By now, most people have heard of AI agents: software that can perform a set of tasks autonomously. But Asana decided to take a different approach when it comes to AI. On Wednesday, the company announced a beta version of what it's calling an “AI teammate” to help with work transitions within organizations.
Paige Costello, head of AI at Asana, said the company chose the name intentionally to signal a shift in how people think about interacting with AI in the workplace. “We believe the future of work is not just humans working with other humans, but humans working with AI,” Costello told TechCrunch.
“And we believe that in that world, it will be equally important to understand what we asked the AI to do, what it did, and how much it cost to get it done.”
Costello said it's important to build transparency and structure into AI, allowing companies to specify and create customized assistants to carry out core parts of their workflows.
That sounds good, but what does it look like in practice? Costello explains that while previous generations of workflow tools were rigidly defined, what sets today's announcement (and generative AI in general) apart is that it offers a more flexible way to move work around the enterprise.
So as work comes in, the AI can assess the current state and determine if it's ready to move on to the next step, or if the work needs to be returned to a human for more information before continuing. For example, if a help ticket has a missing or insufficient description, the AI teammate can send it back to the person who submitted the ticket and ask for the needed information. This could include using generative AI to help a human employee write the ticket before sending it to an AI teammate, who can then route the ticket to the right person for resolution.
Asana clearly has a treasure trove of data about how work moves within the company to train its models, thanks to its work graph, which models how work connects across individuals and departments. But while this all sounds good, we know that AI agents can still hallucinate and won't always understand the true nature of activity.
“The work graph allows us to tell the AI not just how the work is done, but how the work is done in a particular instance. So when you put an AI teammate into a particular workflow, it's assigned specific work. When it's assigned specific work and it knows what information to read, it's much more likely to do the right thing,” Costello said.
But Costello acknowledged that Asana encourages customers to keep humans involved because it recognizes that AI won't always make the right decisions. “I would say that Asana's guiding principle with AI is 'humans are involved,'” Costello said. “Ultimately, we believe humans are responsible for the decisions and accountable for the outcomes.”
This means humans need to be able to oversee and check the AI to make sure it is making the right recommendations, in line with the company's values and ways of working.
To solve this problem, Asana has been looking for workflows that can achieve high accuracy. “We found that by incorporating AI teammates, we can remove a lot of the admin and tracking work within these systems very quickly, with a high success rate. We can also use dynamic variables to effectively get information about the work and about the system within the context of the work,” she says.
That said, the tool is still in beta and there may be some growing pains, especially as companies move beyond experimentation and adopt it at scale. But if data is key to building more accurate models, organizations with insights into how companies work, like Asana, are more likely to succeed in moving the needle through the steps more intelligently.