‘It’s important the culture is there to allow people to test and fail’ — How AI is reshaping workflows

Up until recently a wave of new hires was the clearest sign a startup was scaling. But with the rise of AI that’s become harder to spot, with growth increasingly coming from smaller teams integrating AI systems into their workflows, making them more efficient.

What began as the automation of repetitive tasks such as meeting scheduling, note-taking and transcribing, AI is now being used by startups to influence big decisions and make workflows quicker and smoother.

The question is no longer whether AI systems belong in workflows — but how they can be integrated without losing human judgement and responsibility. 

That was one of the main takeaways from our latest Sifted Talks, where we asked our panel of experts how AI can streamline workflows effectively and how startups can design workflows that become more useful as teams grow.

Our panel of experts included: 

• Marc Wiseler, head of solutions engineering EMEA at workplace software company Notion
• Oana Jinga, cofounder and chief commercial and product officer at warehouse robotics company Dexory
• Gordan Stuart, SVP of product at HR software platform Sage
• Sam Beek, chief product officer at AI video-editing platform Veed

What tasks can be done quicker with AI?

Dexory is using AI to streamline customer service, particularly to speed up response time to tell customers an inquiry has been picked up. After that a team of employees work on a resolution. 

“Our relationship with our customers is the most important thing that matters,” Jinga says.

AI systems can also be useful within R&D workflows to develop and test a prototype, adds Stuart, cautioning that you still need humans in the loop. “When you link up the expertise of our engineers with the pace of AI tools, the result has been exceptional in the R&D space,” he says.

Although some routine tasks can be done quicker and more effectively using AI, other tasks that require making important decisions, such as hiring, should still involve human skill, says Wiseler. “We try to keep humans in the loop as much as possible. I don’t think we’re ready for AI to take over the entire hiring process.”

Strategic decisions within a company can also be assisted with AI, but should be used with caution, adds Beek. “Our CEO locked himself in a cabin, made paper notes and then gave all of that to an LLM,” he says. 

The freedom to make mistakes

In order to implement AI systems successfully, employees must be able to use them freely even if it means making mistakes.

Senior engineers and architects at Sage who are confident using AI systems can go and test how they can improve workflows along with the leadership team, explains Stuart. 

Events such as hackathons and engagement sessions are also hosted to support employees in adopting the technology.

“It’s impossible to force and push people into using AI,” says Wiseler. The most useful way to get employees on board, he adds, is to share the success that AI has brought and allow staff to use it at their own pace.

Employees who are more comfortable using the technology should also help others who may be more reluctant, he says.

Collecting and inputting data effectively

Before the decision to implement AI systems into workflow has even been made, how data is collected and fed to these tools needs to be considered.

“If the data that you input into AI is incorrect then your results are going to be all over the place,” says Jinga. If you have AI tools linked to your CRM, for example, you need to educate and train employees to input data in the right way at the right time, Jinga adds.

Beek says that processes now need to be scrutinised far more to ensure correct data is produced. “If we use AI coding to help us program features, there will be a moment where we review that feature as if it was made by humans,” he adds. “Right before we want to release a feature, we try out what AI has added and if we don’t think it’s good enough, we send it back.”

By doing this, Veed’s AI systems are learning more about what works and what doesn’t. 

The more that these AI tools are used, the more information they’re going to pick up, says Stuart. “The more that I'm using the tool the more it seems to be learning and getting to know the context.

Knowing what kind of AI tools to use

Knowing whether to build new AI systems or use ‘off-the-shelf’ models can be a difficult decision to make.

Building their own AI models is part of the USP of Sage, Stuart says. “We've got a domain specific model that we built for accounting, for example. If we were just regurgitating something that came from mainstream traditional models, there’s no advantage in that.” 

At Veed, AI bots are created in-house to surface information across departments and to entice more AI users across the company. Every department at the company has created a bot that can answer questions for them.

“The HR team has an HR bot for example and over time these bots have reduced a lot of the work that people have to do,” Beek says. “People in HR are often bothered by very simple questions, like ‘how much holiday do we have’, and this is something that these bots can answer so that the team can keep on doing their work.”

Although at Dexory AI tools are often built from scratch, Jinga emphasises the importance of also making use of what is already out there.

“We build a lot of tools ourselves for our customers on the back of proprietary data,” she says. “There's also a variety of things we build in house to support different tooling for the way we look after our robot fleet or even our own databases.”