What is Abrigo's Agentic Platform Experience (APX)?

Abrigo announced the Agentic Platform Experience (APX) on July 8, 2026, an AWS-built agentic platform for banks and credit unions that automates work across lending and risk. For lending, it covers the full life of a loan — pipeline, underwriting, closing, servicing, and portfolio administration — and, per Abrigo, its agents proactively complete tasks including document collection, data review, exception handling, and quality control rather than just tracking them. Abrigo estimates agentic AI can cut manual labor by more than 40%. General availability for lending is expected in Q3 2026. It's a capable orchestration layer. What it can't ship you is a clean loan file for the agent to reason over — that's still yours.

Can AI agents make loan underwriting decisions reliably?

An underwriting agent is only as reliable as the loan file it reads, and loan files are the messy part. The application, bank statements, tax returns, pay stubs, entity documents, and bureau pulls arrive as inconsistent PDFs and feeds where the same field means different things across products and loan officers. Income is the clearest example — gross versus net, base versus variable, a self-employed borrower's add-backs — and it's already the single largest critical-defect category in mortgage QC (ACES, 2025). An agent that reads that file faster doesn't resolve the ambiguity; it commits to one reading of it at machine speed. Reliability comes from defining the loan-file data contract first, not from a better model on top of an undefined one.

What's the biggest data-quality risk in AI lending?

That the agent turns a quiet internal data error into a booked loan or a denial. A transfer between a borrower's own accounts that reads as a deposit inflates income. A three-week-stale pay stub collected before the file moved to underwriting understates it. A bureau pull that returns two different names or addresses for one applicant makes the agent pick one. None of these are model failures — they're loan-file failures the agent inherits and acts on. Because lending decisions are regulated, the cost isn't a bad customer experience; it's a mispriced loan or an adverse action the lender has to defend.

How do you prepare loan data before deploying an AI underwriting agent?

Write the loan-file data contract before you point the agent at the credit decision. Define each field once — what 'income,' 'employed,' 'complete,' and 'verified' mean, per product — so the agent resolves them the same way every time. Structure the intake so documents land as typed fields, not free-text scans the agent guesses at. Reconcile the same borrower across the application and every bureau pull instead of trusting one source. Make the pipeline fail loud when a field is missing, stale, or contradicts another source, so a gap stops the file rather than flowing silently into a decision. Then start the agent narrow, on low-risk, high-volume steps with human sign-off, and widen only where the file is clean.

AI lending agents can run the whole loan. The credit decision still rides on the loan file.

Q: Who is legally responsible when an AI model denies a loan?

The lender — not the vendor and not the model. In Circular 2026-03 (May 2026), the CFPB reaffirmed that lenders using machine-learning underwriting models remain fully responsible under ECOA and Regulation B for giving specific, accurate reasons for any adverse action. A lender cannot claim the black box 'told us to do it,' and a vendor-supplied model doesn't shift the duty: the bank must obtain explainability outputs good enough for ECOA-compliant reasons and validate that those outputs are accurate. That makes loan-file data accuracy a compliance boundary, not just operational hygiene.

On July 8, Abrigo announced its Agentic Platform Experience — an AWS-built agentic system that, for lending, runs the full life of a loan: pipeline, underwriting, closing, servicing, portfolio administration. The line that matters is how Abrigo describes what the agents do. They don’t just track tasks. They “proactively coordinate and complete” work including document collection, data review, exception handling, and quality control. Read that list again. Two of the four items — data review and exception handling — are the words you use when the intake was never a clean data contract to begin with.

TL;DR: AI lending agents can now run the whole loan pipeline, and the orchestration is genuinely solved. But an underwriting agent makes a regulated credit decision on top of the loan file — bank statements, tax returns, pay stubs, entity docs, bureau pulls — which arrives as inconsistent documents where the same field means different things across products and loan officers. Income and employment errors are already the largest defect category in mortgage QC. A faster agent on a messy file doesn’t fix the file; it books the bad loan quicker, and under CFPB Circular 2026-03 the lender — not the vendor — owns the accuracy of the decision. Fix the loan-file data contract first, then run the agent.

This is the same shape I wrote about when an AI customer-service agent turns an internal record into a public promise. Lending is the higher-stakes version, because the promise is a credit decision and a regulator is holding the receipt.

The agent isn’t the hard part. The loan file is.

Adoption is real and moving fast. STRATMOR Group’s mortgage-industry survey found AI and machine-learning use among lenders jumped from 15% in 2023 to 38% in 2024 — more than doubling in a year. And the places lenders point it are telling: 63% use AI for document classification and indexing, 54% for document reading. In other words, most of the value today is spent wrestling the document package into something a system can use. That’s not a side task. That is the job, and it exists because the loan file was never structured in the first place.

Here’s what a lending agent is actually reading. A W-2 borrower’s pay stub, a self-employed borrower’s two years of returns with add-backs, twelve months of bank statements, a driver’s license, an entity’s operating agreement, a couple of bureau pulls. Some of it is typed. Most of it is a scan of a scan. The application says one thing, the paystub says another, the bank statement implies a third, and a human underwriter’s real skill was reconciling those into a defensible number. When Abrigo’s agent “reviews data” and “handles exceptions,” it’s doing that reconciliation — on whatever the file happens to contain.

Agents see the structured slice cleanly and guess at the rest. A transfer a borrower made between their own two accounts reads as a deposit, and the agent counts it toward income. A pay stub collected three weeks before the file reached underwriting is stale, and nothing flags it. A bureau pull returns two different spellings of the same borrower’s name, and the agent picks one and moves on. None of these are the model being wrong. They’re the loan file being ambiguous, and the agent resolving the ambiguity silently — which is exactly what you don’t want it to do on a credit decision.

Speed amplifies whatever the file already is

The defects aren’t hypothetical, and they cluster in precisely the fields an agent “reviews.” ACES Quality Management’s Mortgage QC Industry Trends reports have named income and employment the largest critical-defect category for quarters running; in Q3 2025 it rose to 27.24% of all critical defects, as the overall critical defect rate climbed to 1.79%. Across all of 2025 the critical defect rate averaged around 1.50%. Those are post-close findings on files humans underwrote. The error was in the loan file before any agent touched it.

Now automate the step. Abrigo estimates agentic AI can cut manual labor by more than 40%. Take that at face value — the manual review that would have caught the inflated deposit or the stale stub is the labor you just removed. Speed is neutral. Point it at a clean file and you underwrite faster. Point it at a file where income means three different things depending on the document, and you book a mispriced loan faster, then sell it into a pool where the defect surfaces months later as a repurchase demand. The optimizer got better. The substrate it optimizes over didn’t.

The regulator already named the owner

This is where lending stops resembling every other “agent makes a promise on bad data” story and becomes its own thing. In customer service, a confidently wrong answer costs you a CSAT point. In lending, a confidently wrong decision is a fair-lending exposure — and the CFPB has already said whose exposure it is.

In Circular 2026-03, issued May 2026, the Bureau reaffirmed that lenders using complex algorithms, including machine-learning underwriting models, remain fully responsible under ECOA and Regulation B for providing specific, accurate reasons for an adverse action. The language is blunt about the escape hatches. A lender cannot claim the black box “told us to do it.” A proprietary or uninterpretable model doesn’t excuse compliance. And a vendor-supplied model doesn’t move the duty off the bank — the bank has to obtain explainability outputs sufficient for ECOA-compliant reasons and validate that those outputs are accurate.

Sit with that last clause. Validate that those outputs are accurate. You cannot validate a reason for denial as accurate if the field it rests on — the income figure, the employment status, the debt total — came out of an un-owned document pile the agent interpreted on its own. The regulator has effectively made loan-file data accuracy the lender’s job, in writing, with a date. Deploy the agent on top of an undefined file and you haven’t automated underwriting. You’ve automated the generation of adverse actions you can’t defend.

What the agent does, and what the loan file still needs

The agent step	What it automates	What it inherits from the file	What owning the file adds
Document collection	Gathering and indexing the package	Whatever formats and versions the borrower and LO supplied	A structured intake where docs arrive as typed fields, not scans to guess at
Data review	Extracting income, assets, employment	Ambiguous fields — gross vs. net, transfers read as deposits, stale stubs	One definition per field, per product, resolved the same way every time
Exception handling	Routing mismatches for resolution	A mismatch it may silently pick a side on	A fail-loud rule: contradictions stop the file instead of flowing into a decision
Quality control	Post-hoc checks on the decision	Errors already baked into the inputs	Cross-source reconciliation of the same borrower before the decision, not after

The working version

The fix is the unglamorous one, and it comes before you turn anything on. Write the loan-file data contract. Define what “income,” “employed,” “complete,” and “verified” mean — per product, because a self-employed borrower and a W-2 borrower don’t share a definition — so the agent isn’t inventing one per file. Structure the intake so documents resolve into typed, validated fields instead of free text the model interprets. Reconcile the same applicant across the application and every bureau pull, so a two-name discrepancy is a stop, not a coin flip. And make the whole pipeline fail loud on drift — a missing, stale, or contradictory field halts the file rather than passing quietly into a credit decision no one can later explain to a fair-lending examiner.

That contract is the same layer under every one of these deployments — the integration and data-contract work that decides whether AI does anything useful. In lending it just happens to be load-bearing for compliance, not only for accuracy. Start the agent narrow: low-risk, high-volume steps, human sign-off, and widen only into the parts of the file you’ve actually made legible. Defining the file first is the work; the agent is the last step.

The operator read

Abrigo built a strong lending platform, and “document collection, data review, exception handling, quality control” is an honest description of what these agents do. The catch is in the second and third words of that list: an agent that reviews data and handles exceptions is an agent operating on a file that was never a clean contract — and it will resolve every ambiguity in that file, confidently, at machine speed, into a decision the CFPB says you own. If you’re about to point an underwriting agent at your pipeline and you can’t say who owns the definition of “income” across your products, that’s the conversation to have first.