Every dispute starts with intake. An agent listens to a cardholder, decides what kind of dispute it is, and types in the details by hand. They do this while the customer waits on the line, against FCBA and Reg E clocks, holding 30 to 50 open cases at once. It is the highest-pressure, lowest-margin moment in the whole lifecycle, and it sets everything downstream.
A wrong turn here does not fail loudly. A misclassified dispute clears intake, sits in the wrong queue, and resurfaces weeks later as a missed regulatory deadline. By then the bank eats the loss, the cardholder is still out their money, and a reviewer is reconstructing what happened from notes that never quite agreed.
This was not a speed problem dressed up as a compliance problem. It was the reverse. The decisions made at intake were invisible: who decided what, on what basis, and where the judgment came from. In a regulated workflow, a decision you cannot reconstruct is a liability, no matter how fast you reached it.
Agents read the dispute and chose a type from memory, under time pressure, with no second pair of eyes.
A small mistake at intake compounds quietly downstream, where it is expensive to find and slow to correct.
When a reviewer asked why a case was handled a certain way, the answer lived in someone's head, not in the system.
The obvious move was to automate intake outright. Let the model read the dispute, classify it, fill the fields, and move on. We rejected that, and the reason is the whole point of this case study.
That is the trap in a lot of agentic UX: the system can act, but the person cannot tell whether it should have. Speed helps only if the user can inspect the reasoning, correct the output, and explain the final decision later.
Picture the trust-breaking moment. The model returns a classification in under a second. It is probably right. But the agent has no way to see what it read, why it landed there, or how sure it is. So the agent does the only safe thing under a regulatory clock: they ignore it and start over by hand. A suggestion you cannot inspect is a suggestion you cannot defend, and in this domain that means it gets thrown away.
Automate the decision and agents stop trusting the system. Hold every case for full manual review and the speed gain disappears. Push the model harder and you take accountability away from the person the regulator holds responsible.
Design for calibrated trust. Trust the AI when it is right, review when it is uncertain, override when it is wrong, and make the boundary between AI and human explicit at every step. The human in the loop is not a concession. It is the design.
Pure automation optimizes for the wrong thing. In regulated work, the goal is not to remove the human. It is to make the human faster and more sure, with a record of why.
The model reads the dispute in plain language and pre-fills the attributes of the case: the amount, the date, the merchant, the dispute type, and the rest. Each populated field carries its own confidence. From there the agent takes over. The closed loop below is the spine of the whole system, and the order matters.
Reading, extracting, classifying, and showing how confident it is. Steps 1 and 2. The model proposes. It never disposes.
Deciding. Step 3 is the boundary. Every field is the agent's call, and the audit log records that it was theirs.
The first version showed raw percentages. "85% confident." It read precise, and it failed in practice. A probability gives an agent no instruction. Worse, a high-looking number invited automation bias: people accepted "85%" without checking, even when the underlying field was wrong.
I replaced the number with a plain confidence headline, and paired each level with a next step. The reader engages the reasoning before they ever reach a score. The raw value is still available for anyone who wants it, but it is de-emphasized on purpose.
The agent describes the dispute in plain language. The model pre-fills the attributes, and each one carries its own confidence chip, so the agent knows at a glance which values to accept and which to check. Nothing is committed until the agent decides.
This is the same screen a Fiserv agent saw, rebuilt in the DWS design system. The AI has already read the cardholder's complaint and pre-filled five fields, each tagged with how sure it is. Press start and do the agent's job yourself. Nothing is submitted for you.
A guided 90-second walkthrough of the AI intake: accept what the AI nailed, verify what it is unsure of, and catch the one thing it got wrong. Every decision you make writes itself into a live audit log.
"Human in the loop" is easy to say and easy to fudge. To hold up under a regulatory review, the boundary has to be explicit: who assists, who automates the already-approved, and who actually decides. Here is the split.
AI assists the judgment calls, rule-bound automation handles the approved repetitive actions, the human owns every judgment, and the audit log records who decided what. The boundary between the three is always explicit, never blurred to make a demo look smarter.
Compliance needed a record: who decided what, on what basis, when. That much was a requirement. But because the AI intake was a brand-new feature, there was no historical baseline to measure it against. There were no past accuracy numbers, no prior override rates, nothing to compare to.
So I designed the audit log as the place where the product could learn. The record still satisfied the regulator, but it also showed the team where the AI was trusted, corrected, or overconfident. An override was no longer only a correction. It was a signal about what to fix next.
Each row captures the full interaction: what the AI suggested, how confident it was, what the human did, the value that was committed, and when. A compliance reviewer reads it top to bottom. A product analyst reads it down the columns.
The log made every AI suggestion traceable, and it gave the team a practical way to learn from real decisions.
Read down the override column and the model tells on itself. If high-confidence fields are often overridden, the calibration is wrong. If low-confidence fields are often accepted, the model is under-confident. If one field is corrected again and again, its extraction rule or prompt needs work. If one dispute type drives most of the overrides, it needs more domain support. The audit log turns each correction into a piece of evidence about what to fix next.
Because every decision is structured, a whole class of measurements becomes possible. These are the questions the design lets the team ask once enough cases flow through. They are framed as what the instrument can read, not as numbers I am claiming.
Darker cells mean agents corrected the AI more often. A dark cell under High confidence means the model may be too sure. A dark cell under Low confidence means the field may need better extraction support.
Put the pieces together and a cycle appears. The AI suggests, the agent decides, the log records, the team learns, and the next suggestion can be tuned with real behavior behind it. The record that keeps the work compliant also shows where the AI needs to get better. That is the idea I most want this case study to carry.
High-confidence fields overridden a lot? The calibration is probably wrong. The model is too sure.
Low-confidence fields accepted a lot? The model is under-confident, hiding good answers behind a cautious label.
One field corrected again and again? Its extraction rule or prompt needs work.
One dispute type driving most overrides? That category needs more domain support.
In regulated AI, auditability is not overhead. It is how the product learns. The record you build for review is also the record that shows where the model needs work. Design those needs together.
Confidence has to tell you what to do. A probability is not guidance. Pairing a plain label with a next step beat a precise number, because it changed behavior instead of just reporting it.
The boundary is the product. The problem with agentic UX is not autonomy by itself. It is unclear accountability. In high-stakes work, people need to know what the AI did, what they decided, and what gets recorded.
I redesigned a regulated workflow so AI could accelerate intake without weakening human accountability. The AI classifies, the human confirms, and every decision stays auditable. That record is what let the team learn from real corrections instead of guessing where the model needed work.