Customer story · Energy16 languages, 4 channels, one assistant: DIPO, the official investor assistant for the Dangote Petroleum Refinery IPO
Dangote Petroleum Refinery wanted a single official channel for its IPO that would answer investors' questions, direct them to the right institution and warn them against fraud. CBOT configured its platform, already in production across banking and finance, and took DIPO, the official investor assistant, live within weeks in 16 languages over voice calls, USSD, WhatsApp and web chat.
DIPO was built as the investor communication platform for the Dangote Petroleum Refinery IPO, a national-scale public offering in Nigeria. The issuer announced the platform on its official website as the official investor contact centre.
The task was demanding: millions of potential users, a short window and demand that peaked on specific days. A large share of users relied on basic feature phones without internet access. The system had to speak 16 languages, comply with capital markets rules and be one trusted voice that dozens of banks and fintechs could reach from their own channels.
CBOT configured and scaled its AI Agent platform for this need without writing a new application from scratch. The project started in August 2026, and partner integrations and acceptance tests ran in early September. DIPO went live on 14 September 2026 with all 16 languages active from day one. On launch day, 18 of the 28 targeted banks and fintechs were live, 9 of them on both voice and text. Development was run from Istanbul, and the team was on the ground in Nigeria during launch week.
A physical and a virtual war room were set up for the launch. A 10-scenario fraud and emergency response protocol with defined roles and escalations was tested in a tabletop exercise. Acceptance testing used a script of about 63 questions in 5 languages. CBOT hosted and operated the platform end to end as a managed service.
“The real test of AI is not in the demo room. It is on the day millions of people reach for their phones at the same time. With our team in Istanbul, we took a system that handles 10,000 simultaneous voice calls in 16 languages live in a very short time. It is proof that Turkish engineering plays in the global league of AI.”
A fixed-capacity call centre could not keep up with demand that spiked within a short window; the system had to scale elastically with load. Voice calls and USSD were essential for feature phone users, and WhatsApp and web for smartphone users.
The assistant had to work in 16 languages across the continent, with the same quality in voice and text. When a user switched language mid-conversation, the context could not be lost.
In a capital markets setting, the system could inform and direct; it could not give investment advice, comment on prices, take payments or store raw identity numbers (BVN/NIN). These rules had to be guaranteed in code, not left to a prompt. With fake channels widespread, a single verifiable official voice and an escalation protocol for fraud scenarios were also essential.
Dozens of banks and fintechs had to reach the same assistant from their own channels and see only their own customers' conversations. Telephony and channel data had to stay in the country and comply with Nigeria's data protection law, NDPA 2023. All of this was expected to go live with 16 languages and 4 channels on a very short timeline.
Every channel uses the same engine, the same knowledge base, the same persona and the same set of guardrails. The voice channel runs over a toll-free line, a paid line, an international line and a mobile operator short code; streaming speech recognition and synthesis let users interrupt the assistant (barge-in). WhatsApp offers chat and voice calls through a verified business account, the web has a widget embedded in the official microsite and partner sites, and USSD serves users without smartphones. A custom voice identity was created for the assistant's official English voice.
English, Nigerian Pidgin, Hausa, Yoruba, Igbo, Swahili, Arabic, French, Portuguese, Fula, Oromo, Amharic, Zulu, Somali, Afrikaans and Shona are supported in voice and text. When the language changes, session memory keeps details such as names, references and intent. The 15 languages other than English went live after approval by native-speaker reviewers appointed by the client.
Requests for investment advice or price opinions are refused by rules before the language model generates an answer. Knowledge base layers for before and after regulatory approval switch on with a single toggle. Raw BVN/NIN never enters the conversation store. If a user tries to share an ID, PIN or OTP, a standard security warning is triggered.
Each bank and fintech is a separate tenant with its own widget, its own agents and access only to its own customers' conversations. The assistant recognises the user's bank or broker and hands the conversation to that partner's agent; if no partner is named, it routes to the central team. On handover to a live agent, the context, intent, transcript and an AI summary travel with the conversation.
The programme team and partners follow conversations on real-time monitoring screens: conversation volume by channel and language, the most frequent topics, handovers to agents and security warnings appear on a single dashboard. The reporting and analytics layer brings data from every channel into one place, and each partner sees only its own customers' data. This visibility means capacity and knowledge base decisions on peak days are made with data.
The programme includes a programme management partner, a registrar and operations partner, and a telephony and contact centre partner that provides the phone infrastructure and the live agent team. CBOT is responsible for the AI layer of investor communication and runs the platform as a managed service.
Representative chat scenarios in English, Hausa, Nigerian Pidgin and Yoruba: applying, the investment advice boundary and a fraud warning.
- CLIENT
- Dangote Petroleum Refinery
- Nigeria
- USE CASE
- Official investor communication platform
- Answering questions
- Routing to institutions
- Fraud warnings
- Handover to people
- TYPE
- Voice and Text AI Agent
- Managed Service
- CHANNELS
- Voice calls (toll-free, paid, international line, short code)
- USSD
- WhatsApp (chat and voice)
- Web chat
- LANGUAGES
- 16 languages, voice and text
- AI
- LLM-agnostic engine: CBOT LLM, OpenAI, Azure OpenAI, Anthropic, Llama, Mistral
- Smart routing and automatic fallback
- RAG
- Agentic workflow layer
- PARTNER MODEL
- Multi-tenant architecture
- Separate tenant and knowledge layer per partner
- Co-branding, one recognisable assistant
- INFRASTRUCTURE
- Microservices on Kubernetes
- Autoscaling per service
- Automatic failover
- SECURITY AND COMPLIANCE
- ISO 27001 certified ISMS
- Masking, role-based access, audit logging
- NDPA 2023 compliant data processing
- Telephony and channel data in country
- DESIGNED CAPACITY
- 10,000 voice and 250,000 text conversations at the same time
- More than 500,000 calls in 48-hour peak periods
- Up to 5 million voice conversations and 15 million minutes in a quarter (high scenario)
- New capacity in under 5 minutes
- Target latency: first response chunk in 800 ms, full answer within 1.5 seconds
- Elastic scaling validated by load testing; figures are designed capacity, not actual usage
- TIMELINE
- August 2026 kickoff
- 14 September 2026 go-live
This project showed that in a regulated, high-risk environment, what AI must not do can also be written into code. From feature phones to WhatsApp, from Hausa to Portuguese, every user could reach the same official voice.
The remaining partner integrations were planned to go live in stages after launch day.
More case studies
Every organization's needs are different, but the impact of a well-designed AI experience is shared: faster processes, easier service and stronger customer engagement. Explore the other success stories we have brought to life across different industries.
