Rippling found one engineer burning $50,000 a month on AI, then cut token spend to 37%

Six months ago the pitch was "use AI to move faster." Now the tools have caught up, and they come with receipts. You can measure who is actually getting faster, cap what you spend doing it, and put a human gate on the parts that still need judgment. The vague promise turned into a kit you can pilot on Monday. Let me catch you up.

The deep cut

• Speed without measurement is just spend. Rippling burned 40% of its R&D headcount budget on tokens before it built a way to see it.
• Build gates that tell you no. Oji Udezue's PM skills kill a feature idea that fails three questions, not one.
• Route the review, don't skip it. Claire Vo's Merge Mommy auto-approves low-risk PRs and pings a human for the rest.

The bill nobody was reading

Rippling went all in on AI at the start of the year, same as a lot of teams. Then the CFO put a number in front of the executive team that stopped the room. The company was on track to spend as much on tokens as 40% of its entire R&D payroll. Growth was 80% month over month. Chief Product Officer Matt MacInnis told TechCrunch they were "incredulous."

The real find was who was spending. Roughly 10 to 15% of employees drove about 60% of the cost. One engineer was burning $50,000 a month, often defaulting to the priciest frontier model for simple tasks.

So Rippling built a tool to watch it and cut spend from 40% of headcount budget to about 15%, without cutting usage. Same 600 billion tokens in July as at the peak, at 37% of the cost, because prompts now route to cheaper models that do the job.

The skill that says no

The other half of the story is judgment. Oji Udezue, a three-time CPO, says he keeps seeing PMs speed up their tools but not their product judgment. His fix is a set of open-source PM skills built around one idea: gates that refuse.

Take his vet-a-feature skill. It runs three questions. What do people do today instead? How many have the problem, and how often? Will anyone pay? Fail one and the idea is dead. Then a bar: 3x better than the workaround or nobody switches. Most AI tools never tell you no. This one does.

That is the point for your team. The scarce skill is not building. It is picking this one and not that one, and holding that call yourself.

Put a gate on the flood

When AI writes the code, review becomes the jam. Claire Vo built Merge Mommy, an agent that reads every PR after checks pass, scores it across six risk dimensions like blast radius and reversibility, auto-approves the low-risk ones, and pings a human in Slack for anything real. She cites Intercom, which 5x'd PR approval speed and cut revert rates by putting AI in the review loop.

The safety part matters. Auto-approved PRs can stay SOC 2 compliant as long as the process is auditable and written into your risk policy. You are not skipping review. You are spending human attention on the PRs that earn it.

The plumbing gets AI-shaped

The workflow tools are shifting the same way. Wispr Flow, which people already use for voice-to-text drafting, launched a Granola-styled meeting notetaker that transcribes off system audio, writes action items, and lets you query past meetings. Content teams are wiring the same loop: a radar agent that scores ideas against a brand rubric so you rank the pile instead of adding to it.

Further out, Replit's Amjad Masad says his engineers nearly tripled code shipped per person in six months, and that in three years we will use fewer apps and more agents that use the apps for us. Believe the direction, not the timeline. The near-term move is smaller: pick one workflow and instrument it.

Three questions for your team

• Do we know which people and tasks drive our AI spend, and can we route cheap tasks to cheap models? If not, that is the first pilot.
• Where in our process do we need a gate that says no, and who owns that call when the tool won't?
• If we put AI in the review loop, is the approval trail auditable enough to defend in a compliance review?