AI will cut enterprise G&A headcount in half within three years, and the evidence is already in the results

Nikesh Arora offered a rule of thumb that deserves to be taken seriously: within three years, enterprise companies will likely have half the people in G&A functions, covering marketing, finance, and HR, because of how much process management those roles involve. That is not a hedge or a range. It is a specific claim about a specific timeline from a CEO who runs a large enterprise software company and watches how those companies operate.

The Curative case study, as described by the company’s CEO Fred Turner, is the kind of evidence that makes Arora’s projection credible rather than provocative. Curative’s credentialing department, the function that verifies licenses and qualifications for healthcare providers, used to run on a turnaround of two to three months per credential at a cost of roughly $50 per credential. The company built an internal agent running on Claude that now handles the same process end to end: visiting websites, verifying licenses, reading transcripts, assembling documentation, and stamping for approval. Average turnaround is now 12 hours. Cost per credential is $0.20.

The savings are not limited to credentialing. Turner reports that Curative built an internal CRM using vibe-coded tools, integrated AI agents directly into it, and canceled a $600,000 annual Salesforce contract when the internal tool proved more effective and more tightly integrated into the company’s workflows. The $600,000 went to zero. Separately, a contract agent named Gwen now redlines and signs agreements without routing them to a law firm for review. Each of these represents a G&A function, or a significant portion of one, absorbed by an AI system built in-house. Turner’s conclusion is stated plainly: current-generation models can already handle every back-office task at Curative, with deployment speed as the only remaining constraint.

Today the current gen models can do every back office task we have at Curative. It's just a matter of deploying them. Fred Turner

The deployment point matters. The bottleneck is not capability. If Turner’s description of the current state of the technology is accurate, then the limiting factor at most companies is organizational will, internal build capacity, and the change management required to retire existing vendors and workflows. Those are real constraints, but they are temporary ones.

Steven Pope adds a detail that connects the productivity data to the organizational question. Brand teams, he observes, currently spend roughly 80% of their time on tasks that are mundane and manual, chasing things down rather than doing work that requires judgment or creativity. That figure, if it holds at anything like that magnitude across marketing, finance, and HR, reframes the G&A reduction from a projection about AI capability into a projection about task composition. The question is not whether AI can do knowledge work in general. The question is what fraction of G&A work is process management, and whether that fraction is large enough to justify Arora’s estimate. Pope’s number suggests it is.

The workforce response to that compression may not manifest as mass layoffs in a single visible moment. Adam Mosseri, describing changes already underway at Instagram, points to a new role the company calls product staff: an evolution of the product manager who absorbs work previously distributed across designers, data scientists, and researchers, using internal AI tools to cover the range. The reduction in headcount, in that model, shows up as role consolidation rather than termination. One generalist with the right tools does what a team of specialists previously required. That is a different story than a factory closing, and it is likely to be a harder one to track, because each instance looks like a promotion or a reorganization rather than a workforce reduction.

Dario Amodei, while broadly bullish on AI’s economic effects, argues that mass unemployment at scale does not arrive until 2028 or 2029, after more capable systems are in place. That framing offers a kind of ceiling on the near-term disruption. But it does not contradict the G&A story, which is not about general mass unemployment. It is about a specific class of process-heavy, document-intensive, rules-governed work that current-generation models can already handle. The timeline Arora describes and the results Turner reports do not require superintelligence. They require deployment. That distinction is what makes the three-year window credible, and what makes it consequential for anyone currently planning enterprise headcount.