What a Collections Agent Actually Does All Day, And Why That's the Problem
June 14, 2026 · 9 min read
Debt collections, as practiced today, is a workflow designed around the constraints of 1987. Phone calls, paper files, human memory. The technology has been layered on top, but the architecture of the work has not changed.
A first-principles look at why one of finance's oldest workflows has never actually been solved and what happens when you finally do.
Let's start with something the industry doesn't say out loud.
Debt collections, as practiced today at the overwhelming majority of lenders, fintechs, and financial institutions worldwide, is a workflow designed around the constraints of 1987. Phone calls. Paper files. Human memory as the primary knowledge management system. The technology has been layered on top: dialers, CRMs, ticketing systems. But the fundamental architecture of the work hasn't changed. You hire people. You give them accounts. They call. They log. They move on.
For thirty-plus years, this has been treated as a solved problem. It isn't. It never was. It was just a problem expensive enough to live with and complex enough that nobody wanted to rethink from scratch.
That's changing. But to understand why, you first have to understand what is actually happening inside a collections operation. Not at the dashboard level. At the level of what a human being does when they sit down to recover money on behalf of a lender.
The anatomy of a shift
A collections agent arrives at their desk. They log into their dialer. They log into their CRM. They log into the lender's core banking system or servicing platform. In many operations, they also have a compliance reference document open: sometimes a live portal, more often a shared drive folder with a filename like Policy_v7_FINAL_revised_March.pdf. The rules about what they can offer, when, and to whom change often enough that no one has them memorized. Not often enough that anyone has built a proper system to surface them in real time either.
They pull their queue. Anywhere from 80 to 200 accounts depending on the operation. Each account has a balance, a due date, a number of prior contact attempts, a set of notes from previous agents, and a status. The agent's job is to work through this queue and produce outcomes: a payment, a payment plan, a promise to pay, a hardship arrangement, a dispute flag, or a closure.
They start dialing.
No answer. Log the attempt. Move on. No answer. Log the attempt. Move on. Voicemail. Leave a message or not, policy varies, and the agent has to remember which jurisdiction this account falls under before deciding. Move on.
Then someone picks up.
Now the agent is doing something genuinely complex. They are simultaneously reading account details on one screen, cross-referencing the policy document on another, conducting a real conversation with someone who is likely stressed or evasive or hostile, making a live judgment call about what to offer and when to push versus when to flex, and mentally tracking whether anything about this call triggers a compliance flag: a hardship indicator, a vulnerability marker, a dispute.
After the call, they document it. Manually. In a CRM field designed for a different use case. They set a follow-up. They hope the note is clear enough that the next agent who touches this account, tomorrow or three weeks from now, understands the context.
Then they go back to dialing.
On a good day, a skilled agent makes 60 to 80 outbound attempts. They reach 15 to 20 people. They convert maybe 8 to 12 of those into some kind of positive outcome. The rest are broken promises, missed callbacks, disputes, disconnected numbers, or silent accounts sitting in the queue while the balance ages and recovery probability drops with every passing week.
This is not dysfunction. This is the process performing as designed.
The thing everyone gets wrong about collections costs
When finance leaders look at collections operations, they see a headcount problem. The unit economics are straightforward: X accounts in arrears requires Y agents to work them, at Z cost per agent. Volume goes up, hire more people. Volume goes down, right-size. Linear, legible, budgetable.
This framing obscures where the real money is going.
The visible cost is salaries, benefits, and seat licenses. The invisible cost is everything around the agent that makes them less effective than they should be.
Consider the policy lookup problem. An experienced agent in a mid-sized consumer lending operation works accounts across multiple products: personal loans, credit cards, auto, maybe BNPL. Each product has different recovery rules. Each jurisdiction has different regulatory requirements. The rules change. Agents learn them through training, experience, asking colleagues, skimming that PDF on the shared drive. The gap between what the policy says and what the agent does on any given call is never zero. Every time an agent offers something they shouldn't, or fails to offer something they could have, money is left on the table or compliance risk is created.
Consider the prioritization problem. A queue of 150 accounts is not a flat list of equal opportunities. Some are high-balance borrowers in a momentary liquidity crunch who will pay in full if contacted today. Some are chronic defaulters where the only realistic outcome is a settled arrangement at 40 cents on the dollar. Some are borderline cases where a hardship restructure now prevents a total writeoff in six months. Treating them as a sequential queue, working top to bottom or however the CRM happens to sort them, leaves significant recovery value on the floor.
Consider the follow-through problem. A promise to pay is a contract between the agent and the borrower. It is also a piece of data that lives in a CRM note and depends entirely on the next agent who touches the account finding it, understanding it, and acting on it correctly. The rate at which promise-to-pay commitments collapse, not because the borrower defaulted again but because the operational handoff failed, is one of the dirtiest open secrets in collections. Nobody tracks it rigorously. Nobody wants to.
Add these up: policy leakage, prioritization waste, follow-through failure. You get a number almost certainly larger than the visible headcount cost. Because it's distributed across thousands of calls and never shows up as a line item, it gets ignored. The operation looks like it's running fine. The dashboard says so.
Why this problem has been so durable
There's a reasonable question here. If this is all true, why hasn't it been solved? Collections is not a new industry. There are massive software vendors who have been selling into this space for decades. There are consultants, auditors, and industry bodies. Everyone knows collections operations are inefficient. Why does the problem persist?
Three reasons, and they compound each other.
First, the talent trap. Collections has historically been a high-turnover, low-prestige function. The people who know how it really works are agents and floor managers. The people who make technology decisions are executives. There is a structural gap between operational reality and strategic decision-making that makes it hard to accurately diagnose what's broken, let alone fix it.
Second, the integration problem. A collections operation runs on four or five systems at minimum: a dialer, a CRM, a servicing platform, a compliance tool, and some kind of reporting layer. None of these were designed to work together. Each has its own data model, its own API surface, its own vendor relationship. Building a coherent operational layer on top of this stack requires either a massive custom engineering effort or a willingness to accept the gaps. Most operations accept the gaps.
Third, the problem was never painful enough to force a rethink. Collections is a cost center. When it underperforms, the loss is diffuse: slightly worse recovery rates, slightly higher headcount, slightly more compliance incidents. Nobody's career ends because the promise-to-pay conversion rate is 62% instead of 74%. The organization accepts the leakage as the cost of doing business.
What changes this calculus is scale. As consumer credit portfolios grow, as BNPL and embedded finance push credit access deeper into the economy, as interest rate environments put more borrowers into distress, the leakage compounds. A 10% improvement in recovery rates on a $50M distressed book is a $5M swing. That number gets executive attention in a way that "our agents are spending four minutes per account on policy lookups" never did.
What actually needs to happen
A collections operation is, at its core, a decision engine with a human communication layer on top. Every action: who to contact, when, through which channel, with what offer, under what compliance constraints, with what follow-up cadence. These are decisions. Most of them have better and worse answers derivable from data. The human layer: the actual conversation, the empathy, the real-time judgment about how a borrower is responding. That part is genuinely valuable and genuinely hard to replicate.
The problem is that today, humans are doing both jobs. They are making the decisions and having the conversations. Because the decision-making consumes so much cognitive load: the policy lookup, the prioritization judgment, the compliance check, the documentation. The conversation suffers. Agents are context-switching constantly. Working from incomplete information. Making calls they shouldn't have to make, on the fly, under queue pressure.
The correct architecture separates these two jobs cleanly.
The decision layer: prioritization, channel selection, offer determination, compliance validation, follow-up scheduling, promise-to-pay tracking. Systems built for data problems should handle data problems. Humans are poor at data problems at scale, under time pressure, without complete information access.
The conversation layer: the actual call, the negotiation, the human moment. This is where agents should spend their time, and where organizations should invest in quality. An agent freed from 40% of their shift spent on administrative tasks is an agent who can have better conversations, handle more nuance, and close more accounts.
The technology to do this exists. What hasn't existed until recently is an architecture that integrates it end-to-end into a collection's workflow without a three-year implementation project and a $10M systems integrator.
The agent's job, redefined
The version of collections that is coming doesn't have fewer humans. It has differently deployed humans.
The queue arrives pre-prioritized by recovery probability, with a recommended contact channel, a compliance-validated offer range, and a synthesized context brief of every prior interaction. The agent picks up the phone with everything they need already in front of them.
The policy document is embedded in the system that generates the offer range. The agent never looks it up. They can't misconstrue it. Compliance risk stops depending on memory.
The promise to pay is tracked, validated, and surfaced automatically. If the borrower misses their commitment date, the system flags it before anyone has to check.
What remains for the agent is the only part that actually requires a human. The moment on the call where the borrower says they just lost their job. The negotiation where the right outcome requires reading tone and building trust. The edge case that falls outside any model's training data.
That is a better job. For the lender, it is also a fundamentally different unit economics equation. More accounts worked per agent. Higher recovery rates per account. Lower compliance risk per interaction. Less operational overhead per dollar recovered.
The industry has been accepting a worse version of this outcome for decades. The alternative now exists.
