Case: a fintech in México cut cost per ticket from $6.40 to $0.85 in 12 weeks

A fintech in México moved tier 1 support to an enterprise agent on WhatsApp and web chat connected to its core and its ticketing system. In 12 weeks, cost per resolved ticket went from $6.40 to $0.85 USD and NPS from 42 to 71. The client is anonymized at its request; the figures are verified in the DaCodes Proof Ledger (Sep 2026).

The situation

Dozens of people answering balance, payment and card questions on WhatsApp, with the CRM and the core open in two other windows. Handling time was fine. Cost was not: every conversation carried a person from start to end, and volume grew with every marketing campaign. The team had tried a conversational assistant before. It answered policy questions well and could not see an account, so most conversations ended in a transfer.

What we signed

Primary KPI: cost per resolved tier 1 ticket. Baseline: $6.40 USD, measured from the client's own ticketing data over the weeks before the pilot. Guardrail: customer satisfaction at or above baseline. Consequence: if the target set at Blueprint was not met at the commitment gate, no Scale phase and no retainer.

What we built, in four phases

• Blueprint (first two weeks): workflow map of the most frequent intents, data assessment of the core API and the ticketing export, eval plan built on real anonymized conversations as the first test set.
• Build: integration with the core (read balances, movements and card status; write payment promises and ticket notes), a context layer with product rules and policies, WhatsApp Business and web chat as channels.
• Controlled rollout: traffic moved to the agent intent by intent, A/B against the human queue on the same intents, weekly demos with the KPI on screen.
• Commitment gate at week 12: tier 1 traffic on the agent with human approval on refunds and limit changes, weekly evals on the Agent Console, one page with the number.

The stack, in the client's cloud

AWS Bedrock with AgentCore in the client's account, LangGraph for orchestration, a multi-model setup that routes simple intents to a small model and account actions to a stronger one, MCP as the tool interface to the core and the ticketing system, hybrid RAG over policies, PII scrubbing before any model call, and an audit log of every read and write. Repo in the client's GitHub.

Results at day 90

• Cost per resolved tier 1 ticket: $6.40 to $0.85 USD.
• NPS: 42 to 71.
• Every escalation arrives with the conversation, the account context and the suggested action attached.

Source: the client's ticketing system and the Agent Console, verified in the DaCodes Proof Ledger, Sep 2026.

What the operations team sees

On the Agent Console the support lead sees the KPI against its baseline, the pass rate of the last eval run, spend this month against the alert threshold, the audit log of every read and write, and the queue of pending approvals. Refunds and limit changes wait there for a person. Everything else runs. When a policy changes, the team edits the context layer, the evals run again, and the diff is visible before it reaches a customer.

What made the difference

Not the model. Three decisions did the work. The agent could read and write the core, so it could act instead of explain. The KPI was cost per resolved ticket, not deflection, so nobody had an incentive to close conversations early. And evals ran weekly on real conversations, so policy changes during the pilot were caught the same week they happened, not in the quarterly review.

What we would tell another fintech

Start with the intents where the answer lives in the core, not in a document. Balance, movements, card status and payment promises are the volume, and they are the ones an assistant cannot touch. Write the approval step for refunds and limit changes into the policy before the first conversation, and show it to the regulator's team early. And measure cost per resolved ticket from day one, with the client's own data, so that the number at the gate is nobody's opinion.

What happens now

The retainer covers evals, model updates and policy changes. The next process in the pipeline is collections, on the same platform, with its own KPI Contract.