Konstantin Katsev: The Next Phase of AI Will Be Operational, Not Conversational

The gap between experimentation and value

AI adoption is already widespread, but operational deployment remains much less common.

According to McKinsey’s 2025 global survey, 62% of respondents said their organisations were at least experimenting with AI agents. Only 23% reported that they were scaling an agentic system in at least one business function.

That difference matters. Experimentation demonstrates technical possibility. Scaling requires a system to work inside the organisation’s actual processes, permissions, data and accountability structures.

Research from the Stanford Digital Economy Lab reached a similar conclusion after examining 51 enterprise AI deployments. The decisive difference between outcomes was not simply the underlying model. It was organisational readiness, workflow design, leadership and the willingness to change existing processes.

The bottleneck is therefore moving away from model access. The real challenge is turning intelligence into repeatable execution.

What makes AI operational?

An operational AI system should pass five tests.

1. It must work inside a specific workflow.

   A general assistant may be useful, but a business system needs to understand where a task begins, which actions are permitted and what constitutes completion.

2. The result must be measurable.

   The relevant unit may be time saved, procurement savings, freight utilisation, qualified opportunities, faster resolution or lower operational cost. If the result cannot be connected to a business outcome, the system remains difficult to evaluate.

3. It must handle exceptions.

   Real organisations do not operate through perfect sequences. Information is incomplete, suppliers miss deadlines, customers change requirements and approvals are delayed. Operational systems need defined escalation and human handoff mechanisms.

4. The actions must be auditable.

   As AI moves from producing text to taking action, companies need to understand what happened, which information was used and who remains accountable for the result.

5. The system should improve through operational context.

   The durable advantage is unlikely to come from access to the same foundation model that everyone else can use. It will come from workflow knowledge, proprietary context, integrations and accumulated feedback.

From software that advises to software that executes

This shift can already be seen across several categories represented in Veyra Capital’s portfolio.

In procurement, Zinit is developing AI agents for defined stages of the sourcing process, including supplier discovery, eRFx preparation, bid evaluation, negotiation and contract drafting. The value proposition is not a better procurement chatbot. It is the execution of procurement work within a structured process.

In go-to-market operations, Explee applies AI to researching markets, refining an ideal customer profile, identifying relevant prospects, preparing personalised communication and supporting the process of booking meetings. Again, the important transition is from generating a sales message to helping operate a complete workflow.

The same pattern extends beyond pure software. Fura combines technology with the operational requirements of freight and logistics. Dwelly combines acquisitions and succession planning for UK letting agencies with the modernisation of post-acquisition operations. Tyred connects a digital service layer with the physical delivery of mobile bicycle repairs across London.

These businesses operate in different markets, but they illustrate a common principle: technology becomes more valuable when it is attached to a real operational outcome.

Three transitions investors should watch

The first transition is from interface to workflow.

A conversational interface can make software easier to use. But ease of use alone does not create a durable company. The stronger opportunity lies in owning or improving a workflow that customers repeatedly depend on.

The second is from model advantage to system advantage.

Model performance will continue to improve, but access to capable models is becoming broadly available. Defensibility will increasingly be created through industry knowledge, proprietary data, integrations, distribution, customer trust and the ability to operate reliably in difficult environments.

The third is from adoption to accountability.

It is easy to measure how many employees have access to an AI tool. It is harder, but more important, to measure what the system completed, whether the outcome was correct and how responsibility was allocated between software and people.

This is why operational AI does not mean removing humans from every process. In many important workflows, the winning model will be bounded autonomy: AI executes defined tasks, while people establish objectives, review exceptions and remain responsible for consequential decisions.

What founders should build around

For founders, the opportunity is not simply to add an AI layer to an existing category. It is to identify a workflow where intelligence can change the economics or quality of execution.

The strongest opportunities are likely to have several characteristics:

• A recurring and economically important workflow.
• Fragmented information or significant manual coordination.
• A measurable outcome that matters to the customer.
• Access to context that improves the system over time.
• Clear rules for permissions, review and accountability.
• A distribution advantage or direct route into the workflow.

The next generation of AI companies may therefore look less like standalone technology demonstrations and more like deeply integrated operating businesses.

Some will sell software. Others will combine software with services, transactions, acquisitions or physical infrastructure. What connects them will be their ability to turn intelligence into reliable action.

Conversation made AI accessible. Operations will determine where it creates durable value.