AI pilots are running into margin pressure and governance gaps

Box, Forestay, Naboo and Dust executives said companies moving AI from pilots into production are running into cost, governance and talent problems rather than model capability limits. No deployment budgets, vendor contracts or customer adoption rates were disclosed, but the panel put numbers on the pressure: Forestay’s Jannat Rajan said some companies embedding AI in core workflows have seen gross margins fall from 80% to 90% to 50% to 60%.

The discussion took place during a Sifted Talks session sponsored by Box, which sells content management, collaboration and workflow software. The participants were Omar Davison, solutions engineer at Box; Rajan, a growth investor at Forestay; Lucien Bredin, cofounder and chief product marketing officer at AI events and procurement platform Naboo; and Thibault Martin, ecosystem lead at AI company Dust.

Productivity wins are not the same as production systems

Davison said Box first tries to identify whether a customer is pursuing individual productivity, departmental efficiency or organization-wide gains. For companies that are less experienced with AI, he said the early gains often come from small time savings for employees, while broader deployment requires more controls.

Naboo’s Bredin said the company initially overestimated how much customers cared about the underlying AI model. Customers were more focused on whether the system produced the outcome they wanted, he said. Naboo then trained agents on four years of company data, past booking communications and tone of voice, creating what it calls an “AI Twin.”

According to Bredin, that system now handles about 80% of the organization of an event hosted by Naboo, with an account manager handling the remaining 20% to preserve trust. Naboo did not disclose the number of events involved, the error rate, or the cost of running the system.

Token costs are becoming a finance issue

Martin said AI projects often stall because leadership teams lack a shared framework for security, budgets and accountability. He described AI as another resource competing with employees and outsourced labor, meaning business leaders need to decide where it should replace or support work.

Rajan said pilot programs can sit inside R&D budgets, but production workflows create less predictable token costs. She said some companies are responding by adding finance operations practices and repricing their software. Her view is that companies should avoid sending every request to the same large model and instead use a mix of AI engines, including more targeted models paired with internal data.

Rajan compared the approach to the multi-cloud shift in the 2010s, where customers spread workloads across vendors rather than committing everything to one provider. She also said companies may soon employ specialists focused on token and model spending. That forecast was not tied to hiring data.

Hiring and performance management are changing

Martin said conventional interviews are often a poor fit for AI roles. Some organizations now give candidates AI tools and ask them to design workflows for real business problems, looking for adaptability as much as technical skill.

Bredin said Naboo has made AI use part of performance management. At the company, 10% of an employee’s annual evaluation is tied to how the employee builds, manages and uses AI agents. He also argued against pulling back from junior hiring, saying younger employees are often accustomed to rapid tool changes.

Martin said managing AI agents has similarities with managing people: managers set expectations, define success and give feedback. If an agent produces poor work, he said, the issue is often that humans have not explained the task clearly enough.

On investment criteria, Rajan said she is most interested in founders with domain expertise, proprietary data and a specific problem, citing examples such as taxation, law and industrials. Davison said trust depends on governance, training, change management and human oversight, all of which are slower and less marketable than a model demo but more relevant once AI reaches production.

This story draws on original reporting from Sifted.