In partnership with

Happy Thursday, Fintech Takers!

Spring is almost over and summer is almost here, which can only mean one thing … time for a sponsored deep dive essay!

You know the drill. Sponsored deep dive essays are opportunities for me to collaborate with specific companies on topics that interest them, me, and (hopefully) you. The sponsor provides invaluable input, but each essay is researched, written, and edited by me and is an honest reflection of my thoughts on the topic.

Today’s essay is sponsored by Spinwheel and the instruction they gave me was both simple and compelling: Write the Citrini Report for consumer credit, but do a better job than they did.

I’ve been fascinated by the intersection of AI and consumer credit for a while (as has Spinwheel), so obviously I accepted the challenge!

I hope you enjoy the result.  

- Alex

Was this email forwarded to you?

Subscribe Here

DEEP DIVE

Fixing Consumer Credit’s Human-in-the-Loop Problem

Everyone in financial services is talking about AI right now. And almost every one of those conversations — board meetings, conference panels, vendor pitches, regulatory roundtables — runs through the same core question: 

Where should we use AI, and where should we keep a human in the loop?

"Human in the loop" has become the governor on the car that is AI. A useful brake. A reassurance that we're not getting carried away. Yes, AI is powerful. Yes, we should use it. But let's remember — at least for now — to keep a human in there somewhere. Watching what AI is doing. Catching mistakes. Approving high-stakes decisions.

It's a logical framework for a lot of financial services.

Investment banking, where every deal is a bespoke construction and the work has always required senior human judgment, fits naturally. Commercial lending, where relationships still matter and underwriting still involves various on-the-ground assessments, fits it too. In both, the question of "Where to keep a human?" is a live debate, because there are still humans doing meaningful work everywhere in the process. The AI conversation in those corners of financial services is about deciding what to take from them.

Consumer credit is different.

For most consumer credit products — a $5,000 personal loan, a $2,000 BNPL transaction, a $20,000 revolving credit card line — automation has played a significant role, without humans involved in a meaningful way, for decades. 

The reason for this is simple: the math doesn’t allow for anything else. 

Consumer credit is a volume game played on thin unit economics. Cost-per-decision is the animating force, and any meaningful human time spent on an individual loan application torches the unit economics on contact.

The result of this relentless focus on automation and cost reduction over the last 30 years has been the selective elimination of three different groups of humans from most consumer credit workflows:

1. Lenders’ employees, except for exceptions and large-ticket work. 

2. Customers, except where the system needs free labor.

3. Regulators, except when the volume of consumer complaints gets too high.

Now, at this point, you may be saying, “Wait, why is this problem?”

After all, automation is a good thing, right? We want efficiency. We want lower costs. We want faster workflows. We want to free humans from paperwork and other forms of back-office drudgery … Don't we?

Yes, we certainly do.

The problem is that we didn't pursue automation in consumer credit thoughtfully. We removed humans wherever we could, without ever really stopping to ask whether each removal was the right call.

Apply that approach for long enough, and you produce a set of hidden costs that don't appear on the lender's P&L but show up everywhere else. In markets that don't quite work. In customer experiences nobody would actually choose. In regulatory oversight that's always a step or two behind.

Automation At All Costs

The fastest path to profitability in consumer credit is cost reduction. And the fastest path to meaningful cost reduction, if you are a consumer lender, is to lower your labor costs through automation.

However, the thing we have to remember about automation is that it doesn’t magically make the work go away. Oftentimes, the most effective approaches to automation aren't really about using technology to do the work more efficiently. They're about getting the work off the lender's labor budget, either by pushing it somewhere else, or by deciding the work doesn't have to be done at all.

In consumer credit, this tends to happen in three specific ways:

#1: Push the work onto the customer.

If your goal is cost reduction, the most straightforward solution is to push the work onto the customer.

Identity verification is a good example. If you are getting a loan, the lender first needs to confirm — for the purposes of both regulatory compliance and fraud prevention — that you are who you say you are. 

That work has to happen, but it’s expensive to do it well. So the lender pushes it onto you. Upload a photo of your driver's license. Take a selfie while holding it. Answer a knowledge-based authentication question about a car you sold in 2014. Verify a code sent to your phone. Re-verify all of that with every new lender, because no two of them trust each other's KYC.

Spinwheel calls this pattern "the human as the integration layer," which is the most honest description of it that I've come across. It shows up everywhere once you start looking. Income verification? Upload three months of pay stubs. Shopping for the best rate? Spend a Saturday filling out applications. You get the idea. Even creating a single view of your finances still relies on you to painstakingly connect every account. 

Lenders cleverly refer to this operating model as “self-service” and they treat it as a free input into their workflows, but it isn't free. It's just paid by someone whose time doesn't appear on the lender's P&L. 

And it has a second-order cost that’s even more important: negative selection. The customers most willing to grind through friction are disproportionately the customers most desperate to get the loan. The customers a lender would actually prefer are the ones most likely to give up at the first hint of friction.

#2: Outsource the work to middlemen.

When the customer won’t do the work and the lender can't afford to, the system finds a third party that will. The work gets externalized and the costs, paid by both the lender and the customer, become more pernicious.

Employment verification in mortgage lending is a good example. Mortgage lenders are required to verify a borrower's income before funding the loan, both as a risk-management discipline and because the GSEs and investors buying the loans demand it. Most of the time that can come from verified data pulled directly out of payroll systems. But plenty of cases — non-standard income, last-minute pre-close re-verifications, employers whose documentation doesn't pass muster — still come down to someone picking up the phone and calling the borrower's HR department to confirm they actually work there and earn what they say they earn. Equifax owns the operational infrastructure for most of this last-mile employment verification work; the HR phone numbers, the trained callers, the procedures. Lenders pay the company (a lot) for the service and those costs distort the mortgage market in real ways, all because lenders prefer an API they can plug into rather than employing a team of people.

Lead gen marketplaces — Credit Karma, NerdWallet, LendingTree — are another example. They do the product research consumers would otherwise have to do themselves, and they monetize it by selling leads to lenders. Which means the "best loans for you" rankings are not, in fact, the best loans for you. They're a leaderboard influenced by who's paying for placement. A perfectly rational business model for a company doing work nobody else will pay for, but one that distorts the market into a less efficient shape nonetheless. 

#3: Just don't do the work at all.

Sometimes lenders go further than outsourcing the work to a middleman. Sometimes they just decide the work won't happen at all.

The evolution of the U.S. credit bureau system is an instructive example. 

Customers should be deeply involved in the management of their own credit data: verifying that it's accurate, supervising what's being decided with it, and contesting errors before they propagate downstream. But when the bureaus first evolved into the entities we know today, giving customers that level of visibility and control wasn't economically feasible. Instead, we developed a system that operates almost entirely out of their line of sight.

It is frictionless, on the front end. Customers barely have to lift a finger to grant access to their data and lenders get a relatively comprehensive picture of their credit histories. 

However, when you remove the customer from the loop, you don't just create efficiency. You create a system the customer has almost no ability to correct, contest, or supervise. This results in downstream costs that are real but often underappreciated: bad data customers have to dispute, scoring decisions they can't see into, identity-theft recovery that becomes a part-time job.

And even when the technology emerges to change the status quo and to give customers more visibility into and control over how their data is used by lenders, it doesn’t eliminate the need to make difficult trade-offs.

The promise of open banking, for lending customers, was that they would finally get control and visibility over their financial data, the kind of involvement the bureau model had taken from them. 

And in that narrow sense, open banking has delivered. For the first time ever, a customer can grant explicit, granular permission for how their financial data is used to make credit decisions.

However, the trade-off for this increased visibility and control is that the customer has to do all the operational work themselves: authenticating each account through the lender’s data aggregator of choice, re-authenticating whenever sessions expire, and managing the consent and permissioning for all the third parties that want access.

Open banking corrected one part of the problem and worsened another at the same time. It put the customer back in the loop, in terms of visibility and control. But it doubled down on the customer as the unpaid integration layer. This is why the adoption of cash flow underwriting has been so slow, despite its immense analytic value and its customer friendliness. The additional friction it imposes on the customer cuts against everything else it has going for it.

The lesson of open banking is that consumer lending's automation problem isn't a single problem with one obvious fix. It's a knot. You can't pull on one strand without tightening another. Fixing the bureau-style consumer exclusion problem means making the customer-as-integration-layer problem worse. Better customer experience means relying on more middlemen.

It’s just an endless series of trade-offs.

But it doesn’t have to be.

Not anymore.

Untying the Knot

What's changed isn't the cost of automation. Every prior wave of consumer lending technology lowered the cost of automation. 

What's changed is something more fundamental: the cost of doing the kind of work humans used to have to do.

For decades, the work that lenders couldn't afford humans to do had to go somewhere. Onto the customer. Onto a middleman. Out of the system entirely. The work didn't disappear; lenders just couldn't keep employees doing it.

Large language models (LLMs), and the agentic AI systems being built around them, change that math. AI agents with the right harness can do the kind of work that used to require a human employee. A phone call to HR. A chat window. A web form filled out for the third time this month. A dispute letter sent by mail. If a competent human could do the work, an AI agent can now do it. At financial-services-grade reliability. At scale.

Look at what that does to each of the three patterns.

The customer doesn't have to be the integration layer anymore. Their AI agent can be. The identity verification, the income docs, the form-filling, the product research and comparison shopping, the consent flows all become work the customer's own agent does on their behalf. What used to take a Saturday afternoon becomes a 90-second conversation with an AI agent.

The business case for the middlemen weakens. Employment verification doesn't require an expensive third party tool. A lender's AI agent can call HR; an HR department's AI agent can answer. The verification work happens, the reliability bar gets met, and nobody has to pay a toll to a third party that positioned itself as the only one that could do the work. Same story for lead gen marketplaces: the research and comparison work happens through a chat with an AI agent, not by reading through a sponsored listicle of the best rewards credit cards.

The work that wasn't getting done finally gets done. Putting customers in the loop on their own credit data has always been theoretically desirable but practically unaffordable. AI can make it affordable. An AI agent could monitor a customer’s credit file continuously, flag anomalies, contest errors, supervise sharing decisions, and present the customer with the summaries they need to make decisions. The trade-off between "frictionless for the lender" and "customer in their own loop" stops being binding.

In each case, the work that used to require pushing to one of the three patterns now has a fourth option: do it directly, at scale, with an AI agent.

The Right Humans in the Right Loops

AI plays two roles in the new consumer credit stack.

The first one, which we've been describing, is as a human labor substitute. AI does the work humans used to have to do, or in some cases, work that nobody was doing at all. Customers stop having to be the integration layer. The expensive middlemen lose their rationale. The work that nobody was doing gets done.

The second role is less obvious and more interesting. AI is also a human labor orchestrator. 

By taking the operational work off the table, it creates room for the humans who got sidelined for cost reasons — lenders’ employees, customers, regulators — to come back into consumer lending workflows. Not in the loops they were pushed into out of necessity. In the loops where they actually provide additive value.

The customer comes out of the operational layer entirely. Their agent handles the data assembly, the product research and comparison shopping, the application data entry, the consent flows. The operational labor is gone.

Instead, the customer is looped in at the decision layer, where consideration and consent live. They don’t do the legwork of researching premium rewards credit cards, narrowing it down to the ones that are the best fit, and finding out which of them they’re qualified for. That’s the AI agent’s job. The customer just sees the output of that work — here are the three best premium rewards credit cards for you — interrogates it, and makes a decision, which the AI agent then goes and executes.

For lenders’ employees, there are two opportunities.

The first is that same decision layer, where human judgement really shines. Exception handling. Edge cases. Policy design. The 5% of decisions where the rulebook stops applying and someone has to think. The work that justifies the credit policy analyst’s salary.

The second opportunity, and the more interesting (though less obvious) one is customer service. 

Agentic AI systems are fast and frictionless. However, they can also feel abstract, opaque, and impersonal. They can feel inhuman. A customer whose AI agent just screwed something up doesn't want a smarter agent on the other end of the phone. They want a person who can take responsibility, listen, and fix it. Sometimes the human-in-the-loop isn't there because the work requires it. The human is there because the other humans want them there.

For regulators, the cost equation in consumer lending has been brutally difficult: a handful of humans trying to supervise an industry that processes millions of decisions a day. The math meant they could only ever look backwards — pull bureau data, review exam packages, read complaint patterns, reconstruct what already happened.

With agentic AI, regulators can shift from backwards-looking forensics to real-time supervision. They can watch the system as it operates rather than piecing it together from the data afterward. Anomalies surface before wide-spread harm occurs. The enforcement posture of the industry shifts from forensic to participatory.

This is what the human-in-the-loop question actually looks like in consumer lending. Not "where do we keep a human." Where do we put them back in?

The New Value Chain

Every part of the current value chain in consumer credit serves, in some way, the goal of getting work off the lender's labor budget, pushing it to customers, outsourcing it to middlemen, or skipping it entirely. It’s centered on the lender and the critical third-party service providers that it works with. The customer is shunted off to the side, only called upon when manual work like document uploads for identity verification are required. And the regulators are out on the periphery.

When AI changes that equation, the businesses built around the old patterns will be disrupted. Some companies will lose their moats. Some products and experiences that were infeasible to build before will become possible. The shape of the value chain itself will look different because it will be rebuilt around the customer; every component will sit downstream of their intent and every action taken on their behalf will be accountable to that intent.

Let’s quickly review some potential losers from this disruption:

The verification middlemen. A whole layer of the consumer credit stack exists because lenders need verification of real-world facts that couldn't be automated. Are borrowers actually employed where they say? Are the assets in their bank account really theirs? Does the property match the appraisal? These are last-mile questions that require phone calls, document handling, physical inspections, human-to-human contact. The companies that built operational infrastructure to handle that work at scale built durable moats on it. AI agents change that math. Work that required dedicated teams of trained humans interacting with the real world can now (in many cases) be done by AI agents with the right harness, and the moat that came from owning that operational infrastructure will start to erode.

The marketplaces (in their current consumer-facing form). Credit Karma and its peers are essentially curated browsing experiences for credit products. When the customer doesn't browse anymore — when their agent does the research — the marketplace as a consumer destination loses its primary user. The marketplaces that survive will pivot to being aggregation points for agents: lender-policy clearinghouses that AI shoppers query on behalf of customers.

The frictionless credit bureau model. The bureaus’ value proposition was always "we have the data the lender can't easily get." When the customer's agent can assemble that data from primary sources, with the customer's explicit permission, the bureau's monopoly on credit-decision data loses some of its grip, especially in use cases where real-time data provides a significant lift over static, backwards-looking bureau data. The traditional credit bureaus aren't going away, but they're going to have to compete with consumer-permissioned alternatives in ways they haven't had to before. This has already begun to happen, obviously, but AI will accelerate it.

And some potential winners:

Agent-driven data collection infrastructure. Companies like Spinwheel that can do consumer-permissioned data collection at financial-services-grade reliability — across every interface, from APIs to phone calls to chat windows — become part of a new and more important infrastructure layer. This is the unglamorous work of integrating across the fragmented universe of consumer credit data, doing it with verifiable consumer permission, and standing behind the data with the kind of accuracy that lenders and regulators can rely on.

Intent, accountability, & action infrastructure. Someone will need to build the infrastructure that takes raw AI capability and turns it into something a consumer lending workflow can rely on and action against. Permissioning, attestation, audit trails, intent verification, fail-safes against agents going off-script. Companies are hard at work building the protocols and harnesses necessary to do this for agentic commerce. There’s an opportunity for someone to build the same thing for agentic credit workflows.

Regulators with agentic capacity. Today's regulators are running with one hand tied behind their back: a handful of mystery shops a year, exam cycles of 12 to 18 months, complaint queues years deep. Agentic AI changes all of it. Regulators can mystery shop at industry-wide scale, deploying synthetic borrower personas to test fair lending compliance and pricing consistency across thousands of lenders at once. They can work consumer complaints in real time, surfacing systemic patterns as they emerge. They might even one day move from periodic exams to streaming supervision, watching the lenders continuously rather than dropping in once a year.

This is the new shape of the industry. And it's just the near term.

Where This Goes Next

Everything we've described so far assumes one group at a time gets agentic capabilities while the rest of the system stays roughly as it is. The customer's agent calls today's lenders. The lender's agent reads today's credit bureau data. 

That's a transitional state, not the end state.

The end state is agent-to-agent interactions across the whole consumer credit system. 

As Spinwheel states, “Agents are the next evolution beyond LLMs. We’re not talking about just a financial version of ChatGPT. The future of AI will mean agents that can fetch data, verify it, update it, and take action on it instead of just chatting about it.”

Customer agents talking to lender agents. Customer agents talking to bureau agents. Customer agents filing complaints with regulator agents. Lender agents responding to regulator agents in real time. Regulator agents mystery shopping with bureau agents. Cascading chains of agents handing intent down the line, through brokers and middleware platforms and whatever new layers emerge to fit this new ecosystem.

We can expect that new, agent-to-agent consumer credit ecosystem to be much faster and more complex than our current ecosystem. It’s hard to say exactly what will change, once agents have disintermediated both sides of every consumer credit transaction, but here are a few educated guesses:  

The funnel doesn't just compress. It dissolves. Today's loan application exists because it's the handoff format between humans and machines, but operating at human speed. When the handoff is between two AI agents at machine speed, there's no reason for the structure to persist. The "application" stops being a thing the customer fills out and becomes a thing two agents negotiate. Approval cycles that took days take seconds.

The negotiation surface expands. Agent-to-agent isn't just "read." It's agents negotiating rates, terms, and conditions on behalf of both sides, at machine speed. Multiple lender agents can compete for the customer's business in parallel. The asymmetric information dynamics that have favored lenders for decades may start to compress because the customer's agent has both the time and the analytic horsepower to actually compare offers.

The translation layer disappears. A whole layer of the consumer credit stack exists because human stages of the process need translators between them. Brokers translate between customers and lenders. Loan officers translate between customers and underwriters. Customer service translates between customers and operations. When every stage is agent-mediated, many of those translation jobs lose their reason for existing.

Those changes will, if they come to pass, cause chaos. But there is opportunity in chaos, for those that can recognize it.

That said, it's also important to point out the potential downsides in this agent-mediated future if the risks aren't managed well. And there are indeed some scary-sounding risks:

Intent drift. When a customer's intent passes through one agent, the most likely failure mode is the agent misunderstanding it. When it passes through a chain — customer agent to broker agent to lender agent to bureau agent and back — the failure mode is intent drift. Each handoff is a potential point of misunderstanding. Without infrastructure that captures and verifies intent end-to-end, the customer's actual request gets quietly degraded as it moves through the system.

The new fraud surface. Today's fraud problems in consumer credit mostly involve humans being deceived: by fake websites, by social engineering, by synthetic identities. Agent-to-agent introduces a new category. If you can spoof an agent, you can spoof the chain of trust. Every agent has to be able to verify every other agent it interacts with. Identity attestation across agent boundaries stops being a feature and starts being a condition for participating in the system at all.

The intelligence gap. AI agents cost money to run. Token costs are coming down, but they're not free. The customers who most need agents working on their behalf — thin-file consumers, identity-theft victims, people navigating disputes with lenders — are also likely to be the customers least able to afford the agents that would help them. Without thoughtful design, agentic consumer credit could produce a new shape of financial inequality with the same outline as the old ones.

Closing the Loop

The question of where to involve humans in consumer credit workflows has always, at the end of the day, been an economic question.

When we talk about straight-through processing, what we’re really talking about is removing a lender’s employees from a workflow for a credit product that doesn’t have the unit economics to support their involvement. When we talk about self-service, what we’re really talking about is how much time and effort customers will be willing to give when a lender outsources its required verification work back to them. When we talk about the benefits of low-friction data infrastructure, what we’re really talking about is the trade-off that lenders are making, on behalf of their customers, between convenience and control.

Many of the decisions that lenders have made in these areas over the last three decades have been determined not by what would be most valuable for customers, lenders, and regulators, but rather by what is economically feasible: what can we afford to do?    

AI fundamentally alters this decision-making dynamic by making it economically possible to put the right humans in the right loops. Customers at the decisions that matter. Employees for exceptions and accountability. Regulators in the system in real time.

With AI, we can design a consumer credit system that finds a more optimal point between customer value, lender profitability, and regulatory safety and soundness.

That outcome is not guaranteed. AI brings a boatload of new risks to the table. However, those risks are coming, whether we want them to or not, and the best way to navigate those risks is by leveraging the technology to thoughtfully add humans back into the loop — with the right AI guardrails and frameworks to drive durable success.

ABOUT SPONSORED DEEP DIVES

Sponsored Deep Dives are essays sponsored by a very-carefully-curated list of companies (selected by me), in which I write about topics of mutual interest to me, the sponsoring company, and (most importantly) you, the audience. If you have any questions or feedback on these sponsored deep dives, please DM me on Twitter or LinkedIn.

Today’s Sponsored Deep Dive was brought to you by Spinwheel.

Spinwheel is the future of consumer credit data and payments. Spinwheel tackles inefficiencies in the consumer debt lifecycle by reducing unnecessary friction and costs. Our Debt APIs provide real-time, verified consumer and credit data and the ability to process payments inside of your brand's experience....all with just a phone number and birthdate!

Thanks for the read! Let me know what you thought by replying back to this email.

— Alex