← Back to the newsletter

NewsletterTendencias

The Price of a Resolved Conversation: Falling Token Costs and the New AI Budget in LatAm Financial Services

November 4, 2026 6 min read

Two Years of Cheaper Inference

At Money20/20 USA in late October, PicPay’s CEO shared figures that reframe the cost conversation in Latin American financial services. The unit cost of each AI token at the company is 98% lower than it was two years ago, and 94% of its 4,500 employees already use Nitro, its internal AI tool. In the same session he added a caveat: AI requires management, otherwise token consumption climbs so high that efficiency gains may not cover the bill.

Global data points the same way. According to Stanford’s 2025 AI Index, inference at GPT-3.5-level performance dropped from $20.00 per million tokens in November 2022 to $0.07 by October 2024. Consumption moved in the opposite direction: Ramp’s data shows enterprise token spend grew 13x between January 2025 and early 2026, and Gartner forecasts spending on inference-focused applications at $20.6 billion in 2026, up from $9.2 billion in 2025, overtaking training for the first time. The issue has become large enough that the Linux Foundation is launching a Tokenomics Foundation to bring cost discipline to AI spending.

For banks, insurers and BPOs in the region, the business case for AI now has to show the cost of each resolved conversation across months of production volume. A price per token, on its own, says little about the final bill, and November is when most 2027 service budgets are being approved.

References: Let’s Money, November 2026 / Stanford HAI AI Index 2025, via Workd / HPCwire, August 2026 / Gartner, via Nscale / The Next Web, 2026

AI in Finance: What Three Years of AI Service at PicPay Show

PicPay is a useful reference because its AI has been in front of customers for more than three years. An agent was already talking to customers in March 2023, and in the third quarter of that year all first-level service moved to AI, at a time when the market standard was to use AI only as a copilot. The transition followed a long validation period: the agent ran for six months in mirror mode alongside human service, and the switch only came when the AI’s NPS pulled ahead. The reported result was 45% more NPS points at 3% of the previous cost.

The cost curve came from operating decisions. The company chooses the model for each task, so the choice does not rest with the individual employee. Context is cached once, and each conversation sends only what is new. A separate model rewrites each request before it reaches the main model, and everything runs on a proprietary orchestration layer.

In a financial contact center, the same logic applies at the level of each interaction. Token volume in a service conversation grows with the size of the system prompt, the knowledge retrieved on each turn, calls to core systems, retries and, on voice, the minutes of recognition and synthesis. A balance inquiry and a debt renegotiation can differ by an order of magnitude in consumption, yet most reports count both as one conversation. Operations that keep AI costs predictable tend to route routine contacts (status checks, payment reminders, document requests) to lighter and faster models, and keep larger models for negotiation, complaints or document interpretation. They also retrieve knowledge only when the conversation needs it, and they report cost per conversation next to the result it produced: a payment promise, a renewed policy, a case closed in one contact.

Governance sits in the same design. PicPay opened plugins in external assistants for account queries while keeping authentication inside its own app, so the conversation can happen elsewhere while security stays under the institution’s control. The six-month mirror period is also a model that supervisors and audit committees recognize: a new automated process proves itself against a measured baseline before it takes full volume.

References: Let’s Money, November 2026

Coru Product: ADI Platform and the Cost of Each Conversation

ADI Platform is an enterprise platform for operating AI conversational agents over voice (PSTN inbound and outbound), WhatsApp Business and an embedded web widget, built for contact center operations that want to automate between 30% and 80% of their volume.

The right model for each agent

Every agent defines its main LLM, its temperature (low values are recommended for financial production) and one or more fallback models tried in order. Inference runs through a routing layer that selects providers by cost and latency, multicloud by design and without dependency on a single vendor.

Knowledge and data on demand

Knowledge bases linked to an agent are consulted in real time only when the conversation requires them. HTTP tools and an MCP gateway reach the client’s systems to fetch the specific data a case needs, such as a balance or a policy status, without changes to the core.

Cost visible by client, agent and day

FinOps dashboards show LLM cost per client per day and can trigger an alert when a threshold is crossed. The account view follows cycle consumption (input and output tokens, inbound and outbound voice minutes) against defined limits, with warnings as each resource approaches its cap.

Cost next to outcome

Analytics compares conversations, contact rate and cost against the previous period, filtered by channel, agent and date, alongside custom metrics the agents extract for each use case. In collections, that means payment promises, agreement rate, committed value and D+1 compliance on the same screen as the spend that produced them.

Validation before full volume

Every edit stays in draft until it is tested and published as a numbered version. Versions can be compared side by side and restored immediately, and canary deployment sends a new version to a percentage of traffic while its metrics are monitored, which gives a prompt or model change its own measured trial period.

One configuration for every channel

The same prompt, tools, knowledge base and guardrails serve voice, WhatsApp and web, so an optimization made once applies across all active channels.

Consumption per portfolio

An Organization can run agents for many Clients, each with its own plan and consumption, so a BPO can allocate AI cost to each financial portfolio it serves. Role-based permissions, configuration logs and server-side tenant isolation come standard.

Guardrails that stay on

A mandatory layer protects against prompt injection and personal data exfiltration and cannot be disabled, while a configurable layer defines when the agent responds and when it transfers the case to a person.

Coru Weekly Picks

Film Recommendation: Moneyball (2011)
Billy Beane, general manager of the Oakland Athletics, has one of the smallest payrolls in baseball and a roster to rebuild. With an economics graduate at his side, he starts paying for the statistics that actually produce wins and stops paying for the reputations that do not. The film is a good two hours on choosing the metric that links spending to results, which is the exercise every AI budget now requires.

Book Recommendation: How to Measure Anything, Douglas W. Hubbard
Hubbard argues that anything that matters to a decision can be measured, usually with less data than people assume, and that measurement effort should go to the variables whose uncertainty carries the highest cost. For leaders defending an AI line item, the book offers a practical method to estimate the value of a resolved conversation, an avoided transfer or a payment promise kept, before the numbers are perfect.

Looking Ahead

Unit prices will keep falling. Gartner projects that inference on a one-trillion-parameter model will cost 90% less by 2030, and still expects that cheaper tokens will not produce cheaper enterprise AI bills, because agentic models consume far more tokens per task. Financial institutions and BPOs that use 2027 planning to define model routing by task, retrieval discipline and cost reporting tied to outcomes will be able to scale automation with a budget that holds.

Provider strategy also connects cost to resilience. In Brazil, a Central Bank coordinator warned in September that many regulated entities contracting the same ready-made AI provider could create concentration risk, a point that appears in the preliminary results of an ongoing mapping. An architecture that can switch models and providers by task serves the finance team and the risk team with the same decision.

References: Cockroach Labs, citing Gartner / Let’s Money, September 2026

At Coru, we build the operational intelligence that sits between an AI model and a regulated process. In Latin America, that means conversational systems fluent in local channels and local rules, with compliance-first architecture underneath. Every interaction becomes a resolved case with a record behind it.

Get the next ones by email

Twice a month. One field, unsubscribe whenever you want.

More editions