Ali Sarrafi, CEO and Founder of Kovant – Interview Series

Ali Sarrafi, CEO and Founder of Kovant, is a seasoned technology and AI executive based in Stockholm with a track record of building and scaling high-growth AI companies. Since founding Kovant in late 2024, he has drawn on deep experience in enterprise AI strategy, go-to-market execution, and operational scaling. Previously, he served as Vice President of Strategy at Silo AI following its acquisition by AMD, where he was responsible for shaping enterprise AI strategy and driving large-scale adoption. Earlier in his career, he co-founded Combient Mix, leading the company through rapid growth and a successful acquisition by Silo AI, and has since held advisory and board roles across education and AI startups, reflecting a consistent focus on translating advanced AI into real-world business impact.

Kovant is an enterprise AI company focused on enabling organizations to move from experimental AI use to fully operational, autonomous business processes. The company develops an agent-based platform designed to orchestrate teams of AI agents across complex operational domains such as procurement, supply chains, compliance, and customer operations. By emphasizing secure, enterprise-grade deployment and rapid time-to-value, Kovant positions itself as a bridge between strategic AI ambition and day-to-day execution, helping large organizations embed AI directly into core workflows rather than treating it as a standalone tool or pilot project.

You’ve led major AI initiatives at Spotify, scaled and exited Combient Mix, and later shaped enterprise AI strategy at Silo AI before founding Kovant. What specific gaps or frustrations did you encounter in those roles that convinced you the time was right to build an autonomous enterprise platform, and how did that history shape Kovant’s core design philosophy?

Across my previous roles, a few consistent gaps kept showing up. First, most “vertical” AI tools are effectively captive to a single software stack: they do one thing slightly better inside that boundary, but struggle the moment a workflow needs to span multiple systems. At the same time, enterprise data is scattered across lots of tools, and many automation solutions simply can’t reach it. Layer on years of point integrations and you get classic spaghetti architecture: complexity rises, change gets slower, and teams end up automating individual steps rather than reimagining the workflow end-to-end. The result is that ROI often arrives slower – and smaller – than organisations expect.

Kovant is designed as a response to that reality. Our core philosophy is that agents should behave more like employees: they work across multiple tools, they’re “hired” to do jobs, not to automate a single scripted sequence. That’s why integrations and orchestration are built in, and why we assume enterprise data is often messy and unstructured – it needs a more human-like approach to handle exceptions and ambiguity.

We use foundation agents to achieve speed and scale, while keeping data sovereignty front and centre: enterprises can access and use their own data horizontally without it leaving their premises.

Kovant positions itself as an autonomous enterprise platform capable of running entire operations and departments with AI agents. How do you define “autonomous” in an enterprise context, and how is this different from the automation and agent tools companies are already experimenting with today?

In an enterprise context, when we say “autonomous” we don’t mean “unsupervised”. We mean AI agents can take real actions end-to-end across an operation with clear goals and guardrails, and they’ll escalate to humans when supervision is needed.

What makes Kovant different is our foundation agents. Rather than automating a single, fixed process or following a pre-built sequence, Kovant agents can work as a team (or swarm) on an operation using just instructions and an operations overview we call a blueprint. They’re not designed for one narrow task; they collaborate to solve complex workflows, adapt as conditions change, and hand off to people when the situation requires oversight.

For example an inventory management agent team can perform all of the following jobs without rebuilding them from scratch, including: communicating with suppliers via email, monitoring inventory levels and out-of-stock signals, tracking shipments and purchase orders, updating statuses across systems, creating discrepancy tickets for inventory planners to approve, redistributing inventory between warehouses, and consolidating inventory reports.

So the shift is instead of “chat plus tools” or brittle automations that break at scale, enterprises move from building agents to running them at scale.

Despite massive interest in agentic AI, many organisations remain stuck in pilot mode. From what you’re seeing in real deployments, what are the main reasons companies struggle to move from experimentation to scaled production?

What we’re seeing is that most organisations don’t get stuck in pilot mode because the idea is wrong; they get stuck because the environment is hostile to scaling.

The first blocker is the fragmented enterprise tech landscape. Workflows span loads of systems, data lives in multiple places, and stitching everything together reliably is hard. And agentic AI is often deployed as an add-on to existing tools, rather than as a way to rethink how the workflow should run end-to-end.

There’s also a real architecture and data problem. Many SaaS vendors still try to lock data in, which creates incompatibilities and limits what agents can actually do across systems. And a lot of teams underestimate the fact that most enterprise data is unstructured (emails, documents, tickets, PDFs, chat logs). If your approach assumes clean, structured data, time-to-value becomes long, painful, and difficult to replicate beyond the pilot.

In short: fragmentation, lock-in, and unstructured data create drag – and pilots never turn into production until those realities are designed for.

Reliability is often cited as the biggest blocker to deploying AI agents in the real world. Why do so many agent systems fail once they leave controlled environments, and how does Kovant’s approach reduce issues like hallucinations and unpredictable behaviour?

Some agent systems look great in demos, then fail in the real world because the environment is messy and unpredictable. Data is incomplete or inconsistent, edge cases show up constantly (refunds, disputes, special approvals). Workflows span multiple tools, platforms and integrations that change over time, and permissions vary. When an AI agent is asked to handle a large task and is given too much context at once, the risk of hallucinations and odd behaviour increases.

Kovant reduces this by design. Our unique architecture narrows the problem space, decision space and context that models work with to reduce hallucinations. We also break operations down into narrow, focused tasks for individual agents and steps. That makes behaviour more predictable, and it adds traceability and controllability into the system and can manage hallucinations better. We can see what each agent did, where a failure started, and intervene or escalate when needed.

Hallucinations don’t magically disappear, but by constraining what each agent is responsible for and limiting the context it can act on, we can reduce their frequency and limit their impact. This “narrowed task/context” approach has also been supported in recent work from Nvidia’s research team, which found similar benefits from constraining agent decision-making.

Accountability is a major concern as AI agents begin to take real actions in business systems. How do detailed action logs change the conversation around trust, compliance, and operational risk?

With detailed action logs we can see what happened, why it happened and what happens next.

The detailed logs turn an agent from a mysterious bot working away in the machine into a system you can inspect.

At Kovant, with any AI agent deployment there will be a risk map that the organisation can act on, we have built-in gate keeping for humans for risky actions that means agents can only perform those tasks if a human reviews and approves the decision. All of these are logged the same way as a system of records is logged and are traceable.

We believe it’s important to combine action logs with human oversight and observability to minimise the risk. It means you still get the speed and scale benefits of agents running real operations.

There’s growing discussion about whether AI agents can even be insured due to their opaque decision-making. How does making agent workflows auditable and replayable help address the “black box” problem and open the door to insurability?

The “black box” problem is what makes insurability hard. If you can’t clearly show what an agent did, why it did it, and what controls were in place, it’s tough for anyone, especially insurers, to price the risk.

Our approach is essentially an extension of the accountability setup in the previous answer. We break the decision scope and the impact of actions into smaller chunks, so the model isn’t making one giant, opaque decision that can swing a whole operation. Each step is narrower, more predictable, and easier to evaluate.

We then add detailed logs, observability, and human oversight. For the most important and impactful decisions, we use a human gatekeeper so the agent can only proceed after review and approval. That creates far more visibility into how the workflow behaves in practice.

Making workflows auditable and replayable is the final piece. If something goes wrong, you can reproduce what happened, investigate it quickly, validate fixes, and demonstrate how often human approval is required and where the safeguards sit. In underwriting terms, that turns mysterious AI behaviour into something closer to standard operational risk.

With initiatives like the Agentic AI Foundation aiming to create shared standards for agentic systems, what do you see as the most promising aspects of these efforts, and where do they still fall short for real enterprise operations?

Standardisation is generally a good thing. The AAIF can do the unglamorous but essential work of getting agent systems to speak the same language, which should make integrations easier and reduce vendor lock-in over time.

Where I’m cautious is whose perspective shapes the standards. If most of the work is led by model creators and tech scale-ups, there’s a risk the “standards” optimise for what’s easiest to build or demo, rather than what large organisations actually need to run agents safely day to day.

For real enterprise operations, the gaps tend to be less about connectors and more about control: what an agent can access and change, approval workflows for high-impact actions, auditable logs, and observability so teams can monitor behaviour, investigate incidents, and prove compliance. Enterprises also need practical standards for operating in messy reality: testing against edge cases, handling changing systems, and being able to pause, contain, or roll back actions safely across legacy tools and regulated data environments.

So it’s a promising direction, but the impact will be limited unless enterprise requirements and operational risk controls aren’t treated as an afterthought.

Kovant has already generated significant revenue from large Nordic enterprises while operating largely in stealth. What types of business functions or workflows are proving most ready for autonomous AI agents today?

From what we’ve seen in real deployments, the workflows most “ready” today are the ones made up of reactive white-collar work: monitoring, chasing, checking, updating systems, handling exceptions, and keeping operations moving across multiple tools.

In manufacturing and broader enterprise supply chains, that shows up across:

• Sourcing/procurement: raw materials availability, sustainable sourcing, compliance ops, supplier selection (including dual/multi-sourcing), contract management, supplier risk management, and tendering/bid management.
• Production: capacity planning, production scheduling, maintenance management, quality management, bottleneck management, and loss prevention.
• Warehousing: receiving & inspection, inventory management, stock rotation (FIFO/FEFO), and cycle counting/auditing.
• Transportation / logistics: mode and carrier selection, customs clearance/documentation, tracking & visibility, emissions monitoring, and trade compliance.
• Sales and service: product availability, stockout prevention, sales/returns management, consumer behaviour analysis, plus aftersales areas like repairs, end-of-life tracking, workshop ops, and service contracts.

When enterprises deploy AI agents across critical operations, how do you recommend balancing autonomy with human oversight to ensure control without slowing everything down?

The balance is governed autonomy. You have to let agents move quickly on low-risk work within clear guardrails, and escalate to humans when the action crosses a defined risk threshold.

A lot of  failures come from giving the model too much scope and too much context at once. I recommend breaking operations into smaller, narrowly scoped decisions, where each step has clear permissions and a limited impact radius. That reduces unpredictable behaviour and makes performance easier to monitor and improve.

Then you combine three things: observability, action logs, and human gatekeeping. Everything the agent does should be traceable, so you can inspect what happened and investigate fast. For high-impact or risky actions, you put a human approval step in the workflow, so the agent can propose and prepare, but only executes once a person signs off.

That keeps thing moving quickly. If anything it only slows down ever so slightly at the human oversight step, but that’s an important part of the process. Humans aren’t stuck supervising every click, but they’re still in control of the moments that matter. The result is speed where it’s safe, and oversight where it’s necessary.

Looking ahead, how do you expect the role of autonomous AI agents to evolve inside large organisations over the next few years, and what will separate companies that succeed with agentic AI from those that struggle?

Over the next few years, autonomous AI agents will move from interesting experiments to becoming a real operating layer inside large organisations. They’ll be used for operations, customer service, finance and HR. As reliability, governance and oversight improve, we’ll see enterprises shift from isolated pilots to running agent teams across end-to-end workflows.

The biggest change is that speed, agility, scale, efficiency and costs will become a much more direct competitive lever. I think an “Uber movement” is coming for enterprises. The ones that truly master agentic AI will be able to operate at a fundamentally higher pace than laggards, capture markets faster, and respond to change without the usual operational drag.

What separates the winners isn’t just deploying agents, it’s deploying them well. Governed autonomy, strong observability and action logs, and architectures that narrow decision scope will be key to that. Companies that treat agentic AI as a core operating capability, with the right controls, integration, and ownership, will use it to do more, not less. That’ll free teams to focus on growth and innovation rather than spending their days buried in admin. In short, radical speed and efficiency become a true competitive advantage at enterprise scale.

Thank you for the great interview, readers who wish to learn more should visit Kovant.