Sifted - Giving AI agency

Good morning there,

Imagine a startup that’s profitable with zero employees. Some people believe that might be possible by the end of next year — with the support of AI agents. These automated helpers can be fed a big task and then complete the required steps from start to finish without the back-and-forth prompting needed with chatbots like OpenAI’s ChatGPT. But new research suggests they’re not about to be deployed en masse just yet. News editor Tim Smith digs into that below. 

Elsewhere today, we take a look at the top policy priorities of the rebooted European Startup Network, an umbrella organisation of startup associations.

Plus, next week Sifted’s Éanna Kelly is bringing together a suite of operators from across European tech to look at why data might be the missing link to your scaling efforts in 2024. Don’t miss it, it’s free. RSVP here.

— Anne Sraders, senior reporter

/A message from our sponsor Up!Rotterdam

Startups, are you looking for capital to grow your business?

Whether you're a founder, investor or a corporate changemaker, Upstream Festival in Rotterdam on 28-29 May is a community-driven festival — and your gateway to investment opportunities, cutting-edge ideas, meaningful connections and endless possibilities.

Secret agents: When will AI bots actually get good?

AI “agents” represent one of the most futuristic visions of a generative AI-led future. 

The term describes tools that use the most powerful new AI models — like OpenAI’s GPT-4 — and give them agency. Rather than prompting the system repeatedly every time you want an output, you feed an AI agent one big task, then it does the steps required to get there.

People building this kind of tech predict that AI agents will entirely run the first zero-employee, profitable businesses by the end of next year. There are already startups selling so-called digital workers that say AI can outperform your average junior sales rep.

But a new report out today throws a bit of cold water on that hype. It finds that AI agents are still in their “very early days”, even if some limited use cases are finding traction.

The research comes from Prosus Group, a software developer and investor that's has been working with more than 25 of its portfolio companies to build an AI agent for data analysis.

It lets any worker in the company ask data-related queries about a range of things from customer behaviour to internal systems, without having to use any technical coding language. The agent then goes off, works out where it needs to get that data from, how to write the code query and then questions its work before running the query and preparing an answer.

These steps represent a fairly simple workflow but also a big challenge for today’s AI models — which are known to make mistakes and behave erratically. Prosus’ senior director of data science Paul van der Boor tells me that a lot of work has gone into stitching together software tools with the latest large language model (LLM) technology to create a system that “self critiques” itself enough to provide reliable results.

He says cost is still a big limiting factor in rolling out AI agents at scale in businesses. It’s a problem I’ve heard about before from London-based AI agent startup 11x, which told me that some of its most data-intensive autonomous “workers” cost $12 per hour to run — more expensive than the minimum wage in many countries.

Another limiting factor is improving technology to help LLMs become better at planning. Currently these systems are essentially black box text predictors, which crunch huge amounts of text to statistically assess what word should come next in a sentence. To become more powerful at longer and more complicated tasks, they’ll need improved memory and reasoning systems to control them.

If all this happens, we could be looking at a future economy where whole job titles are automated away, and where super smart AI’s can make better strategic decisions than some top CEOs. Many will hope that their jobs still require a human touch, and aren’t in the firing line just yet.

Others will be hopeful that soon we’ll all be able to put our feet up while the autonomous agents get the work done. Whichever camp you’re in — today’s research suggests we have a while to wait.

 — Tim Smith, news editor

👀 SoftBank is in talks to buy UK chip startup Graphcore, according to Bloomberg, which cited sources close to the matter. Sifted reached out to SoftBank for comment. Graphcore declined to comment.

🤝 The umbrella organisation for Europe’s startup associations is getting a makeover. The eight-year-old European Startup Network, or ESN, today announced a new board, president and CEO. It aims to act as a partner to startup organisations across Europe and help advocate to fix the biggest challenges facing startups. 

• The ESN is also setting its sights on several top policy priorities ahead of the EU elections in June. This includes pan-European tech visas, equivalent stock options programmes across Europe as well as attracting more money from pension and insurance funds to invest in VCs.

B2B SaaS: Here’s how to avoid the chopping block

In a challenging economic climate, where companies are scrutinizing their spend, SaaS tools are often the first to be cut back. Our latest report offers invaluable insights from Europe’s software masters on not just surviving, but thriving, amidst the current market lull. 

Valencia, Spain-based Sesame, a HR management software startup, raised €23m in funding. GP Bullhound led the round and was joined by PSG. 

Munich, Germany-based Insempra, which develops sustainable ingredients for food and beauty products, raised $20m in Series A funding from investors including EQT Ventures, BlueYard Capital, Possible Ventures, Taavet+Sten, Acequia Capital, Henkel dx Ventures, Bayern Kapital and Alante Capital.

Oxford, UK-based Dendra Systems, which uses AI and drones for ecological restoration, raised £12.5m in Series B funding. Zouk Capital led the round and was joined by investors including  Aramco Ventures, Airbus Ventures, Helium-3 Ventures and Understory Capital Management.

Copenhagen-based Electricity Maps, which helps companies track the carbon footprints of their electricity usage, raised €5m in funding. Transition and Revent led the round.

Basel, Switzerland-based Axmed, a B2B medicine procurement marketplace aimed at low- and middle-income countries, raised $2m in seed funding. Founderful led the round. 

Manchester, UK-based GigPig, a live music marketplace, raised £1.3m in seed funding. Haatch Ventures and Notion Capital led the round.

Aalter, Belgium-based Artion, software which automates and tracks legally required company inspections, raised €1.4m in funding. Entourage led the round and was joined by investors including Dovesco, Tilleghem and Belfius.

If you’d like to submit a deal, get in touch. 

For more deals, analysis and M&A insight, become a Pro subscriber to receive our weekly Deals newsletter.

Enjoying our newsletter?

Forward it to your network so they can sign up here to stay informed with industry news and insights, just like you.