ThoughtSpot CEO - ‘most AI agents are confident idiots’

Less than two months after ThoughtSpot launched its ‘team of agents’ for the entire analytics workflow, diginomica got the chance to sit down with CEO Ketan Karkhanis to get a clearer understanding of how he sees agents and ‘intelligence’ for business need play out with buyers that have been grappling with dashboards and data lakes for decades. And his view of how the industry has force-fed users supposedly useful data to date will likely ring true for many enterprise workers:

You've heard about self-service BI. It's such an inhuman thing to do. Who gets up in the morning and says, "I want to build a dashboard today"? I mean, it's crazy. 

We forget that the people that work have to have an experience that is aligned with how they work - which is accessible, which is normal, which is natural. 

Before we dive into Karkhanis’ approach, it’s worth outlining the challenges facing buyers at the moment. On the one hand, enterprises have both invested in Business Intelligence (BI) platforms already, or have been building out new data infrastructure in the form of data lakes or data warehouses, in the hope that they can operationalize insights across teams. This was all happening before the advent of generative AI and the subsequent release of AI agents everywhere. 

Couple this sunk investment with the onslaught of AI features being released by system of record vendors and enterprise platform players (e.g. Microsoft and Google Cloud), and buyers who are struggling to see where the value will ultimately lie when it comes to an engagement layer - or agentic system - it’s understandable that confusion is palpable. And that’s before buyers try to figure out who pays for these internally, how to bring users along on the journey, and try to prioritize governance and risk. I touched on some of this recently, where I spoke about how enterprise buyers are being presented with a new enterprise stack: data platforms, transactional applications and enterprise IT. 

With this in mind, it's understandable that buyers are looking for clarity on where to make strategic investments. ThoughtSpot's pitch is straightforward: you need a proper intelligence foundation before agents can work effectively. And according to Karkhanis, most vendors are getting this wrong.

The 'confident idiots' problem

Karkhanis didn't mince words when describing the current state of AI agents in the market:

Most agents are kind of really confident idiots. What do you do for an agent today? You've defined the problem, you connect it to a tool, and you say, go do this, and then the AI takes over. But do you even have the data to know if you should be doing that or not?

His central critique is that vendors are building "really expensive manual workflows" by connecting systems without the analytical layer that would make those actions intelligent. Karkhanis says: 

It's not AI, it's API.

The example he used was an NPS agent. Without intelligence, it flags a bad score and creates a ticket. With intelligence, it runs correlation analysis across customer segments, predicts which segments might have similar issues, and takes preemptive action across all of them - not just the one customer who complained.

This is where ThoughtSpot's semantic layer becomes important. The semantic layer is ThoughtSpot’s real competitive advantage, in my opinion - it’s the context component the industry keeps talking about. ThoughtSpot uses LLMs to understand user intent and generate search tokens (this problem was solved by OpenAI/Anthropic et al), it then converts tokens to ThoughtSpot's proprietary modeling language (TML) and runs the request through Trust Layer for verification. This request is then executed via an augmented analytics engine with relational search. Because the tokens are converted to SQL, ThoughtSpot argues there are zero hallucinations. He says: 

We have built the technology that converts a token to SQL. What we couldn't do was convert human language to tokens. That was hard. In fact, OpenAI and Google are doing it for free, and they are so good at it. So thank you Google, thank you OpenAI, thank you Anthropic. You just solved my biggest problem for me.

Karkhanis argues that you cannot point an LLM at your data warehouse and expect reliable enterprise analytics. He framed it as a fundamental mismatch:

LLMs, by definition, are probabilistic. Analytics, by definition, is deterministic. When you ask a question, what is my revenue? Do you want a probable answer? Or do you want a real answer? You are trying to mix oil and water.

ThoughtSpot is betting that as enterprises move beyond simple use cases, the semantic layer becomes essential for handling complex analytical queries, row-level security, and industry-specific business logic.

The new action layer

So, if other vendors are building agentic idiots, what can ThoughtSpot do and how does Karkhanis see this new enterprise stack playing out? The CEO’s point is that without this semantic layer, this intelligence foundation that can access both structured and unstructured data, which isn’t tied to one specific vendor, buyers are going to run into blocks. However, he believes that ThoughtSpot could become a system of action that enables enterprise-wide agentic work. But how does this tie in with the AI analytics work system of record vendors are trying to do? He said: 

Does that mean we are going to replace the system of record? No, why would I? Why would I even want to go there? It's a system of record. It's like the filing cabinet at the office. Let it be there. It's fine.

Ouch. I’m not sure system of record vendors would like being described as a filing cabinet, but Karkhanis’ point is a compelling one: 

When we think of actionability, think of it in like two to three simple categories. The simple category of action is connecting things and systems. What does that mean? I get an insight. I want to put it in a JIRA ticket for some developer to work on it, or some customer success team to work on it. That's an action. NPS went low - add it to, let's say, our VIP customer success team's queue to do something with them. That sounds trivial, but that's a lot of work to do, but we can do it today [in Spotter]. 

Beyond connecting systems as action, ThoughtSpot wants to go even further: 

Level two is bringing in what I call the ‘first steps of autonomy’. What does that mean? Level one, we will go make an action in the underlying success team system so that the tier one person will go do something with it. We connected them.

We would be like, "Hmm, based on everything we know, maybe we should send this customer an extra free month contract." Why should I even ask a human being? Seems like the right thing to do to improve NPS? 

So that's level two, wherein I'm not waiting, I'm not asking, I am doing. A human-being may be able to override it. It's still in that system. 

As an example, Karkhanis said: 

In the sales world, maybe the customer who gave us a bad review is a prospect. Maybe we add them to my next customer advisory board to bring back their trust and confidence? Maybe we send them a case study for a reference account in their industry to build back their confidence in us? All these are standard things. When I say them, you're like, "Yeah, that kind of seems logical." Why do you feel it's logical? Because that's what human beings do. We have learned to do this in sales, in marketing, in service.

And if level two is about autonomy, level three is about systems disappearing entirely. Karkhanis said: 

All systems. They start disappearing, wherein they become in the background, where autonomous workflows start running, and the interaction with the system is limited to what I call exception management. 

The competitive landscape

As noted in the introduction, buyers have been investing in data infrastructure for a while now and the enterprise competitive market does create some problems for ThoughtSpot. For instance, every major BI tool is now owned by a larger vendor - Tableau by Salesforce, Looker by Google, Power BI by Microsoft. However, Karkanis has an argument for why customers still choose ThoughtSpot:

Customers want an independent intelligence platform that works with everything. That ecosystem approach.

But he also argued these acquisitions led to stagnation:

When these companies got acquired by these big organizations, they became side projects. I don't think Google wakes up in the morning and thinks about Looker. And they shouldn't, because they've got other things to do.

The result, he said, was an "undercurrent" of dissatisfaction until AI forced enterprises to confront their analytics limitations. He calls this an "upgrade super cycle" like flip phones to smartphones - slow at first, then suddenly everyone switched.

On legacy BI tools with new AI features, Karkhanis had a pointed comparison:

You remember when the iPhone came out? And then BlackBerry said, me, too, me, too, me too. And BlackBerry came up with a slide-out keyboard. It kind of feels like that to me. They have slide-out AI. Their AI is slide-out AI.

However, when I asked Karkhanis about ThoughtSpot's challenges, in the context of the realities of the pressures enterprise buyers are facing, Karkhanis was direct. He said: 

99% of the world has not heard of us. I'm up against a go-to-market machine of 20,000 Salesforce reps. In Munich, they probably have 200 account executives, while I have only two.

Beyond visibility, convincing buyers on pricing at this stage can also have challenges: 

It's very common for customers to say: “our vendor is just going to give it to us because we are buying their hyperscaler”. Do what's right for your company. But just, it's important to recognize that reality - there is a reality of pricing monopolies or contractual monopolies…iinnovation is getting stifled. That may work a little bit, but in the long run, it is impossible.

If the technology doesn't help you compete and grow, why would you buy that technology? You may live with it because you got it free, but at some point, you're like, "What's the point, dude? We are actually losing.

My take

I do enjoy a CEO with a clear view and opinion. Karkhanis has a thoughtful, deep understanding of ThoughtSpot’s value - both to enterprise buyers specifically, but also within the broader context of the decisions being made in the market. The "confident idiots" will likely resonate with lots of buyers right now - agents need intelligence foundations, not just workflow orchestration. The semantic layer case aligns with what we're hearing from CIOs who've tried pointing LLMs directly at data. This is a market still in its very, very early stages. 

Even for ThoughtSpot, its agents don't yet communicate autonomously. Karkhanis says that the company is "three to five years ahead" of its competitors, but it too has plenty of work to do before we can realistically discuss ‘systems disappearing’. ThoughtSpot needs to prove clear ROI that justifies platform investment versus "good enough" embedded analytics that come bundled with existing contracts. And I do think its pitch is compelling.