Vitaly Gordon, Faros AI's CEO, argues most companies spend less than 5% of labor cost on AI, and that teams should prioritize outcomes over token volume.
Gordon says the goal is outcome maxing, lowering the price per outcome, which often means using more AI rather than less.
Gordon says teams must rethink code review, CI/CD and QA because removing one bottleneck shifts pressure to the next, with Faros data showing 76% more changes merging with zero review.
Faros AI's CEO says pay for AI outcomes, not tokens
AI has turned engineering productivity into a question of how well teams use AI, and Vitaly Gordon thinks most companies measure and budget it wrong. Burning tokens or generating code does not create value. Companies should spend more on AI, direct it to teams that turn dollars into shipped features, and fix the review and QA bottlenecks that faster coding exposes.
Takeaways
Optimize price per outcome, not token spend. Gordon rejects token maxing because spending tokens does not by itself yield results. Outcome maxing means lowering the price of each outcome, which often requires using more AI. Lines of generated code are a poor measure. Features reaching production are what count.
Most companies still underspend on AI. In most companies AI spend remains under 5% of labor spend. Gordon cites a claim that Anthropic's Claude Code spend now exceeds its salary spend. Companies not executing at this pace are probably being out-executed by competitors.
Faster coding moves the bottleneck downstream. Gordon calls it the speed trap. Raising throughput pushes problems down the road, and Faros data shows code churn of 861% and QA time up 300%. Anthropic shipped fast enough to break GitHub. There are no free meals, so teams must reinvest in their software development lifecycle.
Zero-review merges signal reviewer fatigue. Faros found 76% more changes merging with zero review. Gordon says many were not truly unreviewed but came from people tired of reviewing, which leads to more production issues. Some open source projects refuse AI contributions, but large companies cannot refuse code.
Allocate AI budgets by results, not headcount. Gordon calls equal per-engineer budgets silly. Teams that turn each dollar into many features should get more, monthly quotas need not be exhausted, and allocation should be automated. One Faros customer raised outcomes two and a half times while its AI budget fell about 20%.
Half of AI spend sits outside engineering. In many organizations about half of AI spend falls outside engineering, Gordon says. The same outcome and cost principles apply to marketing, sales, finance and HR. Faros now serves those functions alongside CTOs, CFOs, developer productivity teams and AI budget owners.
AI will add engineers, not eliminate them. Gordon completely disagrees with job-loss predictions. AI-native leaders like OpenAI and Anthropic keep investing profits in hiring, and AI is still a small share of labor spend, so companies should use more AI alongside humans.
How Faros tests models and workflows
Choosing open-weights or frontier models depends on the task. With releases arriving almost daily, Faros builds private benchmarks and routes each task to a suitable model.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Gordon expects most companies to stay with frontier labs because they trust them and it is easier than managing compute.
Faros makes staff use different harnesses and models, and bought machines to run open-source models in its office to understand regulated customers running on bare metal.
The most sophisticated users spend their time on context and evaluations. Gordon, mostly a Codex user, says Terra 6.1 runs about three times faster than other models and may change how he works.
Agents still fail, deleting databases in YOLO mode and adding code nobody asked for.