How Dot Com Zambia Gives Fleet Managers Instant Answers Across 7.6 Million Toll Records

How Dot Com Zambia Gives Fleet Managers Instant Answers Across 7.6 Million Toll Records

Zambia’s eToll platform held years of toll data that only analysts could reach. Now anyone can ask it a question.

• 7.6M toll records searchable in plain English
• 32 toll plazas covered nationwide
• 24/7 answers, reports and exports on demand

What Dot Com Zambia was up against

Dot Com Zambia runs eToll, the electronic toll card platform used by enterprise fleets across Zambia’s road network. Fleets load prepaid eToll cards, drivers pay at plazas run by the National Road Fund Agency (NRFA) and other operators, and every pass lands in one central database.

That database grew to more than 7.6 million toll records, over 50,000 cards and around 1,400 client companies. The data was complete. Getting answers out of it was not.

• Every question became a ticket. A fleet manager asking “How much did we spend at Kafue plaza last month?” had to find the right report screen, set the filters and export, or ask the support desk.
• Reports were fixed. The admin console offered dozens of report screens, but any question that crossed them (by vehicle, by plaza, by month, against a card limit) meant a spreadsheet and an afternoon.
• Five kinds of users, five different views. Super Admins, Agent Managers, Client Managers, Station Managers and Station Cashiers all need different slices of the same data. Any self-service tool had to show each person exactly their own data and nothing more.
• Fraud signals needed explaining. The fraud engine classifies passes as Clean, Misuse, Fraud Pattern or Rapid Fire. Clients wanted to know why a card was flagged without waiting on an analyst.
• AI had to be safe before it could be useful. Financial data, many tenants and a regulated sector meant any AI had to be read-only, auditable, cost-controlled and impossible to talk into leaking another client’s records.

AI Workflows + Multilingual Voice AI Agent

What impleko.ai built

Impleko.ai designed and built Zuba AI, a business-intelligence assistant embedded directly in the eToll admin console. Users type a question in plain English and get a precise answer, a chart or a downloadable report in seconds.

• Answers questions straight from live data. Zuba AI writes its own read-only database queries, runs them against the eToll data and explains the result in plain language, with totals formatted in Kwacha.
• Knows the product, not just the numbers. A knowledge base built from eToll’s own manuals, tariffs and procedures answers “how do I” questions, such as loading a card, reading a statement or understanding a fraud flag.
• Builds reports on request. “Give me a monthly toll expense report by vehicle for Q2” produces a report with charts inside the chat, exportable to PDF, CSV or Excel, and can be emailed directly.
• Respects every role automatically. Each answer is scoped by the server to the user’s role, company and agency. The AI cannot choose to look further, because the scope is enforced outside the model.
• Hands off to people when it should. When the assistant is not confident, it says so and escalates to customer service by email with the full context, so the client never gets a guessed answer.
  

Governance built in, not bolted on

Because the assistant works with financial data across many client companies, InstantReply.ai built a full AI governance layer alongside it, backed by a written AI Governance Policy and a compliance audit:

• Read-only by design: every query is validated before it runs, and anything that is not a scoped read is rejected.
• Tenant isolation enforced in code: each query must bind to the caller’s own company or agency, supplied by the server rather than by the model.
• No internal details reach users: table and column names are automatically stripped from replies, so customers see business language only.
• Input and output checks: prompt-injection hardening, PII redaction and policy screening on both sides of every conversation.
• Cost and usage controls: per-user and per-company quotas, a monthly spend cap, per-model cost tracking and a global kill switch.

• Full audit trail: every question, query and answer is logged for review, with a retention job that clears old logs on schedule.
  
  Under the Hood

• AI models: Anthropic Claude (Haiku 4.5 and Sonnet) on AWS Bedrock, with streaming responses and prompt caching to keep answers fast and costs low.
• Knowledge base: Amazon Bedrock Knowledge Base with Titan embeddings, synced from S3.
• Backend: Node.js and TypeScript on Express, reading the eToll MySQL database on AWS RDS.
• Admin console: Next.js and React, with Zuba AI embedded as a chat panel next to the existing dashboards.
• Reports: charts rendered in chat, PDF generation with Puppeteer, CSV and Excel bulk exports, email delivery via Amazon SES.
• Delivery: CI/CD with GitHub Actions to AWS, separate development and production environments.
• Quality: an automated suite of 842 tests plus a dedicated evaluation harness that replays real questions against the assistant, alongside structured UAT with the client team.

What changed

• Answers in seconds, not tickets. Fleet managers ask about spend, cards, plazas and vehicles in their own words and get the figure immediately, at any hour.
• Reports on demand. Custom reports that used to mean a spreadsheet and an afternoon are generated, charted and exported from a single sentence.
• Every user sees only their own data. Five roles and roughly 1,400 client companies share one assistant, with tenant isolation enforced by the server on every query.
• Fraud flags explained in plain language. Clients can ask why a card or a toll was flagged and get the engine’s classification explained clearly.
• Hardened against real-world use. Issues surfaced during production review were fixed with structural guarantees rather than prompt rules, and the test suite roughly doubled from 423 to 842 tests.
• Safe to scale. Spend caps, quotas, audit logs and a kill switch give Dot Com Zambia full control over how the AI is used and what it costs.