Credit is civilization's source code — and agents are about to rewrite it

The oldest written artifacts our species possesses are ledgers. Clay tablets from Uruk and Ur, five thousand years old, recording who borrowed barley, how much, and when it was due — older than any poem or prayer we have ever dug up. The strongest evidence says writing itself was invented for accounting. Debt came first. Literature is a side effect of lending.

Eight hundred years later, the Code of Hammurabi — one of humanity’s first legal systems — spends an astonishing share of its 282 laws on credit: caps on grain interest, caps on silver interest, debt forgiveness in flood years. Lending was the first thing we ever regulated at civilizational scale.

I keep coming back to this because it reframes what our industry actually is. Strip the jargon away and a loan is a promise about the future, priced in the present, between strangers. Lending is the technology that lets trust scale past the village — past the hundred-odd people you can personally vouch for — to millions of strangers cooperating across decades. It is, quite literally, civilization’s source code: the trust protocol everything else compiles against.

And for five thousand years, that protocol has had one hard constraint: somebody has to read the tablets. A human being has to look at the evidence, judge the promise, and write down a decision. Every cost, every delay, every bias and every exclusion in the modern credit system flows from that single bottleneck.

That constraint just broke. I want to walk you through what I believe happens next — carefully, with numbers, because the numbers are staggering.

The load-bearing layer nobody looks at

Start with scale. The Institute of International Finance measured global debt at a record $348 trillion at the end of 2025 — roughly three times world GDP. Governments carry about $107 trillion of it, non-financial companies about $101 trillion, households about $65 trillion. US households alone hold $18.8 trillion of debt, $13.2 trillion of it mortgages. Global consumer lending is a $27 trillion machine. Private credit crossed $2 trillion in 2024 and Preqin projects $4.5 trillion by 2030.

These numbers are so large they stop meaning anything, so make it personal instead. Almost everything you touched today was borrowed into existence. The home you woke up in — mortgage. The building you work in — construction loan. The road you drove — municipal bonds. The phone in your pocket — corporate debt financing a supply chain across nine countries, plus receivables financing at every hop. The coffee shop on the corner — a working-capital line. Your salary, in more companies than you’d like to know, clears payroll off a revolving credit facility. The medicine in your cabinet was a decade of venture debt and R&D credit before it was a pill.

All of it is debt as infrastructure — the invisible layer that lets the future pay for the present. Credit comes first and prosperity follows; economists have known this since long before fintech. The places where collateral can’t be pledged and credit can’t flow are the places that stay poor, no matter how talented their people are.

Here’s the thought experiment I use to make the point land. Suppose credit froze tonight — every new loan, every rollover, every facility renewal halted, with money itself still flowing freely. How long before civilization notices? We actually have the answer, because in September 2008 the commercial paper market — the mundane, boring market where companies borrow for 30 to 90 days to make payroll — seized for a matter of days, and household-name corporations came within a week of missing salaries. Days. The modern world runs about ninety days deep on trust, and it refinances that trust continuously, forever. Lending is the economy’s heartbeat.

If lending is civilization’s source code, then every flaw in lending is a flaw the whole civilization inherits. And the flaws are not small.

The machine is magnificent — and broken

Because when you actually inspect this load-bearing layer, what you find is a masterpiece of institutional engineering running on a paleolithic bottleneck: human reading.

Look at the friction first. The Mortgage Bankers Association puts the average cost of originating a single US mortgage at $11,076 — it peaked above $14,000 in 2023. Nearly all of that is the price of processing: document chasing, re-keying, verification, stipulations, committee. ICE’s data says the average mortgage still takes about 38 days to close — and that’s a modern record; a few years ago it was 49. McKinsey’s benchmark for SME lending at banks is grimmer: three to five weeks to a decision, nearly three months to cash. Three months, for a business that needed the working capital this month.

Now look at who the friction lands on, because this is the part that should make everyone uncomfortable. Underwriting has a mostly fixed cost — it costs a bank nearly as much to read the file for a $40,000 loan as for a $4 million one. So the system does the economically rational, humanly corrosive thing: it ignores small borrowers. The IFC’s 2025 estimate of the credit gap for formal micro, small and medium enterprises is $5.7 trillion — $8 trillion if you count informal firms. Seventy percent of MSMEs in emerging markets can’t get adequate financing, and these are the firms that make up over 90% of all businesses and roughly half of world GDP. The Asian Development Bank measures a $2.5 trillion trade finance gap, with SMEs rejected at a 41% rate. The World Bank’s 2025 Findex still counts 1.3 billion adults with no financial account at all. Even in the United States, the CFPB’s latest estimates put roughly 32 million adults — one in eight — outside conventional credit scoring: seven million with no file at all, twenty-five million with files too thin to score.

Read that list again: it is a list of unread people — people and businesses whose evidence of creditworthiness exists (sales ledgers, harvest cycles, mobile-money histories, order books, rent paid on time for a decade) but has never been economically worth a human’s reading time. The cruelty of the current system is precisely that it is rational: when diligence costs thousands of dollars per file, the poor are unprofitable to understand.

And when files do get read, human judgment carries human noise. The landmark Berkeley study of US mortgage data (Bartlett, Morse, Stanton and Wallace) found face-to-face lenders charging risk-equivalent Latino and Black borrowers about 7.9 basis points more on purchase mortgages — roughly $450 million a year in excess interest — and estimated over a million minority applications rejected between 2009 and 2015 that identical white applicants would have had approved. The same study’s algorithmic lenders showed no detectable approval discrimination and about a third less pricing disparity. Two loan officers reading the same file on different afternoons produce different answers. We have simply learned to call that variance “judgment.”

Every one of these behaviors is the rational response to a single constraint: underwriting is a reading profession, and reading has never scaled. Until now.

What an agent actually is — and why it’s not “AI in banking, again”

Banks have had AI for decades: scorecards since the 1950s, neural networks in fraud since the 1990s. So what’s different?

Everything that was called AI in lending until now was a number about the past. A credit score compresses your history into three digits and hands it to a human who still has to do the actual work — read the bank statements, tie the tax return to the P&L, catch the related-party transaction in note 14, check the collateral registry, write the memo, defend it in committee.

An agent is a system that does that work itself: it reads every page of every document with citations back to the source line; it verifies claims against registries, bureaus and bank data; it reasons over your credit policy the way an analyst reasons — but over the entire file, not the sample a tired human has time for; it writes down its reasons in language a regulator can audit; and it acts within governed bounds — assembling the file, chasing the missing document, drafting the memo, flagging the exception to a human whose accountability is real.

The distinction matters because lending’s bottleneck has always been processing evidence, and that is exactly where agents land. McKinsey estimates generative AI can add $200–340 billion of annual value to banking; BCG puts the retail-banking prize at $370 billion of additional annual profit by 2030, with agents’ share of that value nearly doubling by 2028. In corporate credit specifically, McKinsey’s early agentic deployments report 20–60% productivity gains and roughly 30% faster credit turnaround — and credit-memo agents cutting manual workloads by a third to a half. And yet the adoption paradox is stark: EY finds 99% of banking executives familiar with agentic AI but only 31% actually moving to implementation; in Fenergo’s survey, 93% of financial institutions plan agents within two years while just 6% run them today. Translation: essentially everyone can see it coming; almost no one has rewired the lending process around it yet. Which means the interesting part — the part below — is still ahead of us.

The five inversions

Here is what I believe agentic lending actually does to the industry. “Faster loans” is the boring, first-order prediction — the equivalent of guessing the internet would speed up mail. The real story is five structural inversions.

1. The cost of diligence collapses — and the bottom of the pyramid becomes bankable

When an agent underwrites, the marginal cost of reading a file falls from thousands of dollars toward the cost of computation — cents, and falling. Follow that to its conclusion: the $800 loan can receive the same quality of diligence as the $8 million loan. The fixed-cost logic that made small borrowers unprofitable to understand simply dissolves.

That $5.7 trillion MSME gap? I’d argue most of it is a reading-cost gap wearing a risk gap’s clothes. The seamstress with three years of mobile-money records, the farmer with a decade of harvest receipts, the two-person machine shop with a full order book: their evidence was always there. It was just never worth $11,000 of human process to examine. The 41% SME rejection rate in trade finance is substantially a cost-of-diligence rejection rate. When diligence is nearly free, rejection has to earn its keep on risk alone. Entire continents of creditworthy humans become, for the first time in history, legible.

2. Credit stops being an event and becomes a condition

Today, credit is episodic. You apply; you are photographed financially; the photograph is filed; the file goes stale the day it closes. Your bank meets your business three or four times a decade and extrapolates wildly in between. Covenants are checked quarterly. Default arrives as a surprise that was visible for months in data nobody was reading.

An agent never stops reading. The same system that underwrote you can watch the actual cash flows, the actual receivables aging, the actual order book — continuously, with consent, under governance. Underwriting and monitoring merge into a single ongoing act of attention.

What follows is profound: credit limits that breathe with your actual season; pricing that drifts down as your real risk drifts down, without you begging for a review; distress detected as a pattern in week two instead of a missed payment in month six — met with restructuring while restructuring can still work. Default rates fall because the relationship finally, literally pays attention.

3. The loan application dies

The application is an artifact of reading scarcity — a form invented so scarce human attention could consume you in summary. When attention is abundant, the form is absurd.

Within a decade, I don’t think most credit will begin with anyone “applying.” Your agent — your business’s fiduciary software, holding your permissioned data — will negotiate with lenders’ agents continuously and silently. It will notice you’re about to take a purchase order that strains working capital, canvass forty lenders’ agents overnight, exchange evidence under consent, and present you one decision in the morning: “Here are three offers; this is the best one; here’s the reasoning; sign here.”

Notice what this does to market structure. Switching costs collapse — your agent re-shops your entire credit stack continuously, so incumbency stops being a moat and price has to be. Price discovery becomes per-borrower and near-real-time. The subsidy that inattentive borrowers currently pay to attentive ones — the loyalty tax that funds much of consumer finance — evaporates. Some lending revenue pools shrink; I won’t pretend otherwise. The ones that remain will be earned on underwriting skill rather than customer inertia. That is a better industry.

4. The past stops masquerading as the future

Credit history and collateral are proxies — inventions of a world where the actual truth about risk, the forward cash flows and real contracts and real behavior, was too expensive to read. We deny the future capital because the past was too expensive to verify.

Agents invert this. A system that can read and verify everything can price the loom, not the land under it; the order book, not the credit file; the demonstrated discipline of three years of wallet history, not the absence of a bureau record. The evidence is no longer hypothetical. Upstart’s 2025 access-to-credit analysis reports its AI model approving 41% more applicants at 33% lower average APRs than a traditional credit-score model — with the largest gains for Black and Hispanic borrowers. FinRegLab’s independent testing found that machine-learning models using cash-flow data were the most predictive across every demographic subgroup, lifting approvals without lifting defaults. An NBER study of “invisible primes” found that borrowers a traditional model would reject at a 70% higher rate turn out, under alternative-data underwriting, to be excellent credits. And all of that is with today’s models, on yesterday’s data rails. Extend the curve. The 32 million credit-invisible Americans, the 1.3 billion unbanked — these populations were only ever unread.

5. Fairness becomes an architectural property, not an aspiration

This is the inversion I care about most, and the one people find most counterintuitive — because “algorithmic bias” has become shorthand for a dystopia.

But hold the current system honestly in view: judgment that varies by afternoon, by accent, by zip code; disparities measured in basis points that compound into generational wealth gaps; decisions whose true reasons are unrecoverable even in litigation, because they lived in a loan officer’s head. You cannot audit a hunch. You can diff a policy.

An agentic decision, done right, ships with its own evidence trail: every factor, every weight, every citation back to a source document, every policy rule that fired. Bias stops being an accusation to litigate and becomes a defect to detect — a regression you can test for before deployment, measure in production, and patch in the next release. Underwriting fairness becomes an engineering discipline with the properties engineering disciplines have: reproducibility, monitoring, rollback. No human credit committee, however well-intentioned, has ever offered that.

What keeps me up at night

I’m a builder of these systems, so read this section as the confession it is. Four things worry me.

Monoculture. If every lender’s agent descends from the same two or three foundation models, credit errors correlate. Humanity has run diversified, de-correlated credit judgment for five millennia — thousands of banks, millions of loan officers, all wrong in different ways. That diversity was a stabilizer we never appreciated. A monoculture credit system could misprice an entire asset class simultaneously and at machine speed. 2008 took eighteen months to propagate; an agentic mispricing could take eighteen hours. The BIS has been signaling exactly this in its recent AI chapters, and they’re right to.

The fraud arms race. Synthetic identities — people assembled from fragments of real data — already had US lenders holding a record $3.3 billion in exposure by the end of 2024, against human verifiers. Agents will verify better — and generative tools will fabricate better. Documents stop being evidence unless they’re cryptographically anchored to their source. The entire evidentiary chain of lending has to be rebuilt for a world where anything can be forged and everything can be checked. That work has barely begun.

Bias at scale. Inversion five has to be earned. An ungoverned agent trained naively on decades of biased lending history will industrialize that bias, with a confidence score attached. The difference between the utopia and the dystopia here comes down to governance alone: bias testing as a release gate, decision logs as a legal artifact, humans accountable for policy. Fairness-by-architecture is available — and it has to be deliberately chosen, engineered and enforced.

Accountability. You cannot fine an agent, jail it, or make it feel shame — the load-bearing deterrents of financial regulation for centuries. So accountability has to be engineered: every agent action attributable, every decision explainable, every autonomy boundary explicit, a named human owning every policy. And regulation itself must invert — from reviewing individual decisions after the fact to certifying decision systems before deployment, the way we certify aircraft rather than inspecting each flight.

This is, not coincidentally, a description of how we build at Botminds. Blast-door isolation between agent and system of record. Citations on every extracted fact. Evaluation harnesses before autonomy. Audit trails as first-class product, not compliance afterthought. The lesson of every powerful technology is that the guardrails are the revolution. A fast, opaque credit machine would be worse than the slow one it replaced.

The world on the other side

Zoom all the way back out, to where we started — clay tablets, barley, the first trust protocol.

Talent is evenly distributed across the human species. Capital access never has been. That mismatch — between where ability lives and where credit reaches — may be the single largest reservoir of wasted human potential on Earth. Every unread seamstress, every unfunded machine shop, every farmer paying informal-market rates for working capital, every student whose thin file priced her out of an education: that is the compounding, invisible cost of underwriting at human reading speed.

Now imagine the other side. Diligence that costs cents. Credit that watches over a business like a partner instead of photographing it like an examiner. Capital that finds a good order book in Nairobi as easily as a good credit score in Nashville. Trouble met at week two with restructuring instead of month six with repossession. Every decision carrying its reasons, inspectable by the borrower it affects. The interest-rate spread between the world’s connected and unconnected borrowers — mostly a reading premium, all along — grinding toward zero.

I don’t know exactly how fast this arrives. Adoption in banking is never a straight line, and the failure modes above are real. But the direction feels as clear to me as anything in my career: the five-thousand-year-old constraint at the heart of lending — somebody has to read the tablets — is gone. Every prior time the cost of trust fell, civilization got bigger: writing gave us cities, double-entry bookkeeping gave us the Renaissance firm, the joint-stock company gave us the industrial age. The pattern is not subtle.

Lending built the modern world quietly, one read file at a time. Agentic lending will build the next one — and it will read everyone’s file.

Somewhere out there right now is a founder, a farmer, a lab, a town — creditworthy in every way that matters and invisible in every way that’s measured. The most consequential thing our industry will do this decade is finally read them.

That’s the future we’re building toward at Botminds — agentic underwriting with citations, governance and audit at the core, live today in credit teams that have stopped accepting the ninety-second-per-page world. If you run lending and any paragraph above felt personal, I’d genuinely like to talk.

— Ansari