Where should a lender start with AI agents?

What criteria should you use to pick the first process?

• Volume: lots of similar interactions every week. That's where an agent makes a visible difference.
• Clear rules: your team can explain what happens in most cases and when to escalate.
• Reachable systems: the data the agent needs is available via API, a read replica or some other way in.
• Measurable outcome: you know where you are today and can compare afterwards.
• Containable risk: a mistake can be caught and fixed without serious harm; the big decisions stay with a person.
• A motivated internal owner: someone who knows the process and wants to improve it.

Which processes usually fit those criteria best?

Pre-due and early-delinquency collections

Reminders before the due date, a heads-up on payment day, first contacts on late accounts, payment references and recorded payment promises. It's repetitive, time-sensitive and governed by clear conduct rules. See AI collections.

Loan-file building

Requesting documents, checking they're complete and current, and organizing them for credit and compliance. Analysts stop chasing paperwork and the committee gets better files. See AI origination and KYC and AML.

Customer inquiries

Balance, due date, payment references, application status. It's the volume that floods customer service, and nearly all of it lives in the core. See customer service and the UNE.

What should wait?

• Fully automated, end-to-end credit decisions.
• Restructurings, write-downs and complex negotiations.
• Processes whose data is scattered across ownerless spreadsheets.
• Any process where nobody on your side can make time to define rules and review results.

Not because an agent can't take part, but because these need mature rules, earned trust and governance that already works.

Does it depend on the type of institution?

Partly. A consumer-lending SOFOM (a Mexican non-bank lender) tends to have more volume in collections and customer service; a leasing company or an SME lender leans more on file building and application follow-up; a SOFIPO (a regulated savings and loan company) also serves savers. An IFPE (a licensed e-money institution under Mexico's Fintech Law) often starts with customer service and customer verification. What doesn't change are the criteria: volume, clear rules, reachable systems and measurement. See how we work with each license type in solutions by license type.

What comes after the first agent?

Once the first process is running, the connections, the log, the Agentic Office and the habit of reviewing are already in place. The second agent takes less effort, and it's often the one that coordinates with the first: origination hands collections a clean file; customer service passes a payment promise to collections. That's how an agentic ecosystem gets built — one process at a time, with your team in charge.

How do I kick this off?

With an assessment of your operation: which processes you run, how much volume they carry, which systems you use and where time gets lost. You can start by messaging Luz, our Head of Sales, right here, or check out the initial assessment.

Short answers

Should we launch several agents at once?

Usually not. One well-built agent on a narrow process teaches you more than several half-built ones, and it leaves the connections and governance ready for the next.

Does our data need to be perfect before we start?

Not perfect, but reachable and reasonably reliable for the process you pick. That’s exactly what the initial assessment is for.