Scaling AI Agent Deployment Across a Global E-commerce Organization

prosus · E-commerce · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise-wide agent deployment in global e-commerce

Prosus wanted to move from occasional AI assistants to a trusted AI agentic workforce across a 100-country e-commerce organization, enabling non-engineering employees to create domain-specific agents while managing safety, privacy, integrations, adoption, and evaluation.

Constraints:

• Prosus operates across 100 countries and multiple e-commerce verticals.
• The organization expected tens of thousands of agents, not a small centralized catalog.
• Agents range from document-only assistants to senior agents that orchestrate other agents and write back to systems.
• Safety, privacy, proprietary integrations, MCPs, and cost drove the decision to build internally.
• Agent adoption became an organizational change problem after the initial tooling barrier was solved.

Architecture

Prosus built Toqan, an internal agent builder that lets employees across sales, marketing, HR, logistics, and other departments create agents. The platform supports graded agent seniority, from document-working interns to senior agents with tools, integrations, system credentials, database write-back, web access, and agent orchestration. Adoption is driven through demos, upskilling, competitions, and organizational muscle-building, while evaluation is handled case by case through A/B tests, before/after comparisons, and user-provided impact signals across productivity, quality, agility, and independence.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Prosus employees"] --> B["Toqan agent builder"]
  B --> C["Agent seniority model"]
  B --> D["Tools and integrations"]
  C --> E["Business agents"]
  D --> E
  E --> F["Operational workflows"]
  F --> A
  E --> G["Evaluation framework"]
  G --> H["Adoption programs"]
  H --> A

Components

Data flow

Employees identify jobs to be done, create agents in Toqan, connect them to allowed tools, integrations, MCPs, documents, and systems of record, then use them in operational workflows. More senior agents can access systems, write data back, use multiple tools, and orchestrate other agents. Results are measured through A/B tests where possible, before/after comparisons, and user feedback across productivity, quality, agility, and independence.

Why this architecture

Prosus built an internal agent builder because it needed safety, privacy, proprietary integrations, MCP support, use-case optimization, and cost control. The platform had to be accessible to non-engineers, because reaching 30,000 agents required employees across the organization to create and iterate their own agents.

Models used: AI agents · Large commerce models · MCP integrations

Scale

• Prosus operates in 100 countries across South America, Europe, and Asia.
  “We are present in 100 countries across the globe, South America, Europe and Asia predominantly”transcript, stated
• Prosus expected 30,000 agents in production across the group by the end of March.
  “it's exactly the number of agents that we expect to have in production across Prozus by the end of March.”transcript, stated
• By October, 8,000 agents were used weekly and about 20,000 had been created.
  “now in October we had about 8,000 of these agents in production which are used every week but about 20,000 of these agents in total have been created.”transcript, stated
• Prosus envisions every employee having between one and 20 agents.
  “every employee globally will have one two 10 maybe 20 of these agents helping them to do better”transcript, stated
• A restaurant account executive agent was used by about 200 colleagues and did the work of about 30 full-time employees.
  “In this case it's used by about 200 of our colleagues at this moment and it does the work of about 30 full-time employees.”transcript, stated

Alternatives considered (roads not taken)

Use off-the-shelf tools at scale

Rejected because: Prosus did not find tools that met its scale, safety, privacy, integration, MCP, optimization, and cost requirements.

Let engineers create all agents

Rejected because: The goal was for colleagues across functions to create agents, so Prosus had to make agent creation feel ordinary and accessible.

Evaluate every agent with one universal metric

Rejected because: Impact is case-specific, so Prosus uses A/B tests where possible and user-provided before/after evidence elsewhere.

Patterns

tool-use-orchestration (orchestration)

Prosus agents gain capability through tools and integrations across systems of record, help desks, email, calendars, company systems, and the web.

multi-agent-orchestration (orchestration)

Senior agents can orchestrate other agents and handle more complex workflows.

safety-policy-enforcement (governance)

Prosus built Toqan internally to guarantee safety, privacy, and controlled integration with internal systems and MCPs.

eval-driven-development (evaluation)

Prosus evaluates agents through A/B testing, before/after user information, and four impact dimensions: productivity, quality, agility, and independence.

human-in-the-loop-gate (adoption)

Prosus scales adoption through human-led demos, upskilling, competitions, and employee-created agents rather than only technical deployment.

Tradeoffs

Build versus buy

Chose internal Toqan agent builder over external tools.

Rationale: Prosus needed scale, safety, privacy, proprietary integrations, MCPs, optimization, and cost effectiveness.

Capability progression

Chose graded agent seniority over one flat class of agents.

Rationale: Different agents need different tool access, system access, write-back permissions, and orchestration power.

Evaluation approach

Chose case-specific A/B, before/after, and user-supplied evidence over one universal evaluation method.

Rationale: Agent impact varies by use case, so implementation of the four impact dimensions is case specific.

Gotchas (what bit them)

Agent adoption became organizational, not technical.

Root cause: Non-engineers believed agent creation was something only software people could do.

Resolution: Prosus demonstrated creation with sales and HR teams, made agents feel ordinary, and invested in upskilling and competitions.

Higher-seniority agents create higher trust and pollution risk.

Root cause: Agents with credentials and database write-back can affect systems, not just read documents.

Resolution: Grade agents by seniority and tool/integration access so capability and risk are visible.

Internal/external system mixing creates new risks.

Root cause: Agents can access internal systems and the web, creating boundary and data-flow concerns.

Resolution: The speaker flags this as an area requiring discussion and controlled integration design.

Evaluation

Prosus evaluates agents through a developing framework with A/B tests where possible, before/after comparisons and user information elsewhere, and four impact dimensions: productivity, quality, agility, and independence.

Metrics: 30,000 agents expected in production by end of March · 8,000 weekly used agents and 20,000 total created by October · Restaurant account executive used by about 200 colleagues · Restaurant account executive does work of about 30 full-time employees · Impact dimensions: productivity, quality, agility, independence

Outcomes

• Prosus built an internal agent builder, Toqan.
• By October, employees had created about 20,000 agents and 8,000 were used weekly.
• The company expected 30,000 agents in production by the end of March.
• A restaurant account executive agent did work equivalent to about 30 FTEs for about 200 colleagues.
• Agent creation expanded beyond engineers to sales, marketing, HR, logistics, and other functions.
• Adoption programs such as competitions and upskilling helped create organizational AI muscle.

Prosus reports 100-country operations, 30,000 expected production agents, 8,000 weekly active agents, 20,000 total created agents, and one agent doing about 30 FTEs of work.