Corporates rush to invest in AI agents

The next step in the generative AI goldrush is the creation of AI agents that can perform complex, multi-step tasks without human prompting — and corporate investors are rushing to invest in the startups creating these.

Microsoft’s CVC unit, M12, has backed startups like Inworld, which is looking at agents in gaming, while DTCP – the venture unit backed by Deutsche Telekom – invested in Cognigny, a developer voice AI agent for customer service, to name a few.

Corporate interest in these tools is high. Late last year, media and information conglomerate Thomson Reuters acquired Materia, a developer of agentic AI for tax and accounting purposes, which had been a portfolio company of its corporate VC unit, Thomson Reuters Ventures.

Register for free to access our database showing all the recent startup funding rounds backed by corporate investors.

AI agents are a natural extension of large language models, which have been getting bigger, with far more training parameters, larger context windows – the amount of information they can absorb in a prompt – and are becoming multi-modal, able to work across multiple types of media. The next step is programmes capable of handling complex, high-level tasks without human supervision, able to carry out multiple steps without additional prompting.

AI agents in daily life

Agents are not too different from having your computer running in the background, according to Michael Stewart, managing partner at M12, Microsoft’s CVC unit.

“A very simplistic way to think about it is, if you view the work with a computer as giving it instructions during a certain active period of time, and the work continues, do you really need to be monitoring each step all the time?” he says.

Work is already hot in the space, with some early startups generating revenue in the low tens of millions. But this is just the initial phase.

The breadth of their use cases could be as wide as any other AI application.

A classic example of an AI use case is in sales outreach. Unlike the typical robotic process automation we see today, which typically involves things like creating a hit list of prospects from an existing data set and sending off initial emails, an agentic system may source fresh prospects, issue follow-ups and back-and-forths with customers, bringing in a human only when necessary.

The use parameters and their effectiveness would vary from company to company. They could be managed by account executives, and some companies are also playing around with operational dashboard to manage “fleets” of agents.

As they evolve to more complex use cases – another classic one being that of booking travel – more considerations come into play. How much autonomy should these systems have? How empowered should they be? They can search for the best travel packs, but should they be allowed to execute the financial transaction to pay for them? These will require evolution in the legal framework – there is currently no standard by which AI systems can use financial credentials, and businesses are understandably careful about letting them have that kind of access because of security or compliance risks.

Stewart says AI agents could easily take over a range of mundane tasks in people’s daily lives, from filing expenses to arranging meetings and reacting to changes in your calendar.

“There’s a million little time savers you can think of, and none of these are really super hard to code, and they’re not super hard to imagine. This would be something that you would get used to very fast and quickly notice if it was gone,” he says.

The user experience, the exact way that you would interact with them, are not altogether clear yet. Would it be an app on your phone? A device you wear?

“I have a feeling this is just phase one. The next phase is really hard to see. How will people catch onto this? How will people start to notice it? That’s one of the things that’s really, really tough to judge until it’s happening.”

End of the enterprise software business model?

Many of the tasks an agent may be asked to do – generating content or analysing data of some kind – would not require constant uptime or dedicated servers, making the actual demand more sporadic. It doesn’t necessarily lend itself to the subscription-based revenues that make software-as-a-service popular.

SaaS is good for revenue predictability, capital efficiency, initial customer lock-in and other reasons, but AI agents are something that will likely be used as and when needed.

READ MORE

Big software companies under threat as AI undermines business models

“We’re starting to think the demand for AI will just be more interesting when it’s able to tackle a lot of different services when you need. The models, if you look at the actual technology, can kind of do anything, and you’re prompting it sometimes in unique ways,” says Stewart.

The challenge then would not be provisioning the server for the AI, so much as ensuring that the application cues the server or calls the API exactly when needed.

For startups this could mean being able to offer customers services without them having to commit to monthly payments. That would mean that monthly and annual revenue was less predictable, and subject to seasonality or macro factors in way that enterprise software companies haven’t typically had to contend with.

There may still be a place for subscriptions, but in the early days, at least, people will shop around a lot, and competition on price will be high.

Creating agents we love

Currently, the place where there’s perhaps the most movement – because people understand the immediate impact – is in coding.

Something like Copilot, Microsoft’s AI assistant feature, already has similar functionality to what an agent would do – handling context from different apps and generating relevant content, despite not chaining tasks in the same way as one would expect an agent to. Similar things can be said for coding assist programmes like Replit and Anysphere, which may not technically have what can be called agentic functionality but are approaching it.

But, says Stewart, successful AI agents won’t just be ones that are harnessed to particular business tasks or able to show a certain level of productivity gains. Some of the most successful companies may well be simply the ones that create a product people love to use.

“I hope entrepreneurs think about the fun and delight aspect of it. We can get to the super-intelligence and all this other stuff later. It’s more important to grow a business by making products people enjoy.” he says. “Not everything is about generating code and retrieving facts.”