Use casesFinancial institutions & fintech
Early arrears collections
Calls reach customers who would have paid anyway, while the ones who cannot keep up need a plan, not ten calls.
- Kinds of AI
- ModelsSLMsAgentic AI
- Who decides
- Collector
- Where it runs
- Your servers, banking data kept in Türkiye
Model · roll to 60+ days
SLM · track
Hardship, not “won't pay”.
Agent · contact plan, in policy
- TuesdayOne call, 10:00–12:00, then an app message
- FrequencyAt most 1 a day
- Model
- SLM
- Agent
Written limitTerm over 12 months: the collections supervisor approves.
DecidesCollector
1. The desk
Today
- SortEveryone in a delinquency bucket gets the same treatment on the same day.
- ContactCollectors are at capacity; contact rules (hours, frequency, opt-outs) are tracked by hand.
- PlanRestructuring offers are improvised per call.
- OfferThe collector makes the offer, and its affordability arithmetic is not recorded.
With the system
- SortThe roll-rate and cure models put each account on a track: no action, digital reminder, call, or hardship.Model
- ContactThe agent runs the day's contact plan inside the channels, hours, frequencies, and scripts the contact policy allows, records promises, and hands disputes, complaints, and vulnerability flags to a person.Agentic AI
- PlanFor hardship cases the SLM drafts a plan taken only from the policy menu, with a script; affordability figures come from the calculator.SLM
- OfferThe collector decides whether to make the offer, as before; a supervisor approves above the threshold, and the arithmetic is on file.Person decides
What lands on the desk
The collector sees why this account is on the hardship track, the contact plan the policy allows, and one proposal taken from the bank's own menu with its affordability figure from the calculator. Whether to make the offer is the collector's call, with the supervisor where the term requires it.
Model · roll to 60+ days
SLM · track
Hardship, not “won't pay”.
Agent · contact plan, in policy
- TuesdayOne call, 10:00–12:00, then an app message
- FrequencyAt most 1 a day
SLM · proposal from policy menu §4.2
- Convert the card balance into a 24-month installment loan, first payment in 30 days
- AffordabilityInstallment within the policy ratio of declared incomecalculator
SLM · draft script
We see payments have been harder lately. We can spread the balance…
Not allowed
- flaggedLegal referral while a viable plan is open
- flaggedContact on a Sunday or holiday
- Model
- SLM
- Agent
Written limitTerm over 12 months: the collections supervisor approves.
DecidesCollector
2. The system
For your technical team — skip it if you like.
How it works
In
- Payment history, days past due by month
- Promises and the kept rate
- Contact outcomes
- Salary or turnover inflows, bureau refresh
Model · estimates
Roll-rate and cure models
Logistic or GBM, calibrated; cure is the propensity to pay
Reads
- Everything under In
Out
- Self-cure probability
- Roll to 60+ days past due
- Value at risk
- Track: none, reminder, call, or hardship
SLM · understands
Case-plan drafter
Reads
- Hardship and hard cases
- The policy's action menu
- Product terms
Out
- A plan chosen only from the menu
- A script in the bank's tone
- Affordability by a calculator, not the SLM
Agent · prepares
Contact agent
A state graph over the contact policy
Reads
- The tracks and the plans
Out
- The day's contact plan, run in policy
- Approved channels, hours, frequency, scripts
- Promises recorded
- Accounts moved between states
Decision
Collector
Reads
- Restructuring proposals
Out
- The collector decides the offer
- Above the threshold, a supervisor approves
How the agent's work is arrangedState graph
State graph
Each account sits in one state (reminded, promised, promise broken, hardship, referred), and the contact policy writes which moves and contacts are allowed from each, so the agent cannot invent a step.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Cure and roll-rate models
Target
Cure within 30 or 60 days, and roll to the next bucket.
Method and challenger
Logistic against GBM.
Validation evidence
Calibration and a roll-rate back-test.
Monitoring
Champion/challenger with a small randomized share, before any uplift model.
Later: uplift model
Once champion/challenger tests exist, an uplift model shows which action changes the outcome (a T- or X-learner, two standard ways to estimate the effect of an action, judged on a Qini curve of the gain it adds).
Small language model (SLM)
Test set
Hardship cases graded by senior collectors before training: a plan from the menu only, tone, no invented concession.
Dataset
Approved past plans and scripts, with personal data removed.
Continued pre-training (CPT)Skipped
Skipped: collections vocabulary is ordinary.
Supervised fine-tuning (SFT)
On the plan format and tone.
Reinforcement learning (RL)Skipped
Skipped: there is no checkable reward beyond "menu item and calculator figures match", which a rule check enforces.
Where the agentic system stops
Every restructuring offer is the collector's decision, and a supervisor approves above the written threshold.
Show its tools, limits, and record
- Tools and access
- Reads collections and core systems; sends only approved reminder templates through approved channels, inside the contact hours, attempts per day, channels, and scripts the policy owner sets (by default the stricter of bank policy and the asset-management-company rule); books calls in the dialer.
- May not
- Offer a concession, refer to legal, write off, or mention the debt to a third party.
- Who decides and the written limit
- CollectorEvery restructuring offer is the collector's decision, and a supervisor approves above the written threshold. The collections policy owner writes the action menu, contact rules, and concession limits; reminders go out only through the channels, hours, and scripts the bank's contact policy approves. Legal referral, settlement discounts, and write-off are always a named person's.
- Stops when
- A dispute, a complaint, a vulnerability flag, or a "do not call" request: the account leaves the automatic graph for a person.
- Record
- Every contact, template, state change, and promise.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Lower roll rates and charge-offs.
- Collector capacity freed from customers who self-cure.
- Sustainable restructurings, and fewer complaints.
What your institution measures
- Roll rate 30→60 and 60→90
- Cure rate
- Cash collected per unit of cost
- Promise-kept rate
- Contacts per resolution
- Re-default of restructured loans at 6 and 12 months
- Complaint rate
Measured in the Decision Blueprint
Roll and cure rates by bucket set the baseline; your institution sets the target, confirmed by champion/challenger.
Conditions
The rules, the data, and the known error, with its guard.
- Under EBA/GL/2018/06, forbearance must be viable, not a delay, and the EU Consumer Credit Directive 2 requires reasonable forbearance before enforcement.
- In Türkiye, the asset-management-company regulation limits contact hours and attempts; no identical rule was found for banks, so the stricter rule applies, and harassment can be criminal (TCK 123).
- Hardship data is sensitive under KVKK.
Start here
We measure your roll and cure rates by bucket, map the contact rules your collectors follow today, and read a sample of hardship cases against your action menu before anything is built.
Similar desks
- ModelsSLMsAgentic AI
TodayReferred files wait while analysts rebuild the firm from four systems and last year's committee notes.
With the systemOne file with the rating, its reasons, and the authority level.
Credit manager decides
- ModelsSLMsAgentic AI
Early warning on the loan book
TodayThe borrower is current; other banks are already cutting its limits, and you see its quarter only at renewal.
With the systemA file, not a spreadsheet: signals checked in parallel and gathered in one place.
Watch-list committee decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/collections · Printed copy; the website is the current version.