Prosus used Plus One in Slack to let colleagues test generative AI across writing, coding, transcription, image generation, document review, and data enrichment.
2
Human feedback, retrieval from internal knowledge bases, prompt changes, and model selection reduced Plus One's reported hallucination rate from about one in ten answers to below 2%.
3
Generative AI can improve productivity today, but production systems still face cost, reliability, information freshness, retrieval, and agent brittleness problems.
Summary
Paul van der Boor describes how Prosus explored generative AI across a group of technology companies serving about two billion users. Its Plus One tool gave thousands of colleagues access to multiple models through Slack, where they tested tasks such as coding, writing, transcription, document summarisation, image generation, and data enrichment. The resulting usage showed productivity gains and helped teams discover practical applications, including extracting ingredients and allergens from food listings. Paul explains how Prosus used feedback, prompting, model selection, and retrieval from internal sources such as Confluence to reduce hallucinations. He also discusses the economics of serving users at scale, including model routing, token costs for non-English languages, and the choice between large language models and classical models. His assessment of agents is cautious. They can produce impressive results, but they remain sensitive to prompts and model updates. Human review remains important for user-facing applications.
Plus One gave thousands of colleagues a shared way to test generative AI
Prosus built Plus One as a virtual team member available mainly through Slack. Colleagues could ask it to generate images, transcribe audio, review code, summarise internal Google Docs, process PDFs, and answer other requests through one interface. More than half a million questions had been asked in roughly a year. The goal was to discover which tasks were useful before investing in production systems. Paul describes the tool as a practical way for people across the group to experiment with different models without needing to choose or operate each model themselves.
Generative AI helped people work faster and take on unfamiliar tasks
Users reported that these tools helped them work faster, but Paul also describes a second effect: people could do more kinds of work. A finance analyst could write Python scripts with help debugging problems. Marketing staff could ask for more complicated Excel formulas. Engineers used the tools to document code and explain the likely user impact of a pull request. The usage was roughly split between engineering and non-engineering work. Paul says the tools also helped people get started instead of waiting for an answer or searching unfamiliar documentation.
Colleagues discovered business applications through ordinary experiments
One food delivery example began with a question in Portuguese about the ingredients and quantities for one serving of chicken stroganoff. The follow-up request asked Plus One to structure the answer as JSON so it could be applied across millions of menu items. This suggested a way to enrich dishes with ingredients, dietary properties, and related information. Previously, the company might have collected a dataset and trained a smaller model for the task. A general model could perform the first experiment without task-specific training, helping teams find useful applications through repeated exploration.
Feedback and retrieval reduced hallucinations, but the work remained context-specific
Plus One included thumbs-up, thumbs-down, heart, and Pinocchio reactions. The Pinocchio reaction marked answers that appeared to be made up. Paul says the reported hallucination rate fell from about one in ten answers to below 2% by June. Prosus used feedback to identify problem types, then applied prompting changes and connected models to internal knowledge bases such as Confluence. Paul is direct about the difficulty of retrieval-augmented generation. Embeddings, hybrid search, chunk size, and the handling of code versus documentation all depend heavily on the use case.
Production economics depend on model choice, routing, and language
Prosus had to ask whether a system worked and whether it could afford to run at the required scale. Cost reductions came from choosing smaller models, fine-tuning where appropriate, improving prompts and answer generation, and selecting between large- and small-context models. Paul also points out that tokenisation can make non-English use more expensive. In his example, a Spanish translation required about twice as many tokens as the corresponding English case, while Python used more tokens than the same number of English characters. These differences matter across a group operating in many languages.
Plus One included a data-analysis agent that accepted an Excel file, used a Python execution environment, generated a chart, and performed analysis. Paul says agents can create a remarkable user experience when they work. They are also sensitive to model updates and prompt changes. Errors compound when the system performs several executions, and the non-deterministic behaviour can produce different answers to the same request. Prosus considered agents promising, but Paul described them as early technology that was not yet consistent enough for many production use cases.
Productivity is available now, while novel product impact still needs proof
Paul says teams using tools such as GitHub Copilot, legal applications, or Plus One consistently report higher productivity. He contrasts this with flashy experiments that attract attention but may not solve a real business problem. Data enrichment is an example of a less exciting application with clear value because a model can classify or describe content without the earlier requirement to collect a dedicated dataset and fine-tune a model. A proof of concept is easy in many cases. Shipping it cost-effectively and showing an effect on product metrics is harder.
Human review remains part of the system for user-facing applications
Paul recommends experimenting with prompts and model selection before fine-tuning and hosting a model. He also says retrieval-augmented generation is difficult to generalise across code, documentation, books, papers, and other formats. Models still lack current or context-specific information, behave non-deterministically, and have capabilities that teams do not fully understand. For education and other user-facing systems, human review checks quality against existing processes. Plus One also places AI in a shared team setting, where colleagues can inspect answers and provide an additional layer of social quality control.
"One of the conclusions we were able to draw is that it's able to improve productivity the moment you make this available to yourself or start using it or your teams."08:38
Who should watch
You are deciding which generative AI experiments should move from an internal demo into a product used at scale.
Your team is building a Slack-based assistant or retrieval system and needs practical lessons about feedback, hallucinations, and source attribution.
You need to estimate the cost and reliability of language-model features across multiple languages and large user populations.
