Prosus measures AI adoption through trial users, usage frequency, super users, total questions, API calls, and the number of business use cases using its platform.
2
Sean Kenny says enterprise adoption depends heavily on leadership, clear expectations, hands-on training, and a culture where employees can share what they learn.
3
The proposed AI workforce is built around narrow jobs to be done, connected systems, and domain experts who design and monitor the workflows.
Summary
Paul van der Boor and Sean Kenny explain why enterprise AI adoption takes more than making a model available. Prosus started with language models in Slack, which became Toqan, a productivity tool and platform used across a group of about 100 companies. The team tracks whether people try the tool, become daily or power users, and apply it in real business workflows. Sean describes adoption as a cultural and educational problem as much as a product problem. Users need clear examples, reliable first experiences, live training, and permission to spend time learning. The discussion then moves toward an AI workforce built from narrow jobs to be done, integrations, and configurable agents. Sean is careful about the limits. Agents still need better memory, context, reliability, sharing, evaluations, and observability. Existing company data is often too messy to crawl blindly, and many human processes must be redesigned before agents can perform them consistently.
Prosus treats AI adoption as both workforce learning and business use-case adoption
Paul van der Boor divides Prosus's product metrics into two groups. The first measures whether employees try AI, how often they use it, and whether they become super users who ask more than five questions a day. The second measures AI in real workflows through API calls and the number of individual use cases using the Toqan engine. Prosus includes about 100 companies, ranging from large businesses such as OLX, Swiggy, iFood, and Delivery Hero to smaller companies. Paul says the goal is for more than 80% of employees in each company to try Toqan, often by asking a simple work question such as how to summarize or translate something.
Frequent use matters because employees need AI intuition, even when productivity is hard to quantify
Paul says productivity gains are difficult to measure precisely because results depend on the type of engineer and the task. Prosus is convinced the value is real and in the double digits, but it cares less about whether the result is 21%, 23%, 41%, or 42%. Employees also need practical intuition about what AI can do. Lawyers, finance teams, HR, and marketers who understand the technology can contribute more effectively to future products and services. Paul describes this as a culture change alongside productivity improvement. Prosus tracks tool use where it can, including AI-assisted pull requests, active Cursor or Copilot users, surveys, and controlled experiments.
AI use still carries a stigma because people fear that generated work looks like taking a shortcut
Sean Kenny says that when generative AI started appearing, people often hid the fact that they used it. Employees could feel that an AI-written pull request or document was the cheap way out, even when the result was acceptable. Demetrios Brinkmann connects this concern to the doorman paradox, where removing a visible service can remove guidance and trust that people only notice later. They discuss marketing content that is fully automated but damages a brand because nobody checks whether people will read it. Sean's advice is to use AI as a buddy during important work, discuss how to approach a task, and then review the generated result rather than blindly accepting it.
Toqan evolved from a Slack experiment into an agent platform through repeated product questions
The original Toqan team created a Slack app before GPT-3 was widely available because it was difficult to know where language models would fit in business. The app let employees interact with models and observe what they tried. Early versions used intent detection to route requests into four or five types of work. The team later added file handling, file generation, and data analysis, then changed the architecture to an agent with tools. Sean says the agent approach outperformed the earlier intent-routing system and was easier to expand because new capabilities could be added as tools instead of requiring new intent categories and model training.
Leadership expectations and peer learning determine whether companies move past experimentation
Sean found that some portfolio companies adopted Toqan much faster than others. The difference was often a top-down cultural attitude that said employees should try the technology, tolerate the initial time cost, and learn how to use it. Saying that AI is important without policies, guidance, or resources leaves employees uncertain. Teams that experiment quickly and decide whether something works tend to benefit more. Sean also says discovery is technically difficult and agents were not very user-friendly in their early versions. A sharing community prevents every employee from repeating the same mistakes. The team used webinars, hackathons, workshops, training, and a Slack channel where users could ask questions and compare approaches.
Successful onboarding starts with a reliable first win instead of an empty 'ask anything' box
Sean criticizes generic chatbot interfaces that tell users to ask anything. A user may ask the first question that comes to mind, receive an answer the system cannot actually produce, and conclude that the product is useless. Toqan added follow-up questions and a more usable web interface, while the team relied heavily on live support. In one webinar for roughly 200 people, an onboarding problem meant attendees could not access the product. Sean completed the exercise himself, narrating how he was thinking through the task. It became one of the best-received sessions and led to high uptake afterward. He also describes onboarding checklists with useful starter tasks, such as summarizing an article, so users experience a success before applying the tool to their own work.
Power users need advanced capabilities while the wider workforce needs a low barrier to entry
Sean says power users matter because they discover new use cases, give detailed feedback, and can teach other employees. If the product stops advancing, these users may leave. At the same time, the team needs the minimum usability required to get less experienced users started. Sean describes the product cycle as first asking whether a capability works at all, then allowing domain experts to configure it, and only later making it easy and intuitive. A text-to-SQL agent illustrates this progression. The first version proved feasibility. The next challenge was letting data engineers, analysts, and commercial users set it up for their own environments instead of relying on the product team.
The AI workforce idea breaks work into narrow tasks that move information between systems
Sean's future direction is an AI workforce in which employees can build agents for specific jobs to be done. He argues that most work involves creating, transforming, or moving information between systems. User research can produce notes and insights that move into a product backlog. Marketing teams analyze campaign data and turn it into reports, recommendations, or updates to advertising systems. He describes three components: existing systems such as Salesforce, Looker, code bases, and GitHub; agents that complete tasks between those systems; and domain experts who design, configure, and monitor the resulting workflows. The term AI employee is mainly a useful description for a task-focused agent, since actual employees handle much broader and more complex responsibilities.
Agents need scoped integrations, learned parameters, and redesigned processes before they can work reliably
Sean compares the AI workforce idea with Zapier but says ordinary integrations are fragile and often require detailed knowledge of each system. Agents should expose useful actions such as writing an email or updating a Looker dashboard while hiding low-level endpoints. Toqan is grouping many GitHub endpoints into skills, such as managing pull requests, so users do not face an overwhelming list. Sean says the system also needs memory for repeated parameters, such as knowing a manager's name and email. Broad crawling of Google Drive, Confluence, Jira, or Salesforce would create too much noise because company data is messy. The deeper problem is that many human workflows are informal and chaotic. Agents may require teams to define and improve those processes instead of copying them exactly.
