Agent-Based AI Assistants for Enterprise and E-commerce Applications

prosus · E-commerce · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.2 (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 and e-commerce assistants

Prosus built internal and consumer AI assistants across a global e-commerce portfolio while managing hallucination, cost, and adoption.

Constraints:

100-country portfolio
Internal and customer-facing use cases
Hallucination control
Cost per interaction

Architecture

Prosus built a portfolio assistant architecture around Tokan for employees and OLX Magic for e-commerce. Users enter through Slack, web, or APIs; an agentic framework decomposes requests into parallel pieces; orchestration and a router select tools; a reflection/action loop executes tasks such as data exploration with query creation, execution, and diagrams; a separate insights model groups usage conversations for continuous learning. OLX Magic applies the same production-agent pattern to marketplace discovery with specialized e-commerce tools, prompts, and evals.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  tokan-assistant[Tokan assistant] --> slack-web-and-apis[Slack web and APIs]
  tokan-assistant --> agentic-framework[Agentic framework]
  agentic-framework --> orchestration-layer[Orchestration layer]
  orchestration-layer --> tool-router[Tool router]
  tool-router --> tool-set[Tool set]
  tool-set --> data-lakes[Data lakes]
  agentic-framework --> insights-model[Insights model]
  olx-magic[OLX Magic] --> tool-set
  olx-magic --> e-commerce-site[E-commerce site]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Tokan assistant`	Internal Prosus employee assistant used across companies.	`Slack web and APIs`, `Agentic framework`
`Slack web and APIs`	Entry surfaces for users and systems.	`Tokan assistant`
`Agentic framework`	Decomposes requests into executable pieces.	`Tokan assistant`, `Orchestration layer`
`Orchestration layer`	Coordinates parallel execution, reflection, and action loops.	`Agentic framework`, `Tool router`
`Tool router`	Selects available tools for the assistant's current task.	`Orchestration layer`, `Tool set`
`Tool set`	Query creator, executor, diagram creation, marketplace tools, and other assistant capabilities.	`Tool router`, `Data lakes`
`Data lakes`	Enterprise data sources used by data-explorer agents.	`Tool set`
`Insights model`	Groups usage conversations into statistics for continuous learning.	`Agentic framework`
`OLX Magic`	Consumer-facing e-commerce assistant built with production agent tooling.	`Tool set`, `E-commerce site`
`E-commerce site`	OLX marketplace surface where discovery and listing guidance are delivered.	`OLX Magic`

Data flow

A user request enters Tokan through Slack, web, or API. The agentic framework unpacks the request, chooses tools through orchestration and routing, executes and reflects over tool results, and logs abstracted usage insights. For data exploration, the assistant connects to data lakes, identifies tables and columns, creates and executes queries, fixes errors, and can generate diagrams.

3 supporting quotes

“you can access this this um uh stack through slack or through the web or through apis it goes to an agentic framework which decides how to unpack what you want to do in pieces that you can execute in parallel it then figures out which tools you can use it goes”transcript, stated

“goes to an agentic framework which decides how to unpack what you want to do in pieces that you can execute in parallel”transcript, stated

“it can use a variety of tools in this case it use Query Creator executor but eventually also the um um the creation of a diagram all this set of tools is part of the normal let's say uh agent application”transcript, stated

Why this architecture

Prosus moved from general LLM access toward agents because tool use, decomposition, and orchestration reduced hallucination and made assistants useful across enterprise and e-commerce tasks. The insights model and cost tracking exist because adoption at portfolio scale depends on continuous learning and total per-interaction cost, not only token prices.

3 supporting quotes

“we introduce agents so agents is one of the key for make these tools function better somewhere around uh uh beginning of this year so January 2024 we introduced agents and when we talk about agents this is the stack that we use the main components are the framework”transcript, stated

“there is also an important element here which is that of insights there is a another model that lives inside this stack and it is designed to create statistics about usage it is not designed to look at what goes on but we create uh to group that the conversation”transcript, stated

“the cost for interaction this moment it lingers around 25 cents per time”transcript, implied

Scale

Tokan was used by 24 companies, with about 20,000 continuous users and nearly one million requests per month.
“used by 24 companies at this moment uh more than 15,000 mentioned before about 20,000 us it continuously and is getting very close to about 1 million request per month”transcript, stated
Hallucination fell from roughly 10% to about 1%.
“this number of hallucination went from 10% to about 1% where it lingers now”transcript, stated
The iFood data-explorer agent automated about 21% of requests and reduced time to insight by 75%.
“that to the data Lake in English obviously you can see that it has an impact in time saving so it automates about 21% of the request meaning that 21% of the request you can do through the agent to the assistant you don't have to go to the data L create the queries and so on so that's the automation”transcript, stated

Alternatives considered (roads not taken)

Expose a simple ChatGPT-for-company assistant

Rejected because: Prosus argues production e-commerce agents require tools, prompts, evals, and domain-specific systems under the hood.

“everybody out there was trying to build a chat GPT for X for ox or for iFood or for you know any other website or online experience you can imagine um which basically starts with a conversation uh blank canvas where you can start and write and ask about for example”transcript, implied

Judge cost by token price alone

Rejected because: Agent workflows used more calls and tokens per question, so interaction cost mattered more.

“50% at the same time the number of questions went up and I think it is the blue line but the tokens per question went up even further so it's the red line here because these tools require many more calls to the llms many more interactions and because the questions of the users because the tool”transcript, stated

Patterns

`multi-agent-orchestration` (assistant-platform)

The case describes assistants using agent applications with multiple tools.

“in this case it use Query Creator executor but eventually also the um um the creation of a diagram all this set of tools is part of the normal let's say uh agent application and we can see this uh becoming the standard of what our assistants um are going to look like at at present it means after th”transcript, stated

`eval-driven-development` (quality)

Hallucination rates were tracked and reduced as agents were introduced.

“Through iterative improvements including the introduction of agents in January 2024, they reduced hallucination rates from 10% to 1%, achieved approximately 48 minutes of time saved per employee per day, and scaled to nearly 1 million requests per month.”summary, stated

Tradeoffs

Watch total interaction cost, not just token price.

Chose Per-interaction cost tracking for agent workflows over Assuming cheaper tokens automatically make agents cheaper to run.

Rationale: Agentic workflows increased calls and tokens per question, offsetting lower model token prices.

“these tools require many more calls to the llms many more interactions and because the questions of the users because the tool Works become more complicated and complex”transcript, stated

How to measure production agent economics.

Chose cost per interaction over token price alone.

Rationale: Prosus saw token prices fall while agent calls and tokens per question rose, making total interaction cost the adoption metric.

“these tools require many more calls to the llms many more interactions”transcript, stated

Gotchas (what bit them)

Early assistants were hard to trust because users could not tell which answers were hallucinated.

Root cause: The assistant produced a nontrivial share of made-up answers in normal use.

Resolution: Prosus tracked hallucination explicitly and iterated the tool until the rate fell from roughly 10% to about 1%.

“October 22 when we measured this it was about 10% it means one in 10 um it's just an answer that you cannot really trust but you don't know which one”transcript, stated

Evaluation

Prosus tracked hallucination rate, time saved, request volume, and cost per interaction.

Metrics: hallucination rate · minutes saved · requests per month · cost per interaction

“The cost per interaction stabilized around 25 cents despite increased complexity, with token costs dropping 98% while tokens per question increased due to agentic workflows.”summary, stated

Outcomes

The summary reports about 48 minutes saved per employee per day and nearly 1M monthly requests.

Agent introduction reduced hallucination rates materially.