reliable operational intelligence.
José CampilloOn a Tuesday night, a resident messages the city's WhatsApp line. She wants to know which documents she needs for a business operating license and how long the department has to respond. The assistant answers in seconds, with clean spelling and a friendly tone: proof of address, ID, a sketch of the premises and a deadline of ten business days.
The next day she arrives at the service counter with everything in a folder. The clerk explains that she is also missing the civil protection sign-off and that the actual deadline is fifteen days. She shows the screenshot. "The city told me."
Nobody acted in bad faith. The assistant produced a plausible answer, assembled from what a license procedure looks like almost anywhere. What it did not have was a source.
An answer is not a source. AI that serves citizens can draft, summarize and explain, but what it states about requirements, deadlines, fees and rights has to come from a document the institution can point to.
The US National Institute of Standards and Technology (NIST) published its risk profile for generative AI, NIST AI 600-1, in July 2024. Of the twelve risks it identifies, the second is called confabulation: the production of confidently stated but erroneous or false content, colloquially known as "hallucinations."
Confabulations are not an occasional glitch that a patch can fix. They are, the document says, a natural result of how these models are designed: they generate text that approximates the statistical distribution of what they learned. Sometimes that yields accurate answers; other times, answers that sound just as good and are false.
The profile adds two observations that weigh heavily in the public sector:
That last warning matters. It is not enough for the assistant to "cite something." The citation has to exist, it has to say what the assistant claims it says, and someone has to be able to check it.
In February 2024, British Columbia's Civil Resolution Tribunal in Canada decided Moffatt v. Air Canada, a small claims matter that became a reference point for any institution with an automated assistant facing the public.
After their grandmother died, Jake Moffatt asked the airline's chatbot about bereavement fares. The chatbot said they could buy the ticket and apply for the reduced fare afterward, within 90 days. Moffatt bought, traveled and requested the refund. The airline refused: its policy, published on another page of the same website, did not allow retroactive requests.
The airline's defense was that it could not be held liable for information provided by the chatbot. The tribunal called it "a remarkable submission" and rejected it. Its reasoning comes down to three ideas:
The airline had to pay the fare difference, interest and fees. The case creates no obligations in Mexico, but the logic carries over to government with no effort: if the city's assistant says something, the city says it.
The National Law to Eliminate Bureaucratic Procedures (Ley Nacional para Eliminar Trámites Burocráticos), published in the Official Gazette of the Federation on July 16, 2025, applies to federal, state and municipal agencies. Its Article 6 sets legal certainty as a principle: only procedures and requirements expressly established in current law and registered in the Single Citizen Portal for Procedures and Services may be demanded.
Article 51 describes that portal as the national channel for consulting and registering all procedures and services across the three levels of government, with an explicit goal: that no authority request procedures, requirements or conditions not established in that tool.
Article 54 lists what each institution must register for every procedure. Among other things:
For anyone designing an assistant, that article is a specification: it says what citizens will ask and where the official answer lives. Article 53 adds that institutions must keep that information up to date and that failing to correct it can be sanctioned.
An assistant that answers requirements from memory, or from whatever it found online, competes with the official source. One that answers from the procedure's registry and cites it reinforces it.
The difference can be checked on five points:
None of these points depends on which AI model is chosen. They depend on how the documents the model reads are governed. It is the same principle we laid out in a suggestion is not a decision: the value of AI in government is determined by the record it leaves, not only by the quality of the model.
A quick way to evaluate an assistant is to ask it three questions that have an official answer and one that does not:
If the assistant passes the first three and fails the fourth by making something up, the problem is one of design: it cannot tell what it knows from what sounds right. NIST suggests verifying sources and citations before deployment and during operation, not just once.
The bottleneck is almost always upstream of the assistant. Many municipalities have their procedures and deadlines scattered across scanned bylaws, printed sheets and the memory of the service counter. An assistant connected to disorganized sources repeats the disorder with better wording. That is why the first job is not choosing the model, but building the inventory of procedures, services and requirements that the law, in its Article 15, already requires of each simplification liaison.
It is the same logic we explained in a dropdown is not a service catalog: when the base information has structure and an owner, everything built on top inherits that quality.
Agora, Intello's citizen services platform, includes an AI assistant that serves residents over WhatsApp at any hour. It takes reports, explains procedures and requirements, tells people where to go, builds the case file with the photos or PDFs the citizen sends and notifies them when their ticket moves forward. Several of its rules respond directly to what this article describes:
The same criterion applies to phone calls. Agora transcribes the call and proposes a ticket filled in with the report type, address and name that were said, without inventing data that was not said. Staff review it and create it. The AI proposes; the person confirms.
And once a report becomes a ticket, every change is logged with who, when, before and after. If a citizen disputes what they were told, the institution can reconstruct the conversation and what happened next.
In Tribuna, the public safety and civic justice platform, INTELLO AI follows the same line: it helps draft the facts, suggests how to classify offenses and checks the case file for consistency before it is sent, but the decision belongs to the person.
Putting an AI assistant in front of citizens is a decision about official communication, not only about technology. Every answer it gives speaks on behalf of the institution, at any hour and to thousands of people.
The question for leadership teams is not whether the assistant answers well most of the time. It is whether, when it answers a requirement, a deadline or a right, it can point to where it got it. And whether, when it doesn't know, it says so.
If your municipality wants to serve residents over WhatsApp with answers that can be backed up, see how Agora serves citizens from start to finish, how Tribuna uses AI as support rather than as the decision, or request a demo.