How finmid is scaling lending operations with AI

Launching an embedded lending programme is the easy part. What determines whether it scales is everything that happens after: the screening that runs before every payout, the operational request that arrives mid-onboarding, the collections case when a merchant falls behind. Most platforms underestimate this, and most infrastructure partners are quiet about it.

Scaling lending is an operations problem

When finmid started scaling across multiple platforms and markets, the manual processes that worked at lower volume started breaking. Cases took longer and the operations team was spending time executing process rather than improving it.

That is what prompted us to build the Agentic Ops team: a dedicated function that sits between operations and engineering, prototypes solutions, and owns the last-mile implementation of each workflow.

Three operational streams we automated with AI

1. Payout operations

Every merchant who accepted a pre-approved offer through a finmid-powered platform went through a semi-automatic screening process. Each case took between 5-20 minutes. A human would open the case, review the automatic results and perform manual steps on top of it, such as 'googling' the company and it's directors to validate good standing.

In the end, they would make a judgement call based on the information found. This last step of screening was the bottleneck - we saw it provides value in our decision making, yet it was unscalable with the human-only process.

First we considered replacing this step with vendors that specialize in these type of searches, but when running side-by-side tests, we quickly realized that the results were not exactly what we were aiming for.

We needed another approach that gave us better results and could scale at our pace.

Therefore, we supplemented the automatic process with an agentic workflow:

1. 1.

   We look up the merchant's record in official business registries to confirm the business actually exists in public records.

2. 2.

   We match the merchant's data the platform shares with us against what we found in the registry, to confirm the two align.

3. 3.

   We screen the merchant against global sanctions and PEP lists to ensure they are not sanctioned entities.

4. 4.

   We research the merchant online to build a fuller picture of who they are.

5. 5.

   We conduct a final risk review, applying the specific rules mandated by our risk team.

With the introduction of this workflow, the average screening time went down to 2 minutes, and most importantly, we improved the quality and consistency of the outputs while keeping it cheaper than using traditional vendors.

This is only one of the processes automated in the payout operations. Read about the other in How finmid built 10-minute payouts.

2. Collections

Not every merchant pays back their advance on time. In those cases, timely intervention, clear communication and consistency is key. Most companies rely on humans and more traditional approaches, which are hard to scale.

That is why we have an Agent Irina, that can be assigned to tickets by our Operations team. Irina can, for example, detect and analyze all recorded interactions with the merchant, past payment behavior, any prior agreements, communication guidelines and triangulate it against our internal policies.

3. Operational requests

Occasionally, merchants change bank accounts or phone numbers, information that is crucial to know to finmid to make sure the accepted capital offer lands in the right account or two-factor authentication message is sent to the right phone number.

Previously, platform would inform finmid over a shared Slack channel and finmid's Operations analyst would update the information in our back-office. This requires for platform to remember to update finmid every time there is a change and finmid to correctly update the change on our end.

Enter Agent Robert. Robert is a platform-facing agent that detects these operational requests, validates them, coordinates with partner platforms, and executes updates in finmid's back-office automatically.

Results

Each one of these workflows have significantly improved Operations team's efficiency, specifically:

1. 1.

   Our operations team handles double the payout amount daily, without adding more headcount.

2. 2.

   Monthly collection volume handled by the team increased 2.5 times, after the introduction of Irina

3. 3.

   Operations team saves 31 hours monthly by Robert handling manual operational requests

The overall change is not just automation of individual steps, but a shift in how operations run: from manual workflows to structured systems that can scale without adding complexity.

How do we make sure we can trust our agents?

An agent can make a mistake. That is why we have implemented guardrails to make sure we can trust our agents' work.

1. Two-step evaluation

We run a two-step evaluation on anything consequential where there is not a built-in human in the loop step by design. One model acts. A second model evaluates the output.

Additionally, because of the model design differences, we benefit from using multiple ones. Not all LLMs perform similarly on all tasks, and each model will follow the process in slightly different ways with the core outcome goal being the same. Both of the models have to agree before we treat the result as final.

2. New agents do not go straight into production

We run them in parallel with the existing human process first, comparing outputs, reviewing disagreements. Only when accuracy is consistently high do we reduce human oversight. We call these "green flows".

What this means for your platform

Every time a merchant accepts an offer for financing through your platform, something runs in the background: a screening check, a document review, a decision on whether to pay out. The speed and accuracy of those steps affect what your merchants experience.

When cases take 20 minutes, the volume you can process is capped. When vendors miss coverage gaps in your merchant base, legitimate merchants get delayed. When collections outreach is inconsistent, recovery rates fall.

The agents we have described here are not experiments. They are running on live cases, across 30 European markets, for some of the largest platforms in Europe. We built them because the alternative was a ceiling on what our platforms could offer their merchants.

If you are evaluating embedded lending infrastructure, the question is not only what a partner can launch. It is what they can run.

Curious to see how embedded lending could work for your platform? Book a demo.