People sometimes ask when Cardo decided to add the “AI” to its name. The answer is that we never added it. The AI has been there since day one. When we founded the company seven years ago, plenty of people told us we were crazy, that the AI part was a marketing flourish on a fintech idea, that the asset-based finance and private credit market would not be ready for it for a generation, that the buy side would never trust a machine with anything that mattered. We disagreed then, and we disagree now, and what was a conviction in 2018 has become reality in 2026.

Asset-based finance and private credit are no longer niche. Together they represent one of the largest pools of risk capital in the world, financing everything from small-business invoices to consumer auto loans, from houses and solar installations to the GPU clusters powering the AI economy itself. The market is now measured in tens of trillions of dollars.

And yet, the way this market is operated has not kept pace with what it has become. Decisions worth hundreds of millions of dollars are still made on the back of spreadsheets emailed between counterparties. Loan tapes arrive in dozens of incompatible formats. Covenants live as paragraphs of legal prose rather than as executable rules. Risk teams discover deterioration months or quarters after the data was already there to see it.

We are writing this manifesto for two reasons. First, because we believe this transition will be the defining competitive event of the next decade in private markets, and we want to be honest about both the opportunity and the risks. Second, because we have spent the last eight years at Cardo AI building, deploying, and operating exactly this kind of infrastructure with some of the most demanding investors and lenders in the world, and what we have learned is worth saying out loud.

Before we talk about what is broken in this market, it is worth pausing on what this market actually is, because the question I am asked most often by allocators new to the asset class is whether ABF is still niche or whether it has crossed over into mainstream. My answer is that it is rapidly mainstreaming, and the easiest way to see it is to walk through an ordinary day.

You wake up in the morning, and there is a roof over your head, that is mortgage finance. You go to work, and you use the car to get there: auto finance. You order something online: revenue-based finance for the merchants on the e-commerce platforms, and equipment finance for the fleets of trucks. You use AI tools at work: data centres and GPU financing. You pick up your phone: cell-tower financing. The day you have just lived was financed, in some non-trivial part, by asset-based finance.

The honest description of the current state of the market is not flattering. Most asset-based finance and private credit operations today are running on a stack that looks like this: a credit agreement in PDF, a loan tape in Excel arriving via email or SFTP each month, a borrowing base computed in another spreadsheet, a covenant tracker maintained by an analyst, a waterfall that lives in a fourth spreadsheet, and an investor reporting pack assembled in PowerPoint at the end of each quarter. Between these artefacts sit human beings reconciling numbers that should never have been allowed to disagree in the first place.

What unites these failures is that none of them are really failures of judgement. They are failures of plumbing. The credit officer is not making the wrong call; the credit officer simply does not see the evidence in time.

Across the deployments we have lived through with our clients, the value of AI in private credit and ABF arrives in three waves. They are sequential and they are cumulative. Skipping ahead does not work.

Wave one: time-to-decision compression. A loan tape that used to take three days to crack now takes three hours. A credit memo first draft that used to take a week now takes a morning. In a market where the manager who delivers a credible term sheet first wins the deal, time compression is a direct revenue lever.

Wave two: surveillance density. Once data flows through a system rather than an inbox, you do not have to wait for month-end to know how the portfolio is doing. You can watch all credits, all the time, against the same set of indicators. The financial value of surveillance density is asymmetric, small in good times and very large in bad times.

Wave three: structural intelligence. The wave in which the AI is not running specific tasks but is the operating model itself. Every deal a firm has done becomes a queryable precedent. The firm stops being a collection of analysts who happen to share a logo, and becomes a single learning system that improves with every deal it touches.

First: AI in this market is not a feature. It is a layer. It runs underneath the workflows that already matter (onboarding, eligibility, covenants, monitoring, reporting) and it makes them faster, denser, and more honest. If your AI strategy starts with a chatbot, you are starting in the wrong place.

Second: the buy-side problem is fundamentally a data problem. There is no useful AI on top of bad data. Most of the work, and most of the differentiation, is in the layer that turns messy, multi-source, multi-format inputs into a single, validated, traceable source of truth.

Third: the future of credit work is not human-versus-machine. It is human-with- machine, with the human spending their time on the parts of the job that actually require judgement, and the machine handling the parts that never should have been done by humans in the first place. The firms that get this balance right will produce more deals, better decisions, and lower operational risk than the firms that do not, and the gap will widen each year.

Fourth: governance is not a brake on AI. It is the steering wheel. The firms that build proper AI governance: model inventories, change-management, traceability, human override at every consequential decision will move faster, not slower, than the firms that try to skip it, because they will be the ones that supervisors, allocators, and auditors trust

Fifth, and finally: the window in which AI is a competitive advantage is shorter than people think. Within two years, every serious firm in this market will have some version of this stack. The question is who builds the muscle now, while the playing field is still uneven, and who waits until parity becomes the price of admission.

Every credit business is, when you strip away the strategy and the marketing, a data business. The loan tape is data. The credit agreement is data. The covenant is data. The borrower’s financials are data. What separates a manager who can grow from one who cannot is whether their data architecture can absorb new asset classes, new originators, new structures, and new reporting requirements without each of those additions costing six months of engineering.

The brutal reality of asset-based finance and private credit infrastructure today is that almost no one has solved this. When you look inside even very large managers, you find loan tapes arriving in dozens of bespoke formats (Excel files with merged headers, fixed-width text, password-protected PDFs, SFTP drops with no schema attached), fields named differently across originators and dates encoded in incompatible standards.

Most credit platforms are tools stitched together. The better approach is one opinionated system with four layers.

• Ingest and normalise: Every tape arrives differently. The right layer maps fields deterministically, uses AI for harder semantic matches, and validates instantly. Every record carries full provenance, making everything auditable.
• Canonical model: One representation of what a loan, borrower, and cash flow are, regardless of asset class. Built around invariants, so auto loans, CLOs, and solar leases are specialisations of the same model, not separate silos.
• Rules engine: Eligibility tests, waterfalls, covenant calculations must run deterministically. AI belongs at the edges, digitising agreements in, explaining results out. If a covenant test depends on which AI model ran it, you don’t have a covenant test; you have a liability.
• Analytics and reporting: Where AI drafts memos, commentary, and dashboards. But a beautifully written IC memo built on an unverified tape is just a faster way to make a worse decision.

One principle cuts across all four: observability. The system must be loud when something changes: unexpected rows, broken mappings, shifting ratios. Silent failures are how data incidents become regulatory incidents.

If your infrastructure is working, you should be able to do all of this without writing a ticket:

• Onboard a new originator in 1–2 days: mapping their tape is a configuration exercise, not an engineering project
• Run eligibility tests against last night’s data: not last month’s close
• Add a new covenant and back-test it overnight: because it’s a rule, not a script
• Answer complex exposure queries in seconds: because the canonical model already knows what an obligor, leverage, and DSCR are

If you cannot do these things today, you do not have an AI problem. You have an infrastructure problem. AI on top of broken infrastructure is faster broken infrastructure.

The techniques that dominated the past decade, gradient-boosted trees, regularised regressions, hand-engineered features, still work. But three developments are changing what “good” looks like: the maturation of LLMs for unstructured-data extraction, the arrival of tabular foundation models, and the practical viability of multi-agent architectures. Ignoring any of them has already started to be a competitive mistake.

TabPFN and its successors are transformers pre-trained on millions of synthetic tabular datasets. Rather than training from scratch on each new problem, they perform classification or regression in-context, the way a language model is shown a prompt. On the public TabArena leaderboard, TabPFN 2.6 wins roughly 83% of head-to-head comparisons against tuned XGBoost. For private credit, this matters for two reasons: these models excel on small datasets (exactly what most ABF portfolios look like), and they produce well-calibrated probability outputs without post-hoc calibration work.
The limits are real. The current generation works well up to roughly fifty thousand rows. And a 7M-parameter transformer pre-trained on synthetic data is a harder defence in front of a model-risk committee than a decade-proven XGBoost. We are doing the validation work. But anyone claiming it is straightforward has not deployed one in a regulated context.

The pattern that works is consistent: the language model proposes; the deterministic system disposes. LLMs are excellent at first-pass extraction and flagging. They are dangerous as systems of record. Putting an LLM on the path of an authoritative number is an unforced error. Putting one on the path of an interpretation a human will then verify is one of the highest-leverage things you can do in this market.

Thin-data segments: Revenue-based finance, BNPL, legal claims, public administration receivables – asset classes with short histories where traditional ML struggles most. Tabular foundation models are purpose-built for this regime.

Early-warning indicators: Leading signals that fire before the financial statements show deterioration: covenant-headroom trajectories, customer-concentration drift, working-capital-cycle elongation, management-commentary tone shifts.

Causal modelling: Predictive models tell you what tends to happen. Causal foundation models let you ask what would happen, “what is the effect on my book if rates rise 200bps?”, without months of econometric setup.

Institutional knowledge retrieval: Every credit firm has years of decisions, memos, and post-mortems that, if queryable, would be its most valuable competitive asset. The technology to embed and retrieve this corpus responsibly is now mature. The alternative, firm history scattered across laptops and email, is not actually a strategy.

If you ask a buyer in private credit today what they need, you will get three different answers depending on who you ask. The CTO will say “we need better infrastructure.” The Head of Risk will say “we need better models.” The Head of Investments will say “we need more data.” Each of them is right. None of them is right alone. The reason the market has under-served them for so long is that the vendors have organised themselves around these three answers rather than around the underlying problem.

There are software companies that build excellent workflow tools but rely on the customer to bring their own data. There are data companies that aggregate excellent data but expect the customer to bring their own software. There are AI companies that build excellent models but assume the customer has already solved both data and software. In each case, the customer is left as the systems integrator, doing the work that should have been done by the product.

The thesis of the Cardo AI product is that the three are not separable. Software without data is empty; data without software is unusable; AI without both is a parlour trick. The product the market actually needs is a single platform in which the three are designed together – where the workflows speed up execution and reduce risk, the data are properly managed the workflows, and the intelligence layer is built on top of both.

It is easy to claim integration in a marketing document. It is harder to design for it. Let me describe what it actually means in the day-to-day experience of a user.
When a credit analyst uploads a new loan tape, the platform does not just store it. The ingestion layer normalises the file against an asset class specific canonical model, the validation layer flags any rows that fail integrity checks, the rules engine immediately re-runs eligibility tests on the updated portfolio, the covenant tracker updates the headroom calculations, the analytics layer refreshes the dashboards, and the AI layer surfaces any borrowers whose metrics have drifted in a way that warrants attention. None of this requires the analyst to do anything beyond the upload. The system understands what the file means and what to do with it, because the workflow, the data model, and the intelligence are the same product.

1) Determinism for numbers, AI for everything else
Numbers – borrowing-base calculations, eligibility tests, covenant headroom – are deterministic. They have to produce the same answer every time. AI does not belong on this path; rules engines do. Everything else, drafting commentary, extracting covenants, generating IC memos, is where AI dramatically outperforms the human-with-spreadsheet status quo.

2) Traceability is a feature, not an afterthought
Every output must be traceable to its inputs. The IC memo links to the loan tape. The covenant test links to the credit agreement clause. The risk score links to the model version. This cannot be added later. It is also what makes the product work with regulators — firms that built it in have a much shorter path to EU AI Act compliance than firms retrofitting it.

3) Configurability over customisation
Every credit firm thinks its workflows are unique. The covenants and structures are, but what a covenant is, what a borrowing base is, are the same across firms. Express the invariants as platform, the specifics as configuration. One-offs become technical debt. The discipline of asking “is this a new invariant or a configuration of one we have?” is what allows new clients to onboard in weeks, not months.

4) Engineering and domain expertise as a single team
You cannot build for this market with engineers alone, nor with structured finance specialists alone. The engineer who hasn’t lived through a quarter-end close doesn’t know what an asset manager needs. The specialist who can’t write code can’t solve complex problems. The product is what emerges when that gap has already been bridged inside the team, not left for the customer to bridge.

The most common CTO pushback: why can’t we build this ourselves? You can. But consider what you are actually building: a canonical data model you will discover is missing five fields in year three; a waterfall model that works for the first structure and breaks on the second; an eligibility engine that has never seen the edge cases that arrive in year four. The time cost of those discoveries is not a technology cost. It is an opportunity cost measured in deals not done and deterioration not caught.
When you use Cardo, you are buying eight years of decisions about what a canonical data model should contain, what an eligibility test should handle, and what happens when an originator sends something you have never seen before – because we have seen it before, across hundreds of deals and dozens of asset classes.

When most people talk about AI in credit, they imagine a single intelligent assistant, a chatbot, perhaps, or a copilot embedded in a spread-sheet, that helps with one task at a time. That model has its uses. It is not what we believe the future of work in this market looks like, and it is not what the most ambitious firms are deploying.

The challenge with a single-agent model is not just that private credit and ABF involve many tasks; it is that the work is inherently iterative, stateful, and interdependent. A deal is not processed in a straight line from ingestion to output. Data is extracted, normalized, checked, challenged, re-run, compared to policy and precedent, and often revised as new issues emerge. Reading a data room, cleaning a loan tape, computing eligibility, generating sensitivities, drafting an IC memo, and benchmarking against comparable transactions are not simply consecutive steps; they are connected parts of a workflow that loops back on itself until the analysis is robust enough for decision-making. Asking one general-purpose assistant to manage all of this is asking it not only to do many different jobs, but to coordinate them, remember them, and refine them in context.

That is why the architecture that works is multi-agent: a coordinating layer that breaks the problem into components, routes each to the right specialist, and manages state across the process. In ABF, those capabilities are what turn AI from an assistant into an operating model.

The coordinator reads the request, decomposes it into sub-tasks, and routes them. It is the difference between a system that reacts to commands and one that executes plans.

Specialised agents each do one job well – ingestion, logic parsing, spreading, precedent search, drafting. The orchestrator coordinates them; the human reviews the result.

Shared state and memory logs every step, every input, every output – making the system fully auditable and every decision reproducible.

Human checkpoints sit at the points of consequential judgement: before a credit decision is recommended, before a covenant is committed to the rules engine, before an IC memo is finalised. Between those points, the system runs autonomously.

In production today, our multi-agent stack covers the workflows that consume the most analyst hours in a typical credit firm. None of them replaces a human; all of them change what the human does.
Origination & screening: monitors deal flow, scores fit against underwriting criteria, and drafts a preliminary brief before an analyst opens the email.
Data room & loan tape: classifies documents, flags missing or draft materials, normalises the loan tape, and surfaces validation failures.
Covenant extraction & monitoring: reads credit agreements, extracts covenants, proposes mappings for human sign-off, then monitors compliance continuously.
Borrower monitoring: processes each financial package on arrival, compares it to prior periods and the original underwriting model, and flags material drift.
IC memo & sensitivity: drafts the financial section of the credit memo, runs standard stress scenarios, and links every assertion to its source.
LP & investor reporting: assembles the reporting pack, drafts commentary, tailors views by mandate, and presents for sign-off rather than manual assembly.

1. Error compounding: Small errors propagate and amplify across agents. The fix is deterministic validation at every handoff, the next agent doesn’t start until the previous agent’s output passes integrity checks.
2. Observability: A multi-agent system is a small distributed system. We invest heavily in per-agent logging, tracing, and replay tooling, agents whose behaviour can’t be reconstructed after the fact are incompatible with regulated deployment.
3. Governance: Every consequential action needs a designated human owner, a full log of the inputs that produced it, and the ability to reverse it. “The agent is accountable” is not an answer.
4. Autonomy: When the system works well, the temptation is to remove human checkpoints. Resist it. The checkpoints are not friction, they are the accountability mechanism. Firms that removed them had to walk back deployments because no one could explain what the system did or why.

The near-term frontier is cross-firm interoperability, agents coordinating with agents at counterparty firms, compressing deal cycles, and institutional learning loops built on the firm’s own memos, declines, and outcomes. That second capability, over five years, becomes a moat no competitor can replicate.
The more consequential shift is from agents that assist humans to agents that work for them. Users won’t watch agents do work; they’ll delegate outcomes. The agent surfaces what matters and asks for help when it needs it. I believe this is how credit firms will operate within five years.

Better data without better models is just more spreadsheet rows. Better models without better data are faster ways to reach the same conclusions everyone else already has. The edge is at the intersection, applying modern modelling techniques to data sources the rest of the market is still treating as marketing material.

The data that moves the needle isn’t satellite imagery or social sentiment. It’s cash flows from bank accounts and payment processors, already improving predictive accuracy by over 40% in some consumer credit applications. It’s the operational data sitting inside origination, servicing, and collections platforms that has always existed but was too messy to use: how long underwriting took, how often borrowers called servicing, the seasonality of payments. It’s the asset-class-specific signals that never appear on the loan tape, obligor payment behaviour for receivables, platform revenue data for RBF, production readings for solar lease pools. And it’s the firm’s own memos, declines, and post-mortems, sitting unread in documents for years, now indexable through language models. That last one is the data source no competitor can replicate.

Traditional credit modelling resets from scratch with every new data source – collecting, cleaning, engineering features, training, validating, deploying. The cycle takes months. Tabular foundation models collapse the training end of that cycle dramatically. Fitting and evaluating a candidate model takes minutes instead of days. Asset classes too small or too young for conventional ML, BNPL portfolios, revenue-based finance, public administration receivables, become viable. And when experimentation is cheap, the calculus of which data sources are worth acquiring changes entirely. Sources that would never have justified a multi-month modelling project are now worth a one-week pilot. That changes data strategy, which changes what deals a firm can credibly underwrite.

These models work best on small and medium datasets, they are not a replacement for well-tuned models where you have millions of rows, and we use them as a fast experimentation layer and in ensembles. Every data source needs proper consent and documentation; cash-flow data without authorisation is a regulatory liability, not an asset, and this is non-negotiable under GDPR and the EU AI Act. The asset classes where alternative data adds the most value are often the youngest, with no track record through a downturn, models that look strong in benign conditions can be confidently wrong when the cycle turns, and anyone claiming otherwise is overselling. Finally, richer data does not mean fewer fair-lending obligations. Cash-flow features and behavioural signals can correlate with protected characteristics in ways that aren’t obvious: payment timing can carry geographic signal, transaction patterns can carry demographic signal. We run bias testing on every model we ship and treat the results as hard constraints, not a compliance checkbox.