Agentic AI in Lending: What an AI Credit Analyst Agent Actually Does
An AI credit analyst agent decomposes a goal — "assess this borrower" — into steps, calls tools to execute them, carries state between them, and adapts the plan when a step returns something unexpected: a missing year, a mismatched entity, an unreadable page. That is the honest definition. Most products sold as agentic in lending do none of the last part.
The distinction is worth money, because the two things cost differently and fail differently. What YuSight ships is deliberately bounded: a 28-minute sample CAM carrying 142 citations, produced by a pipeline with adaptive steps and hard stops, not by an autonomous analyst. This page is about how to tell the categories apart.
Key facts
- Gartner estimates that only about 130 of the thousands of vendors claiming agentic AI are real, and predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 on "escalating costs, unclear business value or inadequate risk controls" (Gartner, 25 June 2025). Gartner's word for the rest is "agent washing".
- Measured agent reliability is far below marketing. On τ-bench, which tests tool-using agents against simulated users and domain rules, state-of-the-art function-calling agents "succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail)" — meaning under one in four tasks succeeded on all eight repeated attempts (Yao, Shinn, Razavi and Narasimhan, 2024).
- The supervisory perimeter explicitly excludes them. SR 26-2 / OCC Bulletin 2026-13 (17 April 2026): "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance" (Federal Reserve; OCC). That is a gap in coverage, not a clearance to deploy.
- Central banks are saying the oversight model does not scale. The Bank of England's Deputy Governor for Financial Stability, Sarah Breeden, told the ECB's Sintra forum on 30 June 2026 that "our frameworks were not built to contemplate autonomous agents" and that "relying on a human in the loop for all agent actions is unlikely to be realistic" (Bank of England, *Agents of change*).
- YuSight's sample CAM: 28 minutes, 142 citations, every figure traced to source document and page, with analyst-editable spreads and full version history.
What actually makes a system agentic?
Four properties. A system needs all four to earn the word; most lending products have one or two.
1. Goal decomposition. Given an objective rather than a task list, the system works out the steps. "Assess this borrower for a $2.4m term loan" becomes a plan. A system where an engineer wrote the step sequence has skipped this.
2. Tool use. It calls things — a document classifier, a bureau API, a spreading engine, a ratio calculator — and reads the results. This is the property nearly everyone has, and the one nearly everyone points at when asked to prove they are agentic.
3. State across steps. What it learned at step 4 changes what it does at step 9. Not a variable passed forward; an evolving picture. A discovery that the borrower has an unconsolidated subsidiary should cause it to go looking for that subsidiary's accounts.
4. Self-correction against feedback. A step returns something wrong or incomplete, the system notices, and it does something different. This is the hard one, and it is where the marketing and the engineering diverge most sharply.
The uncomfortable observation: properties 1 and 4 are the ones that make a system genuinely agentic, and they are also the two that make it hardest to govern. A system whose plan you cannot predict in advance is a system whose behaviour you cannot fully test in advance.
What does an agent actually do on a commercial file?
A realistic trace, on a three-entity borrower with 47 documents. The middle column is the one that matters: how many of these steps involve a decision the system made rather than a step someone wrote.
# | Step | Genuinely agentic? |
|---|---|---|
1 | Classify all 47 uploads by document type and period | No — fixed pipeline step |
2 | Detect that three distinct legal entities appear in the set | No — classifier output |
3 | Notice FY2024 accounts missing for the holding entity | Yes — gap detection against an expected set |
4 | Decide whether the gap is fatal, and search the file for a substitute (a consolidated statement, a tax filing) | Yes — planning under uncertainty |
5 | Request the missing document from the relationship manager | Yes — action selection, if it can actually do it |
6 | Spread each entity's statements | No — fixed pipeline step |
7 | Compute ratios per entity and on a combined basis | No — deterministic computation |
8 | Detect that combined turnover double-counts an inter-company sale | Yes — anomaly detection driving a re-plan |
9 | Re-run the combination with an elimination entry | Yes — corrective action |
10 | Reconcile bank credits against declared turnover | No — fixed reconciliation |
11 | Investigate a 9% variance by testing candidate explanations against the balance sheet | Yes — hypothesis generation and testing |
12 | Pull the bureau report and compare disclosed debt against the borrower's own schedule | No — fixed step |
13 | Find an undisclosed facility in the bureau data and re-run DSCR with it included | Yes — discovery changing an earlier computation |
14 | Assemble the memo, flag four exceptions, stop | No — assembly |
The fourteen-step trace is an illustrative reconstruction of a realistic file, not measured product behaviour.
Six of fourteen steps involve real agency. Eight are a pipeline. That ratio is roughly what a good system looks like today — and it is worth noticing that the eight pipeline steps are the ones carrying the numbers, while the six agentic ones are about what to look at next. That division is not an accident; it is the safe design. The architecture that keeps figures out of the model's reach is set out in AI hallucination risk in credit analysis.
Why does autonomy compound errors instead of averaging them?
Because steps multiply. This is the single most useful piece of arithmetic in the whole subject.
Take the 14-step run above, and suppose each step is independently correct 97% of the time — a rate most vendors would be delighted to claim.
Probability of a fully clean run = 0.97^14
0.97² = 0.9409 0.97⁴ = 0.9409² = 0.885293 0.97⁸ = 0.885293² = 0.783743 0.97¹⁴ = 0.97⁸ × 0.97⁴ × 0.97² = 0.783743 × 0.885293 × 0.9409 = 0.652836
65.3%. Roughly one file in three contains at least one error somewhere in the chain, at a per-step accuracy of 97%.
Push per-step accuracy to 99% and it becomes 0.99^14 = 0.868746 → 86.9% — still one file in eight. Drop it to 95% and 0.95^14 = 0.487675 → 48.8%, worse than a coin.
This is exactly what τ-bench's pass^k metric measures, and why the gap between "succeeds sometimes" and "succeeds every time" is so wide: under 50% single-attempt success collapses to under 25% across eight attempts.
The architectural response is to shorten the unchecked chain. Put a deterministic gate after every four steps — a cross-foot check, a reconciliation band, an analyst confirmation — and the longest run of unverified reasoning drops from 14 steps to 4. Per-segment clean probability rises from 0.652836 to 0.97⁴ = 0.885293 → 88.5%, and, more importantly, an error that does occur is caught at the next gate rather than propagating into a ratio, a covenant test and a recommendation.
The gates do not make the agent smarter. They make its mistakes local. That is the entire design goal, and any vendor whose pitch runs on end-to-end autonomy has chosen the opposite one.
Where is agentic lending real, and where is it marketing?
The claim | What is usually underneath | The test that separates them |
|---|---|---|
"Agentic document collection" | A checklist comparing uploads to a required-document list, plus an email template | Give it a file with a consolidated statement where separate entity accounts were expected. Does it recognise the substitute, or just report the gap? |
"Autonomous credit analysis" | A fixed spreading and ratio pipeline with an LLM writing the narrative | Remove one year of accounts. Does the plan change, or does the pipeline run and produce a two-year trend labelled as three? |
"The agent decides" | A rules engine emitting a decision, with model-generated prose explaining it | Ask what happens when two rules conflict. A rules engine has a precedence order; an agent would reason about it |
"Self-healing extraction" | A retry loop on the same model with a different prompt | Feed the same unreadable page twice. Does the second attempt use a different method, or the same method again? |
"Multi-agent underwriting" | One model called several times with different system prompts | Ask what the agents disagree about, and what resolves the disagreement. Usually: nothing, and nothing |
"The agent learns from your analysts" | Overrides logged for later retraining that has not been scheduled | When did the last retrain happen, on how many overrides, and what changed measurably? |
None of these underlying things is bad. A checklist that chases missing documents saves real hours. A rules engine with clean precedence is exactly what a credit policy should be. The problem is the label, because a lender who thinks they bought judgement will staff and govern accordingly.
Four tests to tell an agent from a workflow
Run these in a demo, on your file, not theirs.
- The perturbation test. Change one input and see whether the plan changes, not just the output. Remove a year, add an entity, corrupt a page. A workflow produces a different result through the same steps. An agent takes different steps.
- The dead-end test. Give it a task it cannot complete — a facility with no supporting documentation at all. A workflow produces an empty section. An agent says what it tried, why it failed, and what it needs.
- The trace test. Ask to see the execution trace for one file: what it decided to do, in what order, and why. If there is no trace, there was no plan, and if there was no plan there was no agency.
- The variance test. Run the same file three times. A workflow returns identical output. A genuine agent varies — and the variance is your governance problem, which is why you need to see it before you buy rather than after.
Test 4 catches something buyers rarely think about. Non-determinism is the cost of agency, not evidence against it. A vendor claiming both full autonomy and byte-identical reruns is describing a pipeline.
What does the governance gap actually mean for a lender?
Three facts sit awkwardly together.
Creditworthiness assessment for natural persons is classified as high-risk under the EU AI Act at Annex III point 5(b) — "AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score" (EU AI Act).
US model risk guidance, revised in April 2026, places generative and agentic AI outside its scope by name, and the agencies have signalled a future request for information on AI rather than rules today.
And the Bank of England is publicly saying the frameworks "were not built to contemplate autonomous agents", with human-in-the-loop review of every agent action described as unrealistic at scale.
Put together: the thing you are being sold sits in a space where one regime calls it high-risk, another declines to cover it, and a third says the oversight model needs rebuilding. That is not a reason to avoid the technology. It is a reason to be able to describe, precisely and on paper, which components of your stack are deterministic, which are statistical models, and which are generative — because those three go to three different governance processes. The category work is in what is a credit decisioning platform, and the supervisory detail in model risk management for AI in credit analysis.
What level of agency is appropriate in commercial credit today?
A ladder, with an honest view of where the top of the market actually is.
Level | What the system does | Appropriate in credit today? |
|---|---|---|
0 | Fixed pipeline, no adaptation | Yes — this is most production spreading |
1 | Adaptive gathering: detects gaps, requests documents, finds substitutes | Yes, with the request routed to a human |
2 | Adaptive analysis: re-plans when it finds an anomaly, tests explanations, re-runs computations | Yes, if every re-plan is traced and every figure stays cited |
3 | Autonomous recommendation: forms and defends a credit view without a human in the loop | No. Nothing in the measured reliability record supports it |
4 | Autonomous decision and booking | No, and it is not a technology question |
Levels 1 and 2 are where the value is and where the engineering is real. Level 3 is where the marketing is. The gap between them is not a year of model progress; it is the difference between a system that decides what to look at and one that decides what a lender should do, and only the first has a defensible failure mode.
The practical shape of level 2 in production is a system that adapts its investigation and refuses to adapt its arithmetic — the analyst override workflow around it is designed in human-in-the-loop credit review, and the wider sequencing in underwriting automation in commercial lending.
FAQ
What is an AI credit analyst agent in commercial lending?
A system that takes a goal like "assess this borrower", works out the steps itself, calls tools to run them, and changes its plan when something comes back missing or inconsistent. In practice today it does the investigating, while deterministic components do the arithmetic.
How is an agent different from a chatbot on documents?
A chatbot answers the question you asked about the documents you gave it. An agent decides which documents it still needs, goes and gets them, notices when two of them disagree, and does something about it. The difference is initiative, not fluency.
Can an agent be trusted with a credit decision?
No, on the current evidence. Benchmarked tool-using agents succeed on under half of realistic multi-step tasks and are consistent across repeated runs under a quarter of the time. Trust it to gather and investigate; do not trust it to conclude.
How do I know if a vendor's "agentic" claim is real?
Change something about the file and see whether the plan changes rather than just the answer. Then ask for the execution trace. No trace means no plan, and no plan means it was a workflow with a language model in one step.
Why do agent errors compound rather than average out?
Because a multi-step run is a product, not a mean. At 97% accuracy per step, a 14-step run is clean 65.3% of the time. Reliability falls geometrically with chain length, which is why checkpoints matter more than model quality.
Does SR 26-2 cover agentic AI?
No. The April 2026 guidance says generative and agentic AI are not within its scope. That leaves a governance gap you have to fill yourself, not a permission to deploy without controls.
Should the agent be allowed to compute ratios?
No. Arithmetic belongs in code, with the formula visible and the inputs cited. Let the agent decide which ratios are worth computing and which inconsistencies are worth chasing — those are judgement about scope, not calculation.
What is multi-agent underwriting, really?
Usually one model invoked several times with different instructions. It can help by forcing a task to be broken up, but it is not a panel of independent reviewers, because the same blind spots appear in every seat.
Where does agentic AI genuinely save time today?
Document gathering and exception investigation. Working out what is missing, finding a substitute, chasing a variance to its explanation — these are search problems with checkable answers, which is where the technology is strongest.
Key takeaways
- Four properties define an agent: goal decomposition, tool use, state across steps, and self-correction. Most lending products have tool use and call it agentic.
- On a realistic commercial file, roughly six of fourteen steps involve genuine agency — and they are the steps about what to look at, not the steps carrying numbers.
- Reliability compounds downward: 97% per step over 14 steps is a 65.3% chance of a clean run. Shorten the unchecked chain with deterministic gates.
- Gartner counts about 130 real agentic vendors among thousands claiming it, and expects over 40% of agentic projects to be cancelled by end-2027.
- Levels 1 and 2 — adaptive gathering and adaptive analysis — are appropriate now. Level 3, autonomous recommendation, is not supported by any measured reliability record.
- The regulatory position is a gap, not a clearance. Be able to say which parts of your stack are deterministic, which are models, and which are generative.
YuSight is deliberately a level-2 system. It adapts what it investigates — gaps, entity mismatches, variances — and refuses to adapt its arithmetic: figures are extracted with citations, ratios computed in code, exceptions raised rather than resolved silently. The output is a Credit Assessment Memo where a 28-minute sample carried 142 citations, each one resolving to a source document and page, with full version history of what the machine produced and what a human changed. The pipeline behind it is described in how source-cited memo drafting actually works.
See your first CAM in 30 minutes — book a live demo.
This article is general information for credit and risk professionals. The 14-step trace and the agency ladder are illustrative frameworks, not measured product behaviour.
