I spend quite a bit of time thinking about Financial Foundations, particularly in the collections industry. I am also one of the biggest advocates for the very real operational gains AI tools are already introducing into our businesses.

As efficient and cost effective as many of these tools have become, however, I increasingly find myself wrestling with an emerging collision between venture-backed AI thinking and the operational realities of regulated collections.

Over the past several years, venture-backed AI firms have increasingly turned their attention toward the collections industry with ambitious modernization plans. The founders are not traditional collections executives. They come from elite AI and contact-center software backgrounds. Their prior transformational success was built around conversational AI, workflow optimization, agent-assist tools, and intelligent routing platforms for modern customer contact operations.

Collections looks like a perfect opportunity.

The industry is labor intensive. Operationally inefficient. Built on aging systems. Full of repetitive consumer interactions. Burdened with high staffing costs and outdated workflows. Exactly the type of environment modern AI should transform.

So they acquire a collection agency and get to work 

The strategic logic is sound — at least initially. Financially, the results look extraordinary. The venture rapidly installs AI-enabled IVR systems, assignment engines, workflow orchestration tools, and automated contact infrastructure. Headcount is reduced dramatically. Entire layers of collector labor disappeared almost overnight. Gross margins that historically lived in the 20%–25% range suddenly approach 50%.

To venture investors, this looks like proof.

The collections industry appeared massively overstaffed. AI clearly could automate large portions of traditional collection activity, particularly in smaller balance portfolios where interactions were repetitive and economically unsuitable for expensive human collectors.

From the outside, the thesis appears validated. But there is also another side to these transactions that outsiders often miss entirely.

The veteran collection agency founders who sold to the venture-backed AI operators were not blind to the weaknesses of their own infrastructure. Quite the opposite. They lived inside the fragmentation every day. They understood better than anyone:

• the decades-old patchwork systems,
• the fragile custom integrations,
• the manual reconciliations,
• the spreadsheet dependencies,
• the operational workarounds,
• the client-specific exceptions,
• and the growing burden of maintaining aging architectures never originally designed for modern regulatory complexity.

Many of these founders spent decades incrementally adapting their systems to survive:

• evolving CFPB expectations,
• bank audit requirements,
• client-specific controls,
• trust accounting scrutiny,
• litigation demands,
• and increasingly customized creditor reporting requirements.

The systems became operationally resilient, but rarely elegant. And many of the industry founders genuinely believed the partnership with elite AI technologists could finally solve the problem they themselves had long struggled to modernize. Their vision was not simply: “replace collectors with AI.” The deeper aspiration was far more ambitious.

Combine decades of operational knowledge, intimate understanding of reconciliation realities, awareness of regulatory edge cases, and knowledge of where the operational pain truly lived with: modern AI tooling, workflow orchestration, intelligent process mapping, and advanced engineering capability.

Together, perhaps they could finally build the missing layer the industry never truly possessed – a modern, coordinated, flexible system of record capable of surviving the increasingly complex operational and regulatory environment surrounding consumer debt servicing.

For many of the agency founders, this was the real incentive for doing the deal. Not merely liquidity. Not simply margin expansion. But the belief that the industry’s foundational operational infrastructure could finally be rebuilt correctly.

At the same time, the AI technologists viewed the opportunity through a very different lens.

They saw an industry buried under labor inefficiency, repetitive workflows, and legacy operational drag. To them, the visible consumer interaction layer looked dramatically under-optimized: outreach, dialing, texting, workflow routing, assignment logic, RPC optimization, and collector productivity.

And unlike the ARM veterans, they came from environments where rapid iteration, sprint-based development, AI-assisted experimentation, and fast deployment cycles were not merely accepted — they were expected.

From their perspective, many of the legacy workflows appeared unnecessarily manual and technologically primitive.

Both sides believed the other possessed the missing capability required to modernize the industry. The ARM veterans believed the technologists could finally rebuild the fragmented operational foundation underneath collections. The technologists believed the collections industry could be modernized much faster than the incumbents imagined.

Both perspectives contained truth. And both perspectives carried blind spots.

The venture-backed technologists naturally prioritized:

• visible deployment,
• outreach optimization,
• AI interaction layers,
• rapid iteration,
• staffing reduction,
• and measurable short-term operational gains.

Those wins were easier to demonstrate:

• AI IVR,
• texting,
• assignment engines,
• digital workflows,
• conversational automation,
• and collector reduction.

The legacy system-of-record rebuild underneath was far less visible. It required painstaking workflow mapping, reconstruction of invisible control logic, normalization of decades of operational exceptions, deterministic reconciliation architecture, and long-duration implementation cycles with uncertain short-term financial optics.

At the same time, the legacy ARM founders themselves were shaped by decades of operational survival. Their world was not built around startup iteration cycles. It was built around pragmatic incrementalism: patching the latest regulatory issue; satisfying the newest bank audit requirement; adapting to a creditor’s custom process; and layering operational controls onto aging systems one requirement at a time.

One culture came from sprint-based startup development driven by rapid experimentation and AI-assisted iteration. The other came from decades of methodically inserting pragmatic controls to survive another regulator review, client audit, or operational edge case.

Both approaches solved immediate problems. Neither naturally favored the patience required for a long-horizon foundational rebuild of the system of record itself.

And here the deeper operational reality slowly emerged. The founders viewed collections primarily as a contact-center problem. In reality, collections is a regulated cash-control and financial operations business wrapped around a contact center.

That distinction matters enormously.

The visible inefficiencies inside old collection agencies — spreadsheets, manual reconciliations, fragmented workflows, duplicated controls, ugly legacy systems — often exist because they were built over decades of surviving bank audits, client reviews, regulator examinations, litigation challenges, trust account scrutiny, and operational failures.

Much of the “technical debt” was not simply poor engineering. It was operational scar tissue. The AI systems integrated quickly into the visible workflow layer:

• outreach,
• assignment,
• RPC optimization,
• IVR containment,
• conversational orchestration.

But the deeper systems of record remained old, fragmented, and highly customized over decades. And here the real challenge emerged.

The industry may be underestimating how much invisible operational logic lives inside the legacy infrastructure: suppression rules, payment timing dependencies, remittance controls, settlement logic, client-specific requirements, audit trails, reconciliation sequencing, legal workflows, and compensating controls built after years of prior failures.

The danger is not that the AI tools fail visibly. The danger is subtler. Modern AI orchestration layers risk accidentally bypassing old operational controls that nobody fully documented because they evolved organically over decades.

That is survivable in a startup environment. It is potentially catastrophic inside regulated financial services. Large banks and credit issuers audit collection agencies multiple times per year. They require audited financial statements. They review operational controls in detail. They care less about elegant technology than deterministic auditability and operational trust. And unlike startup software customers, they do not tolerate “beta” operational behavior.

At the same time, another strategic contradiction quietly emerged. Although the AI systems dramatically improved internal profitability, collection performance versus competing agencies remained only average. The automation worked best in smaller-balance, highly repetitive portfolios. But in larger-balance accounts requiring negotiation, discretion, hardship evaluation, and nuanced human interaction, traditional agencies still often outperformed through experienced collector populations and broader communication strategies.

So while the AI-enabled agency became more operationally efficient, it did not necessarily become the best-performing agency for creditors.

That distinction matters because creditors eventually notice margins. The largest issuers benchmark agencies constantly, audit their operations, review their financial statements, and understand operating leverage.

If AI materially lowers agency labor costs while liquidation performance remains only broadly competitive, clients eventually ask a rational question: Why are we still paying traditional contingency economics?

This creates a difficult asymmetry for AI-enabled collection agencies. The operational upside initially belongs to the agency. Over time, however, competitive pricing pressure likely transfers much of that efficiency benefit back to the creditor clients through lower contingency fees and tighter pricing negotiations.

In other words, success itself creates pricing pressure. And this naturally pushes technology-oriented operators toward purchased debt.

At first glance, purchased debt appears to solve the strategic problem elegantly. If the agency owns the receivables itself there is no external creditor demanding a share of the AI-created efficiency gains. There is no placement dependency. The economics of superior operational tooling remain internalized.

To a technologist, this looks like a logical extension of the same thesis. If superior AI dramatically lowers the cost-to-collect, why merely service debt for others when you can own the inventory directly?

On the surface, purchased debt appears to be vertically integrated collections. But this is where the business model transition becomes deceptively dangerous.

Purchased debt is not simply “third-party collections with ownership.” It is fundamentally a balance sheet and asset management business. Third-party collections is primarily an operational service model. Purchased debt shifts the center of economic gravity toward:

• portfolio pricing,
• liquidity forecasting,
• yield assumptions,
• capital structure,
• financing access,
• amortization modeling,
• and disciplined asset valuation.

The collection operation itself becomes secondary to buying debt correctly.

This is difficult for many technology-oriented operators to internalize because the AI tools genuinely can improve collection efficiency, workflow automation, segmentation, and labor productivity. But in purchased debt, superior operational tooling cannot rescue consistently poor purchasing discipline.

A debt buyer that marginally overpays in competitive auctions can erase years of operational efficiency gains. A mediocre collector buying portfolios intelligently can outperform an elite AI-enabled operator consistently buying at overly optimistic prices.

And unlike third-party collections, purchased debt introduces:

• substantial capital requirements,
• balance sheet risk,
• financing structures,
• impairment exposure,
• and long-duration cash flow curves requiring continuous reinvestment.

Revenue is no longer simply earned as collections occur. It is recognized over the projected life of the asset itself using assumptions around expected future recoveries. That is an entirely different economic model.

The irony is that the move into purchased debt appears strategically logical precisely because AI success in the third-party collection model creates pricing pressure that limits the long-term economics of the agency business itself.

But escaping the pricing pressure of third-party collections may simultaneously expose the venture to an entirely different category of risk: capital allocation risk.

And that transition — from operational optimization to asset valuation discipline — may ultimately prove far more difficult than automating the contact center layer itself.

None of this means the underlying AI thesis is wrong. In many ways, the opposite may be true. The collections industry likely is ripe for significant technological modernization.

The core operational challenge is not conceptually mysterious. At its heart, collections is fundamentally:

• tracking consumer obligations,
• managing client inventory,
• controlling the movement of cash,
• reconciling balances from payment through remittance,
• and maintaining deterministic auditability across the entire lifecycle.

Those workflows are complicated operationally, but they are not unknowable.

And AI may ultimately be uniquely suited to help map, normalize, monitor, and orchestrate the fragmented processes, spreadsheets, controls, and exception handling logic that evolved across the industry over decades.

At the moment, however, much of the AI investment flooding into the industry is concentrated on the visible consumer-facing layer. A growing ecosystem of slick vendor tools is being aggressively marketed:

• AI IVR platforms,
• intelligent texting systems,
• email collection tools,
• self-service payment portals,
• conversational AI outreach,
• agent-assist overlays,
• and digital engagement workflows.

Many of these tools genuinely create value. They can improve contact rates, reduce staffing costs, automate repetitive interactions, and modernize the consumer experience relatively quickly.

But these are also the “easy” wins. They operate primarily at the communication layer surrounding the actual financial infrastructure. The harder problem sits underneath.

The real long-term transformation requires redesigning the system-of-record foundation itself:

• consumer-level balance tracking,
• client-level inventory controls,
• payment allocation logic,
• remittance workflows,
• reconciliation sequencing,
• audit traceability,
• exception handling,
• and resilient control structures capable of adapting over time.

And unlike the visible outreach layer, these systems are not simply plug-and-play deployments. Even with powerful AI tooling, the complexity comes from the operational interconnectedness of the environment itself.

A modernized platform must handle:

• multiple creditor clients simultaneously,
• consumers with obligations across different clients,
• portfolio transfers,
• legal status changes,
• client-specific business rules,
• trust accounting requirements,
• payment reversals,
• settlement adjustments,
• and remittance structures that often do not align neatly one-to-one with collection performance metrics.

The challenge is not that these workflows are impossible to model. In many ways, AI may ultimately make them easier to model than ever before. The challenge is building resilient infrastructure capable of handling those realities consistently, transparently, and auditably over time.

The collections industry may ultimately resemble the banking industry more than the modern SaaS world. Many of the largest banks still operate core systems on decades-old mainframe infrastructure. From the outside, those environments often appear archaic with fragmented interfaces, layers of middleware, manual reconciliations, operational workarounds, and deeply embedded legacy code.

Yet those systems survived because they proved resilient across regulatory scrutiny, audit requirements, massive transaction volumes, fraud exposure, operational edge cases, and long-duration financial accountability.

The modern AI opportunity inside collections may ultimately look very similar.

The visible consumer-facing layer: AI IVR; texting; portals; digital outreach; conversational interfaces; and workflow overlays can be modernized relatively quickly.

But the true long-term opportunity sits deeper — rebuilding the operational system-of-record foundation itself using modern AI-assisted workflow mapping, orchestration, reconciliation, and control architecture.

And like banking modernization, the economic payoff likely arrives over many years rather than venture-style quarters. That reality creates an uncomfortable tension for many venture-backed AI companies.

Venture economics naturally favor rapid deployment, visible adoption curves, fast ARR growth, and short-cycle proof points.

But foundational financial infrastructure transformation behaves differently. The real work involves mapping invisible operational logic, normalizing fragmented workflows, rebuilding deterministic control structures, preserving auditability, and designing systems capable of evolving safely over decades.

Those projects are extraordinarily valuable once completed. But they are not quick. They require patient capital, operational humility, and a willingness to prioritize long-term infrastructure resiliency over immediate product velocity.

Ironically, that may be why so few firms pursue the deeper opportunity seriously.

The consumer-facing AI layer is easier to demonstrate, easier to sell, and easier to monetize quickly. The invisible operational foundation underneath is harder, slower, less glamorous, and operationally exhausting.

But over time, that foundational layer is likely where the real long-term enterprise value will ultimately reside. The firms that eventually win may not be the ones that simply automate consumer conversations fastest.

They may be the firms patient enough to rebuild the invisible financial operating infrastructure underneath the industry itself.