Use caseRetail banking
1,284 complaints closed in Q3, all on time, and 27 percent filed under Other. One failure hides 208 times in there. Proving it took nights.
Every closed complaint gets a cause from your own listTwo handlers check a sample before anything is reportedThe fix goes to the process owner, with a date
One cause on your list fits · the reasoning sits on the case
Card dispute refund delayed. The customer was promised 10 days and waited six weeks for $389.
source · customer message of 08/14, answer of 08/2027 percent. That is the share of last quarter's 1,284 closed complaints filed under Other. Every one was answered, most on time. The customer's words sit in the messages, nobody has the evenings to read thirteen hundred of them, and the report to the examiners counts complaints without saying why customers wrote. The quarter is handled case by case and never read as a whole.
The team writes down the causes it believes in, and the AI classifies every closed complaint against them, with its reasoning on the case. Handlers check a sample, the process owner confirms the cause, and the head of complaints signs a report that names it and a plan that fixes it.
Dashboards, cause lists, sample reviews: a complaints team knows each of these. Run in this order, they explain a quarter: the causes are written by people, applied to every case, checked on a sample, confirmed at the source, then reported and fixed. The nine steps below go from Dana's dashboard on October 5 to the January figures that show the fix held.
Each complaint was handled in its own case. Dana now reads all of them, through causes her team wrote.
On October 5, Dana Whitfield, head of complaints, opens her home page for the third quarter. Her reports count the closed complaints across channels, the share answered within your deadlines and the share upheld. In the old category field, 27 percent sit in Other.
Dana and two senior handlers write twelve causes in plain language, each with a short description and an example: what they see in the messages week after week. She sets them up as a business task in the console, with the instructions, the model and a confidence threshold under which a person reads the case.
With 1,284 cases, the task runs asynchronously. For each one, the AI reads the customer's messages and the answer sent, and writes two fields: the cause and the reasoning in plain words. 61 complaints fall under the threshold and go to handlers; 23 fit no cause.
Before anything is reported, handlers check the classification and the process owner confirms the cause.
Dana asks the repository in plain words for the upheld card dispute complaints where the answer gave no reason for the delay. The question becomes a query; the result is a table she can export to Excel. Clicking a row opens the case, and she reads three in full.
Dana sends Malik and Jenna, two handlers, a sample of 40 cases in the largest cause, as review tasks. They confirm 36. Three belong to a merchant who refused the dispute, and one is a fraud claim. They move the four, and the cause settles at 208.
Dana meets Nathan Ford, who runs card operations. The card dispute process runs on the workflow engine, so each case shows when it reached each stage. The wait sits after the merchant's acceptance: accepted disputes wait for a weekly batch, and a manual check on amounts above $300 adds two more weeks. Nathan confirms the cause.
The cause goes into the report to the examiners, and the fix goes to the person who runs the process.
Dana fills your quarterly report with the counts from her reports. She asks for a draft of the narrative on the main causes; every figure in it cites the report or the sample it comes from. She rewrites it, Ryan Dupree in compliance reviews it, and she signs.
Nathan commits to two changes in card operations: a daily refund run for accepted disputes, and the manual check raised to $1,500. Dana adds hers: the answer draft's instructions now ask it to explain the refund timing, a change made in the console without a release. Compliance and Dana decide the redress for customers who paid overdraft fees while waiting.
In January, the same classification runs on the complaints closed since the change. Card dispute complaints drop from 69 a month in the third quarter to 18 in January. Dana closes the action with the report attached.
Twelve causes, written by Dana's team, are the only labels in play. Each closed complaint comes back confirmed, sent to a handler, or flagged as fitting none, and each answer starts different work.
The cause and a sentence of reasoning are written on the case, above your confidence threshold.
Then: the case counts in the cause report, and a sample is checked by handlers.Two causes could fit, or the messages are thin. The case goes under the threshold.
Then: a handler reads it and chooses the cause, with their name on it.A complaint about something the team has not named yet. The AI says so and gives its reasoning.
Then: the head of complaints reads them and decides whether a new cause belongs on the list.It answers no customer. It decides no redress. It signs no report. Across the nine steps on this page, the cause list, the sample check, the report and the fix are decided by people, each with their name on the case.
The AI classifies, searches, summarizes and drafts. The control is in the list the team wrote and the threshold set in the console, so the reading holds when the quarter is two thousand complaints long.
Counts from one retail quarter of 1,284 closed complaints. With a shorter cause list or a higher threshold, more cases would go to handlers.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
Complaints belong to the business, get read by compliance and run on IT's systems. Each of the three has to like what they see here.
You know the same failure repeats, and proving it takes nights.
Examiners get a count of complaints and very little about why.
A new category still costs a change request on the reporting tool.
Take the failure your handlers mention most, write it down with one example, and classify last quarter's closed complaints against it. The count that comes back is your first argument.
Your complaint cases, reports and workflows stay as they are. The classification writes two fields onto cases you already have.
See it live
Bring a few hundred anonymized closed complaints and the causes your team believes in. We will classify them with you and see which cause tops the list.