Prosus built Toan as a general, agent-based assistant so employees could discover useful AI applications through everyday work.
2
Adding agents reduced hallucinations, expanded usage beyond engineering, and brought the running cost to about 25 cents per interaction.
3
OLX Magic worked better after Prosus moved away from a blank chat interface and built an e-commerce-specific agent around search, personalization, and product data.
Summary
Euro Beinat and Paul van der Boor describe two AI-agent systems built at Prosus. Toan is an internal assistant available through Slack, the web, and APIs. It connects to company data and lets employees ask questions in ordinary language. Prosus introduced agents in early 2024, which helped lower hallucinations and made the assistant useful to more non-engineers. The system reached more than 15,000 users and close to one million requests per month. Euro says the main benefit is not only time saved. Employees work more independently, try tasks outside their usual skills, and get past blocks. At iFood, an English-language data agent reduced time to insight by 75 percent. Paul then explains OLX Magic, an agent-based shopping experience that suggests filters, compares products, and uses catalog and web search. The team found that e-commerce agents need their own tools, prompts, evaluations, memory, guardrails, and search pipelines.
Prosus gave employees a general assistant so useful applications could emerge from practice
Prosus released an AI assistant across the group so colleagues could experiment with generative AI and find uses the central team could not predict. Euro says early field tests with GPT-3 produced unexpected solutions, so the company chose a collective discovery process. Toan is available in Slack, on the web, and through APIs. It is designed for broad use, connects to company databases, and can be composed into applications used by companies in the group. The aim is to let employees decide where the tool creates value while also finding consumer-facing use cases.
Agents made the assistant more reliable by breaking work into actions and checking the results
Euro describes a stack with an agentic framework, orchestration, tools, and a router. The framework breaks a request into pieces that can run in parallel, selects tools, and repeats an action and reflection loop. A separate insights model groups conversations by subjects such as learning or programming language, without inspecting the conversation content itself, so the team can improve later versions. In Toan's data explorer, the agent identifies tables and columns, creates and runs a query, checks for errors, revises its plan when needed, and can produce a diagram.
Toan became useful to non-engineers as hallucinations fell and usage expanded
In October 2022, about 10 percent of Toan's answers triggered the team's signal that the assistant was making things up. That rate later fell to about 1 percent. Euro attributes the change to better models, users learning where the tool works, and the addition of agents and guardrails. After agents were introduced, engineering usage fell while non-engineering usage rose. The assistant was used by 24 companies, more than 15,000 people, and approached one million requests per month. The usage pattern became closer to the technical and non-technical mix of the wider organization.
Agent interactions became more expensive in ways that token prices alone hide
Euro says the cost of model tokens fell by 98 percent from the early design period. However, agent interactions make more calls to language models, and users ask more complex questions as the tool improves. Over a four-month period, the cost per token fell by about 50 percent while tokens per question increased. The resulting cost per interaction settled around 25 cents. Euro presents this measure as more useful for adoption decisions than the falling price of an individual token, because it includes the complete agent process.
The main benefit for employees is independence and broader capability
Toan estimates saved time by comparing an interaction with the assistant to the best alternative without it. Euro reports about 48 to 50 minutes saved per day, but says users consider other effects more important. They can answer questions without asking a colleague, work in programming languages they do not know well, and get started when they are blocked. The time savings are spread across many small bursts, which are difficult to turn into conventional automation. Euro describes the combined effect as making the organization somewhat more senior.
An English-language data agent moved answers from analysts to the people asking the questions
At iFood, Prosus gave employees a way to ask the data lake questions in English. The agent automated about 21 percent of requests. The larger change came when people in customer support, operations, and marketing answered their own data questions instead of sending them to data analysts. This cleared a large backlog and removed about 109 analyst days each month, according to Euro's figures. The result was a 75 percent reduction in time to insight. The lesson was that organizations have to change how work is assigned, rather than only making the existing process faster.
OLX Magic combines familiar shopping patterns with agent assistance
Paul presents OLX Magic, an agent-based experience for the OLX marketplace. A search for an espresso machine can produce suggested filters, custom carousels, contextual product attributes, and a smart comparison between selected items. Users can describe a requirement such as a milk foamer, and the system turns that description into a search filter. The agent uses the OLX catalog, visual and text search, and the wider web. It also has planning and memory, including context about the user, so the experience can guide product discovery rather than only return a list of results.
An e-commerce agent needs its own search, prompts, evaluations, and guardrails
Paul says the team first tried a blank conversational interface, influenced by the early ChatGPT pattern. Users in a commercial setting expected familiar search behavior, so the product moved through several iterations toward standard e-commerce patterns. The agent's tools, prompts, evaluations, memory, and guardrails had to fit the buying journey. OLX's existing search pipelines were not enough. The team rebuilt retrieval around listing titles, images, and descriptions, then used prompts and language models to filter for items relevant to the user's conversation.
"The best thing we can do is to give the best possible tools to everyone that can test them and can figure out by themselves what it can do for them."04:35
Who should watch
You are deciding whether an internal assistant should be a broad experimentation tool or a narrow application for one team.
Your agent uses company data and you need practical measures for reliability, cost per interaction, and employee impact.
You are building an AI shopping or search product and need to understand why a generic chat interface is not enough.
