As generative AI enters the end of its honeymoon phase, organisations are now moving beyond the experimental phase and getting ready to commit.
But turning early AI use into widespread deployment is where many startups hit a wall.
The biggest challenges they face are operational — increasing costs, governance and trust issues, and the transition from individual AI use to managing autonomous ‘digital workers’.
These were the key talking points during our most recent Sifted Talks in partnership with intelligent content management platform Box, where panellists unpacked how founders can break out of the pilot phase and build resilient, AI-native operations.
Our panel of experts included:
- Omar Davison, solutions engineer at Box
- Jannat Rajan, growth investor at VC firm Forestay
- Lucien Bredin, cofounder and chief product marketing officer at AI events and procurement platform Naboo
- Thibault Martin, ecosystem lead at AI company Dust
Getting AI pilots right before widespread adoption
When working with organisations to adopt AI, Box aims first to determine where they’re getting the most value, says Davison.
“Are they looking to improve on individual productivity? Are they looking at departmental efficacy? Or are they looking at organisational efficiencies?,” he says.
For organisations that are less AI-native, the early value they get often comes from individual productivity, he adds.
“Because they still have guardrails and processes they have to adhere to, the quickest wins often come from saving their team 30 minutes of their day. It's then a growing curve from that point."
When Naboo first launched and started using AI, the company assumed its customers cared about the underlying AI model and how it worked but in reality they cared more about the outcomes it had.
Instead of building a generic chatbot, Naboo trained its agents on four years of data, tone of voice and past booking communication and developed its “AI Twin”.
On a daily basis the AI twin is managing around 80% of the organisation of an event we host and 20% is managed by the account manager to maintain that trust," - Bredin.
Governance and managing financial operations
When AI deployments stall at scaling phase, it’s rarely due to the technology itself, says Martin.
A company’s leadership team all operate under different incentives. Without a clear framework outlining AI security and budget allocation, deployments get stuck in a loop, he says.
“Within companies, you have full-time employees and you have outsourced employees. AI is a new form of resource to which you are reallocating a part of the job,” Martin adds. “You want to make sure business leaders are accountable for allocating AI where it’s needed most and they feel empowered to do so."
While AI pilot programmes can be treated as R&D experiments, rolling out agentic workflows can result in unpredictable token costs — the financial and computational expenses that occur when using AI models.
According to Rajan, companies that once enjoyed 80%-90% gross margins are now seeing margins drop to 50%-60% when embedding AI into their core workflows. To tackle this shift, businesses are implementing specific finance operations practices and are repricing their software.
Rather than routing every internal request through models, organisations should build a portfolio of AI engines and pair lightweight, highly targeted models with proprietary internal data.
“This is very similar to the multi-cloud wave we had in the 2010s,” says Rajan. “The principle is exactly the same: don't lock in entirely with one vendor but diversify and pick the right spend.
“Financial Operations is essential to help manage these kinds of costs," she adds.
We're not far from having specialist employees who focus solely on token spend and model spend." - Rajan
The evolution of talent
Traditional interviews are often not adequate when hiring AI talent and many organisations are replacing standard behavioural questions with AI assessments.
Candidates are given access to AI tools and are asked to design a system of workflows to solve real-life business problems.
“One thing we look for in people is adaptability,” says Martin. “Can they see how their current job will be different six months from now and are they getting ready for that new wave? Two years ago you had no coding agents. The ability to reinvent yourself is a skill that was less important previously because you had more time to adjust."
Encouraging an entire workforce to adopt AI can’t be entirely down to individual efforts. At Naboo, AI adoption isn't optional — 10% of an employee's annual performance evaluation is directly tied to how they build, manage and use AI agents.
“It should be the same in every company and the younger you are, the easier it is,” says Bredin. “I don't always agree with a lot of tech companies that stop hiring juniors. They're used to this kind of change."
According to Martin, there are similarities between managing direct reports and managing AI agents. A manager doesn't do the job for their employees, but defines what success looks like and provides feedback.
“With AI, you need to be clear on what’s expected. This is typically what you would get from managers,” he says. “They lay the foundations for employees to be successful.
If the agent is producing something that's not useful, it’s often down to the humans who've been working with these agents and have not fully explained their expectations." - Martin
Building trust, transparency and value
Investors are increasingly looking for founders who have industry expertise and are using AI to solve specific challenges.
“The companies that are exciting me the most are the ones that have a clear grasp of a particular domain,” says Rajan. “They might have a solution to a very specific problem in international taxation, for example, or in law or in industrials.
"When I meet somebody with genuine domain expertise who's collected data you haven't seen before and they build the intelligence on top of that — for me that's absolutely magic.”
Customers don’t just want a fast answer — they want to feel heard and respected, says Martin.
“Humans are much better at this. You're rarely going to convince somebody of something just because the answer is right,” he says. “Sometimes it takes more and that’s a good guiding principle.”
Trust is earned through human-in-the-loop oversight, strong proprietary data, and training until the AI's output aligns with human standards, says Davison.
“Applying governance, training and change management are all ways we reduce the ‘principle of least surprise’ and ensure AI models’ outputs are reliable and correct,” he adds.
As we guide and provide the instructions, skills and data, we can then roll out and gain the benefits of AI." - Davison
