By Thales & Claude -- CEO & AI CTO, ZeroSuite, Inc.
There is a chartered accountant in Dakar who needs to close the books for a mid-size company operating under SYSCOHADA. She needs to classify an unusual transaction, determine the correct Plan Comptable OHADA account, and generate the journal entry -- all before the quarterly filing deadline in three days.
Her options: consult a colleague (who is equally busy), search through the 1,200-page OHADA Uniform Act on Accounting (which she has read, but not memorized in its entirety), or hire a consultant at 150,000 FCFA per day. None of these options scale. None of them are available at 11 PM when she is finishing the books.
Deblo Pro exists to be her fourth option: an AI advisor that speaks French, understands SYSCOHADA, and costs a fraction of a consultant's fee. But "an AI that knows accounting" is not enough. Professionals work in specific domains with specific vocabularies, specific legal frameworks, and specific expectations. A generic chatbot that conflates IFRS with SYSCOHADA or confuses French tax law with Ivorian tax law is worse than useless -- it is dangerous.
So we built 101 of them. One hundred and one specialized AI advisors, organized across 14 professional categories, each with a distinct identity, domain knowledge, and behavioral profile. This is how.
---
The Problem with Generic AI Assistants
Every major LLM provider offers a general-purpose assistant. ChatGPT, Gemini, Claude -- they are all brilliant generalists. But generalism fails professional users in three specific ways.
First, context switching. A professional's questions exist within a domain. When a Senegalese auditor asks "What are the requirements for the commissaire aux comptes report?", the answer depends on OHADA's Uniform Act on Commercial Companies, the national auditing standards of Senegal, and the specific entity type. A generic AI does not know which framework to apply and often defaults to French (metropolitan) standards or IFRS, both of which are wrong for this context.
Second, tone and register. African professionals -- particularly in francophone countries -- expect formal French with vouvoiement (the formal "you"), proper legal citations, and structured deliverables. They do not want casual conversation. They want a colleague who produces artifacts: memos, journal entries, compliance checklists, and legal analyses.
Third, trust calibration. A generic AI that confidently applies French tax law to a question about Ivorian tax law is not just unhelpful -- it is a liability. Professional advisors must know what they do not know and cite their sources precisely.
---
The Agent Architecture
Each Deblo Pro advisor is defined as a ProAgent in the frontend catalog. The TypeScript interface captures everything the system needs to present, select, and activate an agent:
export interface ProAgent {
id: string;
nom: string;
fonction: string;
systemPromptSlug: string;
icon: any;
category: string;
accroche: string; // Hook -- one-line pitch
description: string; // Full description
usages: string[]; // 3-5 example use cases
probleme: string; // Problem statement
solution: string; // How this agent solves it
isActive: boolean;
tags?: string[];
}The systemPromptSlug is the critical field. When a user selects an agent, this slug is sent to the backend and mapped to a specific set of behavioral instructions that are injected into the system prompt alongside the base Pro prompt.
The agent catalog lives in a single file: agents.ts, at 1,865 lines. It is the largest single file in the Deblo frontend, and for good reason -- each agent requires a complete identity definition with name, function, hook, description, five usage examples, a problem statement, and a solution statement. This verbosity is intentional. The agent selection UI presents all of this information to help the user choose the right advisor.
---
14 Categories, 60+ Agents
The agents are organized into 14 professional categories:
export const PRO_AGENT_CATEGORIES: ProAgentCategory[] = [
{ id: "admin", name: "Administration", color: "amber" },
{ id: "finance", name: "Finance", color: "brand-green" },
{ id: "commerce", name: "Vente & Marketing", color: "orange" },
{ id: "tech", name: "Technique & IT", color: "indigo" },
{ id: "legal", name: "Juridique & Qualite", color: "purple" },
{ id: "operations", name: "Operations", color: "teal" },
{ id: "rnd", name: "R&D & Innovation", color: "red" },
{ id: "education", name: "Formation & Education", color: "blue" },
{ id: "media", name: "Medias & Creation", color: "pink" },
{ id: "sante", name: "Sante & Bien-etre", color: "emerald" },
{ id: "agri", name: "Agriculture & Ressources", color: "lime" },
{ id: "social", name: "Developpement Social", color: "cyan" },
{ id: "telecom", name: "Telecommunications", color: "violet" },
{ id: "gouvernement", name: "Gouvernement", color: "rose" }
];Here is what a typical agent definition looks like -- the accountant advisor, one of the most-used agents:
{
id: "comptable",
nom: "Conseiller Comptable",
fonction: "Comptable / Responsable financier",
systemPromptSlug: "sp_comptable_responsable_financier",
icon: CreditCard,
category: "finance",
accroche: "Accelerez vos analyses financieres et saisissez l'essentiel des chiffres !",
description: "Un allie pour vos clotures comptables, analyses budgetaires et rapports "
+ "financiers qui vous fait gagner un temps precieux.",
usages: [
"Automatisation de rapports financiers reguliers",
"Analyse et justification des ecarts budgetaires",
"Preparation de documents fiscaux et declaratifs",
"Elaboration de tableaux de bord financiers",
"Verification de la conformite des ecritures comptables"
],
probleme: "Vous passez trop de temps sur les analyses financieres et "
+ "la preparation de rapports comptables ?",
solution: "Consultez le Conseiller Comptable pour obtenir rapidement des analyses, "
+ "des ecritures et des rapports conformes aux normes SYSCOHADA.",
isActive: true
}The Finance category alone includes: Comptable (general accounting), Fiscaliste (tax advisory), Controleur de Gestion (management control), Analyste Financier (financial analysis), Auditeur Interne (internal audit), Banquier (banking and financial advisory), Analyste Credit (credit risk), Recouvrement (debt recovery), Microfinance (microfinance institutions), Inclusion Financiere (financial inclusion), and Mobile Money (mobile financial services). Eleven agents in a single category, each with distinct expertise.
---
The Backend: Agent-Aware System Prompts
When a user selects an agent and sends a message, the backend receives the agent ID in the chat request. The agent ID is mapped to a function label through PRO_AGENT_FUNCTIONS:
PRO_AGENT_FUNCTIONS: dict[str, str] = {
# --- Administration & Direction ---
"dg": "Directeur General / CEO",
"secretaire": "Secretaire / Assistant administratif",
"assistantegestion": "Assistante de Gestion / Office Manager",
"planificateur": "Planificateur / Organisateur",# --- Finance, Comptabilite & Audit --- "comptable": "Comptable / Responsable financier", "fiscaliste": "Conseiller fiscal / Comptable", "controleur": "Controleur de Gestion", "analystefin": "Analyste financier / Controleur de gestion", "auditeur": "Auditeur Interne", "banquier": "Banquier / Conseiller financier", "creditanalyste": "Analyste credit / Charge de risque", "recouvrement": "Charge de recouvrement / Contentieux", "microfinance": "Agent de microfinance / IMF", "inclusion-financiere": "Specialiste en Inclusion Financiere", "mobilemoney": "Agent de services financiers mobiles",
# --- Juridique & Conformite --- "juriste": "Juriste / Conseiller legal", "compliance": "Responsable Conformite / Compliance", # ... 40+ more entries } ```
The selected agent's function label is injected into the system prompt, narrowing the LLM's behavioral scope. Instead of being "Deblo, a general professional assistant," the model becomes "Deblo, acting as a Comptable / Responsable financier." This single injection changes the model's response style, vocabulary, and reasoning priorities.
The base Pro system prompt, at version 7.0, is structured as XML for strict model compliance:
<deblo_pro_system>```
Several things are notable about this prompt. It mandates vouvoiement -- the formal "vous" -- which is non-negotiable in a professional French context. It prohibits generic introductions ("Bien sur !", "Excellente question !") and closing summaries ("En resume...", "J'espere avoir pu vous aider..."), forcing the model to deliver information directly. And it includes an explicit legal disclaimer for applied tax, legal, and audit advice -- a compliance requirement that protects both the user and ZeroSuite.
---
Domain Knowledge: SYSCOHADA, OHADA, and Beyond
The professional domains that Deblo Pro serves are specific to Africa and largely absent from mainstream AI training data. Consider the frameworks:
SYSCOHADA (Systeme Comptable OHADA) -- The unified accounting standard for 17 West and Central African countries. It replaced the Plan Comptable 1957 and OCAM systems. SYSCOHADA mandates the Plan Comptable OHADA, a chart of accounts with specific numbering (Class 1 through Class 8), journal entry formats, and financial statement templates. An LLM trained primarily on IFRS and US GAAP data will default to those frameworks unless explicitly guided otherwise.
OHADA (Organisation pour l'Harmonisation en Afrique du Droit des Affaires) -- The supranational legal framework governing business law across 17 member states. Its Uniform Acts cover commercial companies, securities, insolvency, arbitration, accounting, cooperatives, and labor law. When a Deblo Pro user asks about corporate governance requirements, the answer must reference the Acte Uniforme relatif au droit des societes commerciales, not the French Code de commerce.
National tax codes -- Each OHADA member state has its own tax legislation. Ivory Coast's Code General des Impots (CGI-CI) is different from Senegal's (CGI-SN), which is different from Cameroon's. A tax advisory agent must ask for the user's country before citing tax rates or article numbers.
The domain knowledge overlay system was designed for these requirements. The PRO_DOMAIN_PROMPTS dictionary maps domain slugs to additional system prompt text that is appended when the relevant agent is active. In the current architecture, these overlays are intentionally left empty -- the domain-specific knowledge will be injected via a RAG (Retrieval-Augmented Generation) pipeline once the legal document corpus is indexed. The agent system is the behavioral layer; RAG will be the knowledge layer.
---
The Agent Picker Modal
The user-facing agent selection experience is the AgentPickerModal -- a full-screen overlay with search, category filtering, and favorites.
The filtering logic supports three modes:
1. Search by name or function -- A text input at the top filters agents by matching against both nom and fonction fields, case-insensitive. Typing "comptab" matches "Conseiller Comptable" by name and "Comptable / Responsable financier" by function.
2. Category tabs -- Horizontal scrollable tabs for each of the 14 categories. Tapping "Finance" shows only the 11 finance agents. Tapping "All" resets the filter.
3. Favorites -- Users can star agents they use frequently. Starred agents appear at the top of the list regardless of category filter. Favorites are persisted in localStorage (web) or AsyncStorage (mobile).
Each agent card in the grid displays the icon, name, function, and the accroche (hook). Tapping a card opens a detail panel with the full description, usage examples, and a "Start conversation" button. The detail panel also shows the problem/solution framing, which helps users understand whether this agent is appropriate for their current question.
---
The Credit Model for Pro
Pro conversations cost more than K12 conversations because they use more capable models, generate longer responses, and invoke complex tools (PDF generation, Excel creation, email drafting). The credit model reflects this:
- Standard complexity: 1 credit = 3,000 tokens. Minimum balance: 3 credits.
- Complex requests: 1 credit = 2,000 tokens. Minimum balance: 8 credits. Triggered by file generation, multi-step analysis, or web search.
The organization system allows companies to create credit pools. A firm of five accountants can purchase a bulk credit package and distribute it. Each member has their own account but draws from the organization's pool. The org_memberships table tracks role (admin, member) and the credit allocation logic handles both individual and pooled balances.
This model works for the African professional market because it avoids subscriptions. A chartered accountant who uses Deblo Pro intensively during quarterly closings and barely at all in between should not pay $30/month for 12 months. Pay-as-you-go, with bulk discounts on larger packages, matches the usage pattern.
---
Why 101, and Why It Matters
The number 101 is not arbitrary. It represents a systematic mapping of the African professional landscape. We did not start with "let us build 101 agents." We started with a question: "What are the professional roles that exist in a typical West African economy?" Then we cataloged them.
Administration: directors, secretaries, office managers, planners, facility managers. Finance: accountants, tax advisors, controllers, analysts, auditors, bankers, credit analysts, debt recovery, microfinance, financial inclusion, mobile money. Sales and marketing: sales managers, customer service, product managers, digital marketers, community managers, trade representatives. Technology: developers, system administrators, data analysts, project managers, UX designers, cybersecurity. Legal: lawyers, compliance officers, notaries, IP specialists, labor law, contracts. Operations: logistics, supply chain, procurement, quality, maintenance. R&D: researchers, laboratory managers, innovation managers, engineers. Education: trainers, curriculum designers, assessment specialists. Media: journalists, graphic designers, content creators, video producers, photographers. Health: doctors, nurses, pharmacists, nutritionists, sports coaches, wellness advisors. Agriculture: agronomists, agricultural advisors, climate specialists, water resource managers. Social: social workers, community development, local administrators. Telecom: network engineers, product managers, retail. Government: public servants, local administrators.
Each of these roles has specific vocabulary, specific regulatory context, specific deliverables, and specific expectations for how an AI assistant should behave. A logistics manager does not want accounting journal entries. A tax advisor does not want marketing copy. The agent system ensures that the AI's behavior matches the professional context.
---
The Road to RAG
The current agent system is behavioral -- it shapes how the model responds, but the knowledge comes from the LLM's training data. This has a known limitation: the model's knowledge of SYSCOHADA, OHADA Uniform Acts, and national tax codes is incomplete and occasionally outdated.
The next phase is RAG integration. We are building a document corpus that includes:
- The full text of all OHADA Uniform Acts (9 acts, thousands of articles)
- National tax codes for the 6 initial target countries (Ivory Coast, Senegal, Cameroon, Burkina Faso, Mali, Benin)
- SYSCOHADA Plan Comptable with full account descriptions
- Professional standards for auditing, legal practice, and tax advisory
- Regulatory updates and case law
When RAG is active, a user's question will trigger a semantic search against this corpus. Relevant document chunks will be injected into the prompt alongside the agent's behavioral overlay. The model will then generate responses that cite specific articles, reference specific account numbers, and quote specific legal provisions -- with source attribution.
The 101 agents will become the behavioral interface for a knowledge system that is both deeper and more current than any single LLM's training data. That is the plan. The agents are the first half.
---
This is Part 10 of a 12-part series on building Deblo.ai.
1. AI Tutoring for 250 Million African Students 2. 100 Sessions Later: The Architecture of an AI Education Platform 3. The Agentic Loop: 24 AI Tools in a Single Chat 4. System Prompts That Teach: Anti-Cheating, Socratic Method, and Grade-Level Adaptation 5. WhatsApp OTP and the African Authentication Problem 6. Credits, FCFA, and 6 African Payment Gateways 7. SSE Streaming: Real-Time AI Responses in SvelteKit 8. Voice Calls With AI: Ultravox, LiveKit, and WebRTC 9. Building a React Native K12 App in 7 Days 10. 101 AI Advisors: Professional Intelligence for Africa (you are here) 11. Background Jobs: When AI Takes 30 Minutes to Think 12. From Abidjan to 250 Million: The Deblo.ai Story