Banxico: almost half of companies in Mexico already use AI. Why do only 4 in 10 see higher productivity?
Banxico's new survey shows AI adoption doubled in a year, but most companies use it for content, not for operations. What the data says and how to turn usage into results in your company.
September 21, 2026 · Lixto Labs Team · 4 min read
Banco de México (Banxico, Mexico's central bank) published the results of its Monthly Regional Economic Activity Survey (EMAER, for its Spanish acronym) on artificial intelligence, reported today by Bloomberg Línea. It's the most rigorous snapshot we have of how Mexican companies use AI, and it brings one piece of good news and one piece of bad news.
The numbers
- 48.5% of companies with more than 100 employees already use AI (June 2026). A year earlier, in September 2025, the figure was 24.3%: adoption doubled in nine months.
- 64.8% expect to use it within the next three years.
- Size matters: 69.9% of companies with more than 1,000 employees use it, versus 43.7% of those with 101 to 250.
- By region, it ranges from 45.8% in the north to 51.2% in the center.
- In manufacturing, the main use is data processing (50.1%); outside manufacturing, visual content creation (64.2%) dominates.
And the part that matters:
- Only 39.1% report positive productivity impacts so far, although 60.2% expect to see them in the future.
- 67.5% have not seen significant effects on employment.
The takeaway: adopting AI is not the same as operating with AI
The data tells a coherent story. Most companies "use AI" to generate images, draft text or summarize documents. That's useful, but it doesn't move operational productivity: it doesn't collect faster, it doesn't fill orders on time, it doesn't prevent stockouts.
This isn't just a Banxico finding. EY Mexico reports that barely one in ten companies takes AI to an advanced stage, and other studies this month put the share achieving significant economic value at 6%. The gap between the 48.5% that use AI and the 39.1% that see results is, almost exactly, the gap between desktop AI and AI connected to processes.
And there's a data point the survey doesn't measure but that can be inferred: companies with fewer than 100 employees don't even appear. If adoption is 43.7% among companies with 101 to 250 employees, it's lower in the typical SMB. That's not necessarily bad: it means SMBs can still skip the "AI for making presentations" stage and go straight to the one that delivers results.
Where the uncaptured productivity is
In the small and mid-sized companies we work with, the return shows up when AI touches a process with these three characteristics: it's repetitive, it has structured data and its outcome is measured in pesos or hours.
| Process | "Desktop" AI | AI connected to operations |
|---|---|---|
| Sales | Draft a follow-up email | Detect unanswered quotes and follow up with current price and stock |
| Collections | Summarize an account statement | Send reminders, reconcile applied payments and send a payment link |
| Inventory | Make a sales chart | Flag products below minimum and propose the purchase order |
| Invoicing | Explain what a payment complement is | Issue the CFDI and the REP from the recorded order and payment |
| Management | Ask a chatbot for ideas | Ask "how much did we sell per branch this week?" and get the real figure |
The difference isn't the model; it's that in the right-hand column the AI has access to the data and permission to act, with clear rules.
How to be in the 39% (and not the 61%)
- Choose a process, not a tool. First ask "which process costs us the most hours or the most money?" and then "which AI solves it?".
- Get the data in order before the prompt. If customers, quotes, inventory and invoices live in five spreadsheets, the first AI project is consolidating them.
- Measure before and after. Average days to collect, lead response time, stockouts per month. Without a baseline there's no productivity to report.
- Start narrow and with permissions. An agent that only sends reminders about overdue payments, with a log and human approval for anything sensitive, reaches production in weeks.
- Train the team on the workflow, not the tool. The challenge the surveys mention most is training; it gets solved faster when the AI lives inside the system the team already uses.
Sales director: How many of last week's orders were invoiced late, and why?
Assistant: Seven of 64. Five were waiting for stock of the 20-liter presentation and two had incomplete customer tax data. Would you like me to show you the products with stock below minimum?
That kind of question — with a real figure and a cause — is what the survey calls "productivity impact." In Centella, the assistant answers and acts on the same sales, inventory, invoicing and collections data, with each user's permissions.
Frequently asked questions
Does the survey say AI is eliminating jobs? Not for now: 67.5% report no significant effects on employment and only a minority expect reductions. What does change is the type of work: less manual data entry and follow-up, more customer attention and closing.
My company has 30 employees. Does this apply? Yes, and with an advantage: you have fewer systems to integrate and faster decision-making. A single well-automated process can show up in next month's cash flow.
Where do I start if we already use ChatGPT or Copilot? Keep them for what they're good at — drafting, summarizing — and pick an operational process for the next step. The key question is: where does the data the AI would need to act actually live?
Half of large companies already use AI; fewer than half of them see results. The difference lies in connecting it to operations. Request a Centella demo, explore our AI integration services or tell us which process costs you the most.