Traditionally, scaling a startup would mean hiring more staff. But in the AI era, that’s changed.
From marketing to coding, any team or department can be enhanced by the technology, allowing startups to scale without forking out for the extra salaries.
But is it as simple as logging onto ChatGPT. Of course not. That’s why, in our latest Sifted Talks, our on-demand virtual panel series, we spoke to the experts about practical strategies for using AI to solve real pain points.
We were joined by:
- Omar Davison, solutions engineer at Box, an intelligent content management platform.
- Ben Blume, partner and investor at Atomico, a venture capital firm built by founders.
- Rhiannon White, CEO of period-tracking app Clue.
Here are the key takeaways:

1/ Roll out AI internally then expand out to customers
White said AI is particularly useful for Clue to translate or transform content so it resonates with another country or area — otherwise known as content localisation. “We have an enormous amount of high-quality scientific content that we want women to be able to read in their own language,” she said. But while AI has helped speed up and streamline workflows, White noted they had also encountered some “watch outs” along the way.
For Davison at Box, the biggest shift was that internal AI development had become a playground for the product team. Whenever a new AI model came out, they had access to it within six to 12 hours and could start testing capabilities. If useful for their teams, it could be offered to clients too.
“The range of innovation and ambition for what people are looking at doing with AI is fantastic — very much for internal operations and growing their own businesses, but also for what their products are delivering to their customers.” — Ben Blume, Atomico
2/ Use isolated models to combat some of AI’s flaws
When asked if Clue was vibe coding — generating code based on natural language prompts — White said: “Vibes are not good enough when it comes to taking and treating people’s personal health data with the care and attention that it deserves.”
For example, when Clue used an AI tool to generate new images from slides that featured only women, the AI repeatedly added men into the images instead, often placing them in leadership roles.
To combat challenges of this nature, including privacy concerns, Davison suggested using isolated models.
“It’s a self-contained model that doesn’t talk to the world, it doesn’t remember any of the sessions or have memory of the conversations you’re having. That way, it’s only looking at the context you’re providing in the moment, for that moment.” — Omar Davison, Box
3/ Set aside time to experiment with AI in the workplace
Davison said one big mistake startups make when using AI is bringing the tech from use in a personal context to a business context and assuming it’s safe.
“One big issue is not having the right principles and guidelines in place before sharing with the world,” Davison said. “When we try to democratise tools that are freely available in our personal lives — like large language models — and bring them into a business context, we need to be much more safe, secure and thoughtful about how we use them.”
However he added that this shouldn’t stop teams being encouraged to explore how they can use AI. “The best companies empower that kind of productivity,” he said.
Blume agreed, saying that so much innovation comes from “play”.
“Google used to have this idea of spending 20% of your time on your own projects. I suggest doing the same with AI, spending one day a week just experimenting with how you could use it in your business.” — Blume
4/ Take extra time before releasing AI tools to customers
Does not using AI hold a startup back from an investment perspective?
“It really depends on what the business is doing,” said Blume. “If you're not putting AI at the centre of your product strategy, you need a good reason. There can be very good reasons, but the bar that ‘it's not a relevant piece of our strategy’ is a pretty high one.”
Davison said it depends on the context. Focusing on internal productivity is safer because you're only dealing with your company's data, and you've got those guardrails in place. Whereas dealing with a customer's data, where that might be revenue-generating, holds a higher value of risk.
White emphasised building AI slowly before releasing to customers.
“It's better to take a few extra months than to have a tool tomorrow that breaks trust. It's not like this is one moment and if we miss it, it's gone.” — Rhiannon White, Clue
5/ Hire people with a strong understanding and knowledge of AI
White said the success that the company has had with AI has been because its adoption has been driven by individual teams focusing on their niches.
As a result, it’s been important to “put the support in place to upskill and support our existing team,” White said. “I really wanted our teams to be able to do this themselves.”
Blume said he had seen more companies appointing someone from the management team as an internal “AI champion,” someone who creates forums for people to share ideas. However, he didn’t believe that roles like prompt engineer or agent manager were real jobs.
Davison added that while Box does not have prompt engineers, they have made AI hires.
“If you're building customer-facing services and products that include AI, then absolutely you need people with strong understanding and knowledge. But when it comes to internal use, you can be a little more relaxed.” — Davison
6/ Build with the idea you may need to pivot
How can founders build AI systems that scale as the tech develops?
“Whatever you’re building needs to be made with the future in mind — and with the idea that you may need to pivot,” said Davison. “Build a framework rather than commit to a singular focus. That’s what’s going to help you stay on top of whatever the future throws at you.”
White offered two principles: “One, choose a really specific use case — be defined and narrow about where you're going to start. And two, give your team the room and the space to play and to experiment.”
“Think in terms of the core capabilities of these models — text generation, summarisation and image generation. Use those as building blocks to figure out what will make your product genuinely more compelling, with real user value. Then assume each of those building blocks will only get better.” — Blume
Like this and want more? Watch the full Sifted Talks here: