Vadym Shashkov: 'Our Job Is To Keep Expanding What AI Agents Can Do on Their Own, While Keeping Them Under Control'

According to McKinsey, 2026 saw agentic AI get tested at scale across workflows, but fewer than 10% of companies managed to deploy it widely enough to see a real economic payoff.

'The main challenge is getting AI to reliably execute an entire chain of actions — without losing data, while performing operations correctly inside enterprise systems, and handing off to a human at the right moment', explains Vadym Shashkov, a specialist in architecting autonomous multi-agent AI systems for regulated industries.

Earlier in his career he built solutions for legaltech and the enterprise sector, and in 2025 he became co-founder and CTO of Superagent AI, a startup that built a platform of nine specialised AI agents covering the entire insurance sales cycle.

We spoke with him about why most AI pilots never make it to real-world deployment, how to give agents more autonomy, and which decisions need to stay with a human.

— Why do AI agents perform well in experiments but rarely get adopted into companies' day-to-day operations? What does it take for a company to trust them with real work?

— The problem is that a prototype is usually tuned for a single scenario and monitored by hand. Real-world work is messier: the agent gets incomplete data, runs into failures and unusual requests, has to talk to several enterprise systems, and takes actions where a mistake could have financial or legal consequences.

A convincing prototype shows that an agent can perform a task in principle. A production system has to do much more: operate reliably across different scenarios, respect access controls, and hand complex cases to a human at the right moment.

Every action also needs defined evaluations, benchmarks, and guardrails, so the company knows what the agent can do, what it must not do, and whether each step produced the intended result. The entire process should be traceable: a human must be able to see every action the agent took and review how it reached the outcome.

That is why a strong test result does not mean an agent is ready for production.

— Your career moved through three tech projects in sequence: the legaltech startup AXDRAFT, where you automated legal document workflows; the development company Code & Cakes, which built enterprise products for clients; and now the insurance startup Superagent AI.

What problems did you solve at each stage, and how did that experience feed into your latest product?

— At AXDRAFT, the system handled the drafting, negotiation, and collaborative editing of legal documents. It also had to prepare large contracts as quickly as possible, freeing lawyers to focus on complex cases and higher-value legal work.

The main requirements were accuracy, predictability, and the ability for several people to work on the same document at once. At Code & Cakes, the scale of the work grew. We built enterprise platforms across different industries that had to handle heavy load, integrate with other systems, and run reliably inside processes that mattered to the business.

On top of the technical architecture, I was now responsible for timelines, the team, and business results. At Superagent AI, the system is responsible for an entire sequence of connected operations.

Specialised agents make and receive calls, gather customer information, enrich it with data from public and private sources, prepare insurance quotes, help complete the policy sale, and continue supporting the client afterward.

Here the challenge was tying several agents together through a shared context and figuring out what each one handles on its own versus when it brings in a human. So the scope of what AI was responsible for kept expanding — from a single document, to an enterprise platform, to an entire insurance sales process.

— Tell me about your time at AXDRAFT, where as the company's first engineer you built from scratch one of the few real-time collaborative editing engines that existed at the time — comparable only to Google Docs, Word Online and Ironclad.

Why isn't a regular editor good enough for legal documents?

— For a legal document, simple real-time syncing is not enough.

Several lawyers may be changing wording, data, and automatically generated clauses at the same time, and every change has to appear immediately for the relevant participants without compromising the consistency of the document. Some sections needed to be visible to everyone, while others could only be accessed by specific users.

Certain clauses also had to be reviewed and approved by several lawyers in a defined sequence. At the time, there was no existing product that combined real-time collaboration, granular access controls and sequential legal approvals. We built all three into the contract automation workflow.

As a result, document processing time dropped from 30 minutes to under five. That saved Sandoz 3,333 work hours and $300,000 a year, and BNP Paribas 10,250 hours and $922,500.

That project taught me that users trust automation when they can trace every change and check the final document themselves.

— You also led a project integrating GPT-3 into a live legaltech product, where the model explained complex legal language in plain terms. How did you minimise the risk of errors when using a neural network in a legal context?

— As a Y Combinator company, AXDRAFT gained access to GPT-3 before the technology became generally available.

During an internal hackathon, we built a prototype that explained complex legal language in plain terms, and later integrated the feature into the platform. The model did not alter the document or make decisions.

An explanation could sound convincing and still be inaccurate, so users always saw the original wording alongside it and could check the result themselves. That experience taught me that the level of autonomy given to an AI system should depend on the potential cost of a mistake.

The more serious the consequences, the more tightly its actions need to be constrained and the more clearly we need to define when a human makes the decision.

Later at Superagent AI, we expanded the range of tasks the system handles independently, but kept the same rule: ambiguous or high-risk cases go to a specialist.

— You then became a co-owner of the development company Code & Cakes and led the build of a news platform for well-known media entrepreneur Patrick Bet-David. VTNews.ai analysed content from more than 130,000 sources and compared how different outlets covered the same events.

How did you build a system that could keep up with that volume of news?

— VTNews.ai had to continuously collect and analyse articles from a huge number of sources, group together material about the same event, and surface results to users immediately.

So from the start we designed the platform's architecture to handle around 15,000 users per minute. The real stress test was election night in the 2024 U.S. presidential election, when both the volume of incoming news and the number of visitors spiked at the same time.

We had to hold performance steady even as news volume and user traffic both surged together. We pulled it off — the platform kept analyzing data and updating results without interruption.

— Why did you choose insurance for your new project after legaltech and enterprise platforms?

— Insurance brought together everything I'd worked on before. Like legaltech, it's a heavily regulated industry where precise language and control really matter. At the same time, it's a large operational business — calls, staff training, data collection, calculations across different insurers' systems, policy issuance, and client retention.

In the past, software products usually automated one step at a time, with a human tying everything else together. By 2025, the quality of language and voice models, orchestration tools, and the ability to train systems on industry-specific data made it possible to design much longer chains of actions.

At Superagent AI, I brought together what I'd learned automating expert work at AXDRAFT with what I'd learned building reliable enterprise systems at Code & Cakes.

— You started Superagent AI with a single agent, and now it's a platform of nine. How did you get to a system that covers the entire sales cycle?

— At insurance agencies, phone systems, CRM, SMS, email and carrier portals are often disconnected from each other. Even if a voice agent takes the call and gathers information, an employee still has to manually enter it into a rating system, get a quote back, and reach out to the client again.

To automate the whole process, we grew the product from one agent to nine and built Platform 2.0, which ties together the main communication channels and the work of the AI agents. The goal was never to pile on as many features as possible.

We kept expanding what the system could handle, step by step, until it covered the full insurance sales cycle. In the end, the platform became an operating environment where individual agents work together as one team.

— How does the human role change as AI takes over functions traditionally handled by insurance agents? Which decisions do you think need to stay under a specialist's control?

— Assessing complex risk, untangling ambiguous client needs, and decisions with serious financial or legal consequences all need to stay under a specialist's control. At the same time, people will spend less time on routine tasks and more time setting the rules the system follows, monitoring quality, and handling edge cases.

My view is that the next stage of enterprise software will look a lot like managing a team.

Each AI agent will have a defined role, its own access permissions, quality metrics, and clear conditions for when it escalates to a human. Our job is to keep expanding what AI agents can do on their own, while keeping them under control.

That's how AI moves from a handful of impressive standalone features to systems a business can actually trust with real work.