The landscape · Finance
Finance: what AI is really doing
Brief me in 2 minutes
AI in Finance.
AI in finance right now means agents running the boring middle of the business: contract review, bookkeeping, loan-document chasing, compliance evidence-gathering, not a customer-facing robot advisor. The real money is in production at places like JPMorgan and Rogo's bank clients, while a lot of the 'agentic payments' and full loan-underwriting-by-AI talk is still pilot-stage plumbing between big payment networks and single-digit-employee startups. None of it removes the human who signs the loan, the audit, or the trade, it just changes how many hours that human spends getting there.
The game: we show you 6 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 6
Banks move agentic AI from chatbots to running back-office and dealmaking work
Real now
JPMorgan is quantifying software it has already deployed at scale, and Rogo has named, paying bank clients in production, not pilot-stage promises.
How it works. JPMorgan now runs more than 400 AI and machine-learning use cases it values at $1.5-2B a year, and its COIN contract-review platform, which used to save 360,000 lawyer hours annually, now processes 1.2 million hours of contract review a year. On Wall Street, agents built by startups like Rogo run multi-step dealmaking tasks, deal screening, CIM generation, buyer outreach, data room diligence, inside major banks without a human doing each step by hand.
Overstated: The $1.5-2B and 400-use-case figure is JPMorgan's own count of its own software, not an audited number, and 'without human intervention' still means a banker reviews the output before anything reaches a client.
Shift 2 of 6
Accounting goes agent-native, and it starts with the entry-level rung
Early
The funding and the workflows are real and running, but both companies still keep a licensed human in the loop before anything ships, so it's automating hours, not the job.
How it works. Startups like Basis and Integral run agent-based workflows, not copilots, that do the mechanical work of tax prep, reconciliation, workpapers and payroll, then hand off to a licensed professional for sign-off. Integral's agents now do end-to-end bookkeeping for over half its clients and cut monthly turnaround from weeks to hours; Basis just raised at a $1.15B valuation to do the same across tax, audit and advisory.
Overstated: "End-to-end" still means a licensed professional signs off before anything goes to a client or a regulator, so now this replaces hours, not headcount.
Shift 3 of 6
Loan underwriting moves from AI-assisted software to agents that do the paperwork themselves
Early
It is live and originating real loans, but only in two states and at a size too small to say the approach holds up at scale.
How it works. Instead of surfacing a risk score for a human loan officer to act on, new lenders let agents run the whole pipeline. Chestnut's agents shop mortgage rates across 100+ lenders, chase and verify borrower documents, and originate the loan end to end, targeting the $12,000-plus per loan that manual origination costs now.
Overstated: "Automates 99% of mortgage operations" is the company's own claim, live in only two licensed states, from a company that has raised in the hundreds of thousands, not proof the model holds at national-lender scale.
Shift 4 of 6
Payment rails get rebuilt for agents paying agents, in fractions of a cent
Mostly hype
The rails exist and have named partners, but none of the sources disclose any actual agent-to-agent payment volume moving through them.
How it works. Mastercard's Agent Pay for Machines and Ant International's Agentic Mobile Protocol let AI agents pay each other automatically at machine speed, settling across cards, bank accounts and stablecoins, with agent permissions recorded on-chain. Startups like Natural are building the orchestration layer in between, the wallet and permission system a company's agents use to actually move money.
Overstated: "Always-on machine payments" is still mostly partner-announcement plumbing: Mastercard names 30-plus partners but discloses no transaction volume, and Ant's protocol is a few months old with ten wallet partners, not a functioning market yet.
Shift 5 of 6
Fraud and compliance go agentic, and the bank's own agents become a new attack surface
Early
These are funded, live products with real customers, but the harder claim, that an agent can be trusted to act instead of just flag, is still unproven at any of them.
How it works. Sardine's agentic risk platform profiles devices and scores fraud risk in real time instead of running static rules; deepidv adds deepfake, synthetic-media and document-forgery detection with sub-150ms risk scoring; Comp AI's agents draft policies, gather audit evidence and run vendor security assessments continuously instead of once a year. Because banks now run their own AI agents in production, Arcade.dev exists purely to authorize and audit what those agents are allowed to do.
Overstated: "Agentic" is doing a lot of work in these announcements: most of what's shipped is agents gathering evidence and flagging anomalies for a human reviewer, not agents making the compliance decision itself.
Shift 6 of 6
Wealth management AI goes after the advisor's inbox, not the client chatbot
Real now
Hazel shipped, has a disclosed adoption number, and does checkable admin work rather than promising to replace a human advisor.
How it works. Altruist's Hazel platform drafts client emails in an advisor's own tone, auto-creates and syncs CRM tasks after meetings, transcribes and summarizes meetings in real time, and answers advisor questions against the firm's own records through a feature called "Ask Hazel," rather than generic financial advice. It targets an advisor's admin time, not the client relationship itself.
Overstated: Flashy "AI financial advisor" demos get the headlines, but the money is quietly going into back-office paperwork automation, a duller story than the industry likes to tell about itself.
The scary part
What happens to jobs in Finance
What AI is already taking:
- First-pass review of commercial loan agreement clauses AI Agent Corps ↗
- End-to-end monthly bookkeeping close for small business clients Tech.eu ↗
- Shopping mortgage rates and originating the loan paperwork Fondo ↗
- Gathering audit evidence and running vendor security assessments for compliance SiliconANGLE ↗
What stays human:
Signing off on anything that touches a client's money, a loan decision, or a regulatory filing stays human: Basis and Integral both route agent output through a licensed professional before it ships, and Rogo's own client base is analysts directing Felix, not analysts replaced by it. Judgment on the file that doesn't fit the pattern, a borrower with a strange income history, a transaction that trips a rule for the wrong reason, still needs someone who understands why the rule exists.
Get better · 15 minutes
The ability to catch what an agent got wrong before it ships, which means knowing the underlying accounting, credit or compliance logic well enough to audit a shortcut instead of just accepting the answer.
Try this this week:
Pick one task you already do by hand this week, reconciling a batch of invoices or drafting a client update, and watch the public product walkthrough for Payouts.com's Digital Employee or Altruist's Hazel doing that same task on their own sites, then write down the one step where you'd still check its work before sending it out.
Your move
The open lane is authorization and audit trails for agents already in production, not another agent itself, that's why Arcade.dev exists and banks are the ones paying for it. Bookkeeping and mortgage-origination copilots are already crowded, Basis, Integral and Chestnut are all chasing the same workflow across different YC batches; the harder, less-crowded problem is what happens when one of these agents is wrong and a regulator asks who signed off.
Watch for a named, paying enterprise customer running a real workflow, not a pilot MOU, that's the gap between Rogo (five banks, disclosed production use) and half the 'agentic lending' pitches that cite a seed round and a demo video. Red flag: any pitch that claims full automation of a regulated decision, credit, audit sign-off, AML clearance, with no human step named anywhere in the flow.
This quarter, pick one back-office workflow you already do by hand, reconciling a batch of invoices or spreading financials for one loan file, and run it through a live agent product (Payouts.com's Digital Employee or Basis both publish demos) to find exactly where it hands back to a person, then check whether that handback point is where your actual risk lives.
The valuable skill is auditing a machine's shortcut, not producing the first draft yourself, which means knowing the underlying accounting, credit or compliance rule well enough to catch a plausible-sounding wrong answer. First step: take one AI-generated output your team already produces, a credit memo, a reconciliation, a compliance evidence pack, and write down every place you'd need to check by hand before you'd sign it.
What's actually changing
- 1
Banks move agentic AI from chatbots to running back-office and dealmaking work
How it works. JPMorgan now runs more than 400 AI and machine-learning use cases it values at $1.5-2B a year, and its COIN contract-review platform, which used to save 360,000 lawyer hours annually, now processes 1.2 million hours of contract review a year. On Wall Street, agents built by startups like Rogo run multi-step dealmaking tasks, deal screening, CIM generation, buyer outreach, data room diligence, inside major banks without a human doing each step by hand.
Why it matters. This is a headcount and cost story, not a chatbot novelty, and it moves revenue-generating work into pipelines banks now depend on and budget around.
What's overstatedThe $1.5-2B and 400-use-case figure is JPMorgan's own count of its own software, not an audited number, and 'without human intervention' still means a banker reviews the output before anything reaches a client.
JPMorganRogoRothschild & CoJefferiesLazard
3 receipts
- JPMorgan runs 400+ AI and machine-learning use cases valued at roughly $1.5-2B a year on an $18B 2026 tech budget.
JPMorgan runs 400+ AI and machine-learning use cases and pegs the measurable value at roughly $1.5–2B a year
ValueAdd VC, 27 June 2026 ↗ - JPMorgan's COIN contract-review platform, which once saved 360,000 lawyer hours a year, now processes 1.2 million hours of contract review annually.
Their COIN platform — Contract Intelligence — reviews commercial loan agreements, identifies key clauses, and flags issues that previously consumed 360,000 lawyer hours per year. Now it handles 1.2 million hours annually.
AI Agent Corps, 27 Mar 2026 ↗ - Rogo's Felix agent runs multi-step dealmaking tasks without human intervention and is in production at Rothschild & Co, Jefferies, Lazard, Moelis and Nomura.
Felix is capable of handling complex, multi-step financial tasks without human intervention, spanning deal screening, CIM generation, buyer outreach and data room diligence.
Fintech Global, 29 Apr 2026 ↗
- JPMorgan runs 400+ AI and machine-learning use cases valued at roughly $1.5-2B a year on an $18B 2026 tech budget.
- 2
Accounting goes agent-native, and it starts with the entry-level rung
How it works. Startups like Basis and Integral run agent-based workflows, not copilots, that do the mechanical work of tax prep, reconciliation, workpapers and payroll, then hand off to a licensed professional for sign-off. Integral's agents now do end-to-end bookkeeping for over half its clients and cut monthly turnaround from weeks to hours; Basis just raised at a $1.15B valuation to do the same across tax, audit and advisory.
Why it matters. This is the work junior accountants used to cut their teeth on, and firms are automating it while the profession is already short on new CPAs.
What's overstated"End-to-end" still means a licensed professional signs off before anything goes to a client or a regulator, so now this replaces hours, not headcount.
BasisIntegral
3 receipts
- Basis raised $100M at a $1.15B valuation to automate accounting workflows across tax, audit and advisory with an agent-based architecture.
Basis today revealed that it has raised $100 million in new funding on a $1.15 billion valuation
SiliconANGLE, 24 Feb 2026 ↗ - Basis's platform combines large language models with rules-based controls to run multistep accounting processes as agents, not suggestions.
agent-based architecture that executes multistep financial processes using large language models combined with rules-based controls and domain-specific accounting logic
SiliconANGLE, 24 Feb 2026 ↗ - Integral's agents now prepare bookkeeping end-to-end for more than half its clients, cutting monthly turnaround from weeks to hours.
Its agents now prepare bookkeeping end-to-end for more than half of its clients, reducing average monthly accounting turnaround from weeks to hours.
Tech.eu, 16 Sept 2026 ↗
- Basis raised $100M at a $1.15B valuation to automate accounting workflows across tax, audit and advisory with an agent-based architecture.
- 3
Loan underwriting moves from AI-assisted software to agents that do the paperwork themselves
How it works. Instead of surfacing a risk score for a human loan officer to act on, new lenders let agents run the whole pipeline. Chestnut's agents shop mortgage rates across 100+ lenders, chase and verify borrower documents, and originate the loan end to end, targeting the $12,000-plus per loan that manual origination costs now.
Why it matters. If a small startup can originate loans at a fraction of the manual cost, the fee structure the mortgage-broker industry runs on gets squeezed.
What's overstated"Automates 99% of mortgage operations" is the company's own claim, live in only two licensed states, from a company that has raised in the hundreds of thousands, not proof the model holds at national-lender scale.
Chestnut
2 receipts
- Chestnut says its AI agents can automate 99% of mortgage operations.
AI can automate 99% of mortgage operations
Fondo, 9 Feb 2026 ↗ - Manual mortgage origination costs lenders and borrowers more than $12,000 per loan, or over $61B a year industry-wide.
These manual processes cost lenders – and ultimately their borrowers – $12,000+ per loan ($61B+ per year).
Fondo, 9 Feb 2026 ↗
- Chestnut says its AI agents can automate 99% of mortgage operations.
- 4
Payment rails get rebuilt for agents paying agents, in fractions of a cent
How it works. Mastercard's Agent Pay for Machines and Ant International's Agentic Mobile Protocol let AI agents pay each other automatically at machine speed, settling across cards, bank accounts and stablecoins, with agent permissions recorded on-chain. Startups like Natural are building the orchestration layer in between, the wallet and permission system a company's agents use to actually move money.
Why it matters. This is infrastructure for commerce with no human clicking "buy," which changes who captures the transaction fee and who is liable when an agent overpays or gets scammed.
What's overstated"Always-on machine payments" is still mostly partner-announcement plumbing: Mastercard names 30-plus partners but discloses no transaction volume, and Ant's protocol is a few months old with ten wallet partners, not a functioning market yet.
MastercardAnt InternationalNatural
3 receipts
- Mastercard launched Agent Pay for Machines so AI agents and connected systems can pay each other automatically, including payments worth fractions of a cent.
Called Agent Pay for Machines, the new network is meant to make smaller money transfers easier when, for example, an AI agent wants to access data piecemeal from a website.
Fortune, 10 June 2026 ↗ - Ant International's Agentic Mobile Protocol went live in April 2026 with ten wallet partners and seven acquirers.
The Agentic Mobile Protocol went live in April 2026 and already has ten wallet partners and seven acquirers.
Fintech Global, 18 Sept 2026 ↗ - Natural raised funding to build an orchestration layer that lets a company's AI agents autonomously move and hold money.
an agent orchestration layer that enables AI agents to move and store funds
TechCrunch, 20 July 2026 ↗
- Mastercard launched Agent Pay for Machines so AI agents and connected systems can pay each other automatically, including payments worth fractions of a cent.
- 5
Fraud and compliance go agentic, and the bank's own agents become a new attack surface
How it works. Sardine's agentic risk platform profiles devices and scores fraud risk in real time instead of running static rules; deepidv adds deepfake, synthetic-media and document-forgery detection with sub-150ms risk scoring; Comp AI's agents draft policies, gather audit evidence and run vendor security assessments continuously instead of once a year. Because banks now run their own AI agents in production, Arcade.dev exists purely to authorize and audit what those agents are allowed to do.
Why it matters. Compliance teams that used to review a stack of evidence once a year now have software watching continuously, but the attack surface now includes the bank's own AI agents, not just outside fraudsters.
What's overstated"Agentic" is doing a lot of work in these announcements: most of what's shipped is agents gathering evidence and flagging anomalies for a human reviewer, not agents making the compliance decision itself.
SardinedeepidvComp AIArcade.dev
3 receipts
- Sardine's agentic risk platform helps banks detect and prevent fraud without adding friction for genuine customers.
an agentic risk solution designed to help banks detect and prevent fraud without creating unnecessary friction for genuine customers
Fintech Global, 21 May 2026 ↗ - deepidv's fraud suite does real-time transaction risk scoring in under 150 milliseconds alongside deepfake and forged-document detection.
real-time transaction risk scoring with sub-150-millisecond response times, and multi-agent AI compliance monitoring for AML, KYC, and CFT requirements
GlobeNewswire, 11 Mar 2026 ↗ - Arcade.dev gives production AI agents governed access to thousands of tools so enterprises can control what those agents are allowed to do.
it provides access to more than 8,000 MCP tools that those agents can use to automate work tasks
SiliconANGLE, 15 June 2026 ↗
- Sardine's agentic risk platform helps banks detect and prevent fraud without adding friction for genuine customers.
- 6
Wealth management AI goes after the advisor's inbox, not the client chatbot
How it works. Altruist's Hazel platform drafts client emails in an advisor's own tone, auto-creates and syncs CRM tasks after meetings, transcribes and summarizes meetings in real time, and answers advisor questions against the firm's own records through a feature called "Ask Hazel," rather than generic financial advice. It targets an advisor's admin time, not the client relationship itself.
Why it matters. The value shows up in an advisor's calendar, not in a slicker robo-advisor app, which is why an incumbent custodian is winning this fight faster than app-only startups.
What's overstatedFlashy "AI financial advisor" demos get the headlines, but the money is quietly going into back-office paperwork automation, a duller story than the industry likes to tell about itself.
Altruist
3 receipts
- Altruist launched Hazel, an AI platform for advisors that drafts client emails in the advisor's own tone and syncs CRM tasks automatically.
Draft client emails in each advisor's tone of voice
Altruist, 9 Sept 2025 ↗ - Hazel's "Ask Hazel" feature answers advisor questions using the firm's own recorded conversations, emails, documents and CRM data.
answers questions instantly based on a firm's recorded conversations, emails, documents, and CRM systems, as well as market and regulatory information
Altruist, 9 Sept 2025 ↗ - Industry analysis says 2026 AI value in wealth management concentrates in back-office standardization rather than client-facing personalization.
In 2026, AI personalization will matter far less at the client-facing layer than in the back office.
WealthManagement.com, 17 Apr 2026 ↗
- Altruist launched Hazel, an AI platform for advisors that drafts client emails in the advisor's own tone and syncs CRM tasks automatically.
Who's doing it
11 companies to know, 8 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
Comp AIcompliance automation
Agents that prepare and continuously maintain SOC2/ISO27001 compliance evidence, for companies going through security audits.
Latest · 17 Sept 2026Comp AI's agents handle onboarding, gather evidence, watch controls and run vendor security assessments instead of a human compiling an audit binder. SiliconANGLE ↗
$34M Series A, 17 Sept 2026 source ↗
IntegralSME accounting, tax and payroll
AI-native accounting firm that runs bookkeeping, tax and payroll for European SMEs.
Latest · 16 Sept 2026Integral's agents now handle end-to-end bookkeeping for over half its clients, cutting turnaround from weeks to hours. Tech.eu ↗
€18M Series A, 16 Sept 2026 source ↗
Naturalagent payments infrastructure
Orchestration layer and wallet infrastructure that lets a company's AI agents autonomously move and hold money.
Latest · 20 July 2026Natural built an orchestration layer so AI agents can transact with both humans and other agents. TechCrunch ↗
$30M Series A, 20 July 2026 source ↗
Payouts.comfinance operations agents (AP/AR, payouts)
Finance-ops platform whose AI agents run accounts payable, collections, treasury and month-end close, for finance teams.
Latest · 14 July 2026Payouts.com launched Digital Employee, long-running agents that OCR invoices, match purchase orders and call internal APIs. Fintech Global ↗
Arcade.devagent authorization and governance
Authorization and governance layer that controls what a bank's production AI agents are allowed to do.
Latest · 15 June 2026Arcade.dev gives production AI agents governed access to thousands of tools so enterprises can control what those agents do. SiliconANGLE ↗
$60M Series A, 15 June 2026 source ↗
Sardinefraud and risk
Agentic risk platform that profiles devices and scores fraud risk in real time, for banks and fintechs.
Latest · 21 May 2026Sardine's agentic risk platform detects and prevents fraud without adding friction for genuine customers. Fintech Global ↗
$25M Series C extension, 21 May 2026 source ↗
Rogocapital markets, dealmaking agents
Agent (Felix) that runs investment-banking workflows like deal screening, CIM generation and diligence, for banks and advisory firms.
Latest · 29 Apr 2026Rogo's Felix agent runs deal screening, CIM generation, buyer outreach and diligence without human intervention, live at five major banks. Fintech Global ↗
$160M Series D, 29 Apr 2026 source ↗
deepidvidentity verification and fraud
AI-native identity verification and fraud detection suite, for businesses doing KYC/AML.
Latest · 11 Mar 2026deepidv launched a fraud suite with deepfake detection, forged-document checks and sub-150ms transaction risk scoring. GlobeNewswire ↗
$1M Seed, 11 Mar 2026 source ↗
Basisaccounting and audit
Agent-based software that runs tax, audit and advisory workflows for accounting firms.
Latest · 24 Feb 2026Basis's agents run multistep accounting processes with rules-based controls, not just suggestions for a human preparer. SiliconANGLE ↗
$100M Series B, 24 Feb 2026 source ↗
Chestnutmortgage underwriting
AI-native mortgage lender whose agents shop rates and originate home loans directly, for homebuyers.
Latest · 9 Feb 2026Chestnut says its agents can automate 99% of mortgage operations, live in Texas and Colorado. Fondo ↗
Altruistwealth management, advisor tools
Custodian and AI platform (Hazel) that automates admin work for financial advisors.
Latest · 9 Sept 2025Altruist's Hazel answers an advisor's questions against the firm's own client records through a feature called Ask Hazel. Altruist ↗
Your move
Our read, built on the receipts above.
The open lane is authorization and audit trails for agents already in production, not another agent itself, that's why Arcade.dev exists and banks are the ones paying for it. Bookkeeping and mortgage-origination copilots are already crowded, Basis, Integral and Chestnut are all chasing the same workflow across different YC batches; the harder, less-crowded problem is what happens when one of these agents is wrong and a regulator asks who signed off.
Watch for a named, paying enterprise customer running a real workflow, not a pilot MOU, that's the gap between Rogo (five banks, disclosed production use) and half the 'agentic lending' pitches that cite a seed round and a demo video. Red flag: any pitch that claims full automation of a regulated decision, credit, audit sign-off, AML clearance, with no human step named anywhere in the flow.
This quarter, pick one back-office workflow you already do by hand, reconciling a batch of invoices or spreading financials for one loan file, and run it through a live agent product (Payouts.com's Digital Employee or Basis both publish demos) to find exactly where it hands back to a person, then check whether that handback point is where your actual risk lives.
The valuable skill is auditing a machine's shortcut, not producing the first draft yourself, which means knowing the underlying accounting, credit or compliance rule well enough to catch a plausible-sounding wrong answer. First step: take one AI-generated output your team already produces, a credit memo, a reconciliation, a compliance evidence pack, and write down every place you'd need to check by hand before you'd sign it.
Other industries
- Software engineeringSpaceX buys the leading AI coding startup outright, for $60 billion
- HealthcareAmbient scribes go from pilot project to clinical infrastructure
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing and resident calls move to voice AI agents that never sleep
- Sales and marketingCustomer support agents now get paid per resolved ticket, not per seat
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- InsuranceClaims move from adjuster-first to AI-first, humans supervise the exceptions
- CybersecurityAutonomous AI agents take over penetration testing and vulnerability hunting
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- Retail and e-commerceShopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
- EducationNYC and LA ban AI for students, not for teachers
- Media and creativeAI video splits into model labs that train and aggregators that resell
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare across a dozen-plus US cities
- Energy and data-centre powerGrid rules get rewritten mid-buildout, state by state
- GovernmentGSA turns AI adoption into a government-wide subscription, then flips 'free' to metered
- Agriculture and foodAutonomous tractor fleets learn to run the tools behind them, not just the wheel