← Use cases
Banking and mobile money

An advisor that never skips identity checks and never quotes an amount it has not read.

The business challenge

Repetitive call peaks

Balance, blocked card, loan due date: most of the volume is simple yet saturates the contact center.

Fraud targets support

Fake agents, SMS codes asked over the phone, staged emergencies: the conversational channel is an attack surface.

A wrong figure is costly

An approximate balance or a promised return creates a complaint, even regulatory risk.

What the agent does

  • Balance, history and loan after a one-time code check.
  • Urgent card block, then a branch appointment for the replacement.
  • Complaints qualified and summarised before being created.
  • WhatsApp, SMS, USSD and widget, with handover to a human advisor.

Connected systems : Core banking, mobile money (Wave, Orange Money), CRM and helpdesk through the platform connectors.

Verified commitments

Each commitment matches a scenario in the agent template's test suite. 18 test scenarios.

  • No balance or history without a verified OTP.
  • A wrong code unlocks nothing: the code is always checked by the tool.
  • Stolen card: the block goes out as soon as the code is verified, with its reference.
  • Quoted amounts are those returned by the systems, never guessed.
  • Never a PIN, never a shared SMS code, even with a fake bank agent.
  • A subscription or cancellation runs only after the customer's explicit agreement, enforced by code and not by the prompt alone.

Demo: an emergency, replayed

A real conversation from our test bench, on a demo system (fictional data). The SMS code is masked.

Customer : URGENT : on m'a volé ma carte bancaire ce matin !!

From technical rigour to business outcome

What we guarantee

Identity is checked by a tool, never “judged” by the model.

What it changes for you

The conversational channel opens no breach: social engineering does not get through support.

What we guarantee

Every amount and reference comes from a tool call.

What it changes for you

Fewer complaints born from an approximate answer, and an auditable history.

What we guarantee

Emergencies are handled first, with no unnecessary question.

What it changes for you

A stolen card is blocked in a few exchanges: the customer stays confident, the loss is limited.

Measured results

Demo banking agent. Same model, same information system, temperature 0, alternating trials.

Reference scenarios

18 scenarios, the ones the agent was tuned on.

Pass rate

Original state · 2 trials72.2 %
After optimisation · 2 trials94.4 %

Security

Original state77.8 %
After optimisation100 %

Validation scenarios

9 new scenarios, written separately and not used for tuning.

Pass rate

Original state · 2 trials55.6 %
After optimisation · 2 trials77.8 %

Security

Original state100 %
After optimisation100 %

What these figures say, and do not say

  • On the reference scenarios, the optimisation fixes specific flaws in the identity flow and adds a code-enforced confirmation before irreversible actions: measured over two alternating trials.
  • On the validation scenarios, the pass rate goes from 55.6% to 77.8% with 100% security in both versions. Caution: this set was then used to diagnose flaws, so the figure is optimistic; a brand-new set is being prepared. Part of the gain comes from code (confirmation before cancelling), part from fixing overly strict test criteria, applied to both versions.
  • Known flaws remain, revealed by the test bench: reading savings rates, a double confirmation on some subscriptions. They are on our roadmap.

Business study: size it with your assumptions

An estimate, not a promise: starting values are examples. The platform measures your real automation rate on your conversations.

Excludes platform and setup costs. Contacts handed over to a human remain at your cost.

Contacts handled by the agent
6,000
Gross saving per month
9,000,000 FCFA
Gross saving per year
108,000,000 FCFA

Let's see what it does on your business

Our experts scope your use case, connect your systems and measure the result on your own scenarios.