Human… or AI?

They couldn’t tell.

Last Tuesday in The Bleeding Edge – They Couldn’t Tell It Was an AI, we explored some research showing how natural conversation with large language models (LLM) like OpenAI’s GPT-4 have become. 

So good, in fact, that most of the time, humans believed that they were communicating with another human. But it was an AI all along.

Even more exciting is that the research didn’t use the most advanced artificial intelligence (AI) at the time the study was done.

Now, the industry is anxiously awaiting the release of OpenAI’s advanced Voice Mode, which has been deployed to a limited number of test users. Its promise is to add emotions and tone, as well as very humanlike natural conversations. This will represent a major upgrade to OpenAI’s multi-modal LLM GPT-4o.

It’s this kind of technology that will make us humans feel comfortable. So comfortable, that we’ll forget that we’re communicating with an AI.

So comfortable, that we’ll feel like we’ve formed a bond with our AI “partner.”

While avid users of LLMs are focused on the release of the advanced voice mode, and for that matter OpenAI’s GPT-5 (or whatever it is called), there is a major undercurrent right now in the industry around something called agentic AI.

AI with Agency

Agentic AI, or agentic reasoning, is kind of like it sounds.

The technology, the AI, is given agency. It is given the authority or directive to solve a problem or complete a task through a series of steps.

This differs from today’s LLM technology, which provides users a zero-shot response. When we use something like ChatGPT, we give it a prompt, and then it returns us a complete response. The response is based on the information from our prompt, along with its pre-trained knowledge, and returned in a matter of seconds.

An agentic workflow is quite different. It is an iterative process, where an agentic AI uses a more human-like workflow to accomplish a task.

By way of example, here’s a zero shot or non-agentic approach – for a large language model like ChatGPT – to writing a paper…

We – the user – prompt the AI, which draws on its knowledge base from its training to produce an output.

If the output isn’t what we are looking for, we simply edit and revise our prompt, and we try again.

To state the obvious, it’s us, the user, that is doing the iteration. We edit and improve our prompts, iterate, and hopefully get to something that we’re looking for when working with this kind of generative AI technology.

But when we’re working with an AI agent (or agentic AI), it’s the AI that is doing the iteration. It’s the one that is doing all the work. And rather than a zero-shot response produced in seconds, it may take minutes or even hours to complete a given task.

To use the example a second time, let’s say an agentic AI is tasked with writing a paper based on a prompt. It would follow a series of steps, like the ones I have outlined below:

  1. Understand the user prompt/task

  2. Determine if any specific research is needed to complete the task

  3. Write a draft of the paper

  4. Review the draft as a critic

  5. Determine what needs to be improved in the paper

  6. Perform additional research if necessary

  7. Write a new draft of the paper

  8. Review paper again

  9. Write final paper

I’m sure that as we read the above series of steps, it felt normal – like the same approach we’d take if we were doing the work ourselves. It should feel that way, as an agentic workflow is very similar to human problem-solving.

An AI agent is different from an LLM though. It is given the agency to use an LLM for some of its workflow where necessary.

A simple way to think about it is that the AI agent is in charge, it’s the boss. And it uses other forms of AI or software tools to complete the task that it was assigned. Those tools can include a speech generator, the ability to make phone calls, and software that can be directed to interact with websites.

The research on agentic AI picked up last year, so it has been no surprise to see the activity in industry happening this year…

But well before the excitement over agentic AI started picking up last year, one of the more interesting AI unicorns had begun working on this technology. It’s been “on it” since 2021.

The Promise & The Potential of AI

The company, called Adept, focused its energy immediately on developing artificial intelligence for agentic workflows. 

Its purpose was to design AI in such a way that it could be used to complete both simple and complex tasks requiring multiple steps.

The applications for this kind of technology are wide ranging, from both enterprise applications to everyday consumer needs.

Just imagine the value in being able to complete the following prompts:

  • Locate all the inventory of Product# 892345-88 in the warehouse and aggregate in aisle 5 section 4.

  • Find the best travel schedule for me to be in Tokyo for meetings on Monday and Tuesday, Seoul for meetings on Wednesday afternoon, Taipei for meetings first thing on Friday morning, and to return to New York in time for a birthday party Saturday morning. Travel should be in business class, and seat preference is window seats, meal preference is chicken or fish for flights. Book travel once schedule if finalized.

  • Clean up the kids’ bedrooms, return toys to their place, dirty clothes to the laundry room, and then clean up the kitchen.

Being able to complete multi-step tasks like the above is exactly the point of agentic AI.

And it’s this kind of AI that will have an immediate and profound impact on both workplace productivity and our daily lives. Using generative AI to draft an e-mail or generate a quick image from a prompt is like a parlor trick compared to an “agent” that actually gets tasks done for us autonomously.

It’s this technology that will help us both complete mundane and time-consuming tasks to free up our time, as well as help us achieve some real breakthroughs in immensely complex problems.

Even the investors in Adept were telling… in terms of how its technology would be used. 

Three software giants, Atlassian Ventures (program management), Service Now (IT and customer service management), and Workday (human resources and finance management) all invested in Adept.

Those investments made perfect sense. All three companies would benefit from Adept’s technology as it was developed. Having a strategic stake in Adept would allow those companies to keep abreast of this bleeding edge technology.

But Adept had one challenge…

After raising about $64 million in April 2022, and $350 million in February 2023, it didn’t quite have enough funding to develop a full stack of artificial intelligence technology… including an LLM to compete with the likes of OpenAI, Google, Anthropic, and xAI.

It was at a fork in the road.

Either it needed to raise at least a billion dollars to compete with the big players, or it needed to refine its technology focus.

Tomorrow, we’ll find out which path Adept took.