AI process automation: 10 processes your company can automate today

Almost everything written about AI automation fails at the same point: it names the category (automate your back office) and never gets down to the process. And the process is where everything is decided: automating invoice capture and automating payment approval do not differ in degree, they differ in that one works in three weeks and the other lands you in an accounting problem.

All four come from studies with a comparison group and they point to the same thing: AI genuinely speeds up certain tasks and makes others worse. Automating well comes down to knowing which side each step of your process falls on. None of these measurements were taken in Mexico, and none of them measure accounting savings: they measure cases resolved, minutes and accuracy rates.

Classic automation executes rules: if field A says X, move the record to B. It works perfectly with structured inputs and breaks the moment someone sends the information in an email worded however they like, or in a PDF scanned sideways. That is where AI comes in: it does not replace the rules, it gives them an input only a human could produce before (reading a messy document, understanding an ambiguous email) and lets the rest of the flow stay deterministic.

That matches what we see in companies already using it. In the working paper from the Center for Economic Studies at the U.S. Census Bureau, using data from the BTOS survey, the dominant generative uses among workers are writing, document analysis and information search, and 65% of the companies where workers use AI keep that use limited to three tasks or fewer ⁵. Careful with the convenient reading: that measures what people use most, not what is most automatable.

The criterion that works best has four filters: high volume and low variety, verifiable output (you know whether the result is right without opening a debate), data that already exists in an accessible system, and errors whose cost is bounded and reversible. If it fails all four, do not automate it. If it fails only the fourth, automate the preparation and leave the decision to a human.

It also helps to look at where the value concentrates. McKinsey estimated that generative AI could add between 2.6 and 4.4 trillion dollars a year to the global economy across the 63 use cases it analyzed, with close to 75% concentrated in four areas: customer operations, marketing and sales, software engineering and R&D ¹⁰. That is modeled potential, not money captured. Actual adoption only partly lines up: among U.S. companies that use AI, the most common functions are sales and marketing (52%), strategy and business development (45%) and IT (41%) ⁵. Two of those three are not among the four areas where McKinsey concentrates the potential value.

For each one: what hurts today, what the AI does, what you need before you start and what stays unsolved. The order runs from lower to higher implementation difficulty.

It pays to go by studies with a comparison group rather than by vendor promises. In the study with 5,172 support agents at a U.S. software company (most of them working from offices in the Philippines), access to a generative AI assistant raised cases resolved per hour by 15%, with the improvement concentrated among the least experienced, lowest-performing agents ¹. One detail that almost never survives the summaries: the gains were larger among agents based in the Philippines, partly because the assistant improved their English fluency; that mechanism does not apply to a Mexican team working in Spanish.

In knowledge work, the field experiment with 758 workers recruited with Boston Consulting Group found that, across eighteen tasks inside what the model does well, those who used GPT-4 completed 12.2% more tasks, 25.1% faster and at significantly better quality ³. And in software development, three experiments with 4,867 developers produced 26.08% more tasks completed, with a standard error of 10.3%: a positive but imprecise effect ¹².

The same pattern shows up in small and mid-sized companies. An OECD survey of SMEs in seven countries (four of them G7; Mexico is not in the sample) found that, among those already using generative AI, only 29% report using it in their core activities: the predominant use is peripheral, supporting operations without deeply redesigning the productive processes. The most reported benefit is improved employee performance, ahead of cost savings ⁷.

And it is worth calibrating the full-automation talk. In the World Economic Forum survey of more than a thousand employers worldwide, in 2025 humans alone performed 47% of tasks, technology mainly performed 22% and a combination handled 30%, and by 2030 respondents expect roughly even proportions ⁹: stated expectations, not a measurement of what is automated. Stanford HAI's 2026 AI Index (2025 data) reports that 88% of the organizations surveyed globally say they have adopted AI in some form, with agent deployment still in the single digits in almost every function ⁶. Plenty of use, little end-to-end automation.

The available data does not support that reading, at least not yet. In that same working paper from the Center for Economic Studies at the U.S. Census Bureau, among AI-using companies that report any change in their task structure, 66% use it only to augment tasks and not to replace them; 52% of users report no effect on tasks at all. Headcount reductions are rare: 95% of users report no employment change from AI, and the 2% reporting a decrease is a slightly smaller share than the 2.3% reporting increases ⁵.

That does not mean the horizon is static. The McKinsey Global Institute estimates that, by 2030, activities accounting for up to 30% of the hours worked today in the United States could be automated (against 22% without generative AI) and that 12 million additional occupational transitions could be needed, with the declines concentrated in office support, customer service and food services ¹¹. Those are technically automatable hours, not eliminated jobs, and it is not a figure about Mexico. What we see in the first year is something else: the team stops doing data entry and starts reviewing exceptions.

1. 1

   Measure the process before you touch it (1 week)

   How many documents or emails per week, how many minutes each, how many errors. Without a baseline, the project can only be defended with anecdotes. This is the step most often skipped and the one that costs the most.

2. 2

   Automate a single step, not the whole process (3 weeks)

   Pick the step with the highest volume and the lowest consequence: almost always extraction or classification. Run it against real data, in parallel with the current process, with someone reviewing every output. That is where you see the real accuracy rate, which is never the demo's.

3. 3

   Define the confidence threshold and the fallback (2 weeks)

   Which cases go through on their own, which go to human review and what happens when the system is not sure. An automation is judged by how it fails, not by how it succeeds: without an exception path you do not have a system, you have a bet.

4. 4

   Connect the rest of the flow and watch the cost

   Integration with the ERP, the CRM or email; monitoring of latency, cost per document and quality; alerts when the accuracy rate drops. Only once this is stable do you move to the second process on the list.

An honest closing note on timelines: if someone offers to automate six processes in parallel in the first quarter, they are selling scope, not results. One process automated well and running stably is worth more than six half-done, because the second one takes half the time once you have already solved access, permissions and observability.