What to automate, what to keep human

In May 2026 I sat on a panel at Latitude59 in Tallinn called "AI in Your Org: What to Automate, What to Keep Human", with the chief of staff at Hostinger and a co-founder of Sera Leads. Three very different company sizes: 900 people, 150, and 10.

Below are my answers, pulled out of the discussion and lightly edited. I've left their parts out rather than paraphrase them, so watch the full panel if you want the whole thing.

What percentage of your workflows are already AI-assisted or automated?

I'd be surprised if there was a significant chunk of Pactum that wasn't touched by AI at all. Maybe not every person all the time, but it's pretty ubiquitous.

What tools are you using, and how do you keep track of AI usage?

I have a dashboard in Looker showing weekly actives, how much they're using it per week, broken down by function and depth of usage. Claude Code is our main surface for AI agents internally, so that's relatively comprehensive.

There are other tools available in principle that are rarely used in practice. We have a Lovable credit sitting somewhere worth 25k that I haven't claimed, because nobody is really using it. We have ChatGPT, Gemini, Slack AI, which I don't know if anyone uses, some Salesforce agent platform credit we're mostly not touching. We're investing in Claude as the main way to bring all our systems together and as the main AI collaborator in the work.

Are there tools people use that you don't know about?

Probably. Shadow IT is a real category and there's definitely some of it. Our approach is that when we see something, I'd rather not forbid it and instead work out how we can use it, because organic adoption usually means there's something valuable there. At this point I'd be surprised if there were anything significant, because almost anything you might want is already available, or will be if you ask.

Have you created the 10x employee?

Our engineers were 10x to start with, so we aim for 100x. That's a bit of a joke, but it is genuinely hard to 10x anyone. Take any role and decompose it into the things that person does every day. Maybe 20% can be done almost entirely automatically by Claude or some automation. Then there's the part where you have to sync with other people, either in writing or in a meeting, and there's a cost to that which is very hard to avoid. You can use Wispr Flow to dictate instead of typing. But if you're meeting a customer, you can't turn an hour into five minutes just because you're better at using AI. At least not today.

So for most people, it turns out a large share of the job, 60% or 80%, is quite hard to automate with Claude, or even hard to do entirely virtually without a physical body, a charismatic face and a warm handshake. Which is weird. But it's part of why it seems unlikely to me that we'd cut headcount 10x anywhere. Most jobs in most places, at least in most startups, are probably relatively safe.

When does AI productivity become an operational risk?

Two case studies.

The simpler one: someone in our go-to-market organization posted a competitive analysis deck to Slack, about how we compare to competitors and how to talk about the differences with customers. One of our senior executives commented on it, and I'm paraphrasing, something like "I believe most of what is in here is wrong, you should not rely on AI for things like that." What went wrong is that the person trusted Claude to do it for them, it looked legit, and they accepted it. For competitive analysis you need to understand what customers and potential customers are actually saying and thinking, not what's on a marketing website. Claude went with the marketing website, and the output was garbage, and the person shared it as the outcome of their work. That's embarrassing to me. It doesn't matter what tool you use, you're still accountable for what you produce, and we try to hire people who understand that and own the outcome.

The second one is harder, because the first was AI slop accepted uncritically, and this one happens even when everyone is using AI well. Imagine a hundred sales calls transcribed in Gong, connected to Claude, and a hundred people in the company go and investigate those calls with Claude. They get a hundred different narratives and a hundred different answers, each focused on a different aspect. That slowly melts the structure of the company apart, because you no longer have a common understanding of what customers are saying or what matters. It's more hidden than the first problem and I think it's more dangerous. Kristjan, our co-founder and chief scientist, gave a half-hour talk on exactly this pattern at the same conference.

Isn't the compounding failure mode the real threat, one agent controlling another until it breaks?

Let me push back a bit here. Imagine you have a sales manager agent with a team of account executive agents, each working with an SDR agent, doing outreach and emailing customers together. Unless you're reading every email and staying very on top of that process, you'll need some form of trust. Many things could go wrong. They could say random things to a customer, sell things that don't exist in the product.

Now replace the word "agent" with the word "employee". That's already happening today, and the tools you'd use are one-on-ones, dashboards, occasionally analysing what's actually going on. So I don't think it's a fundamentally novel problem. It might even be an easier one, because it's quite intrusive to screen-record every person's laptop 24/7, and much less intrusive to see what every agent is thinking, its reasoning trace, which actions it called. Maybe this gets solved by programming agents rather than programming humans, which is very hard.

How do you teach employees not to trust AI too easily?

If I had a solution I'd tell you. I'm not sure it's even specific to AI. You can rely on Google too much, you can rely on Wikipedia too much. As I said, it doesn't matter what tool you use, you're still accountable for what you produce. That's what we should focus on: is what you're doing actually what you're supposed to be doing, and are you doing it well? Whether you used AI is somewhat secondary.

What should stay human?

You left the last forty seconds for the title of the panel, I see.

Someone needs to own the result. Eventually that's the CEO, and the CEO needs to make other people accountable, usually executives, who have managers. In a smaller company there's less structure to it, but someone has to own it: making sure that whoever is doing the work, human or agent, is doing it well, that the process is well described, that we know what we want out of it. That seems hard to stop keeping human for a long time.