SaaS Says: The Missing Before

The Missing Before

A $3 billion company's marketing team got asked what share of its traffic was coming from LLMs. They guessed 20%. Then 30%. Someone said 50%. Measured, it was 1%.

The gap between those numbers is the year's real story. 2025 went to proving adoption. 2026 is the year somebody asked what the adoption returned, and the honest answer needs a baseline almost nobody wrote down. Stephan Bajaio, who co-founded Conductor and now runs VibeLogic, has the tidiest statement of the problem. Without what he calls the pre, "it's really hard to prove that what you did actually got you the outcome."

Engineering got the same message with a budget attached. LinearB CTO Yishai Beeri published a mid-year benchmark drawn from millions of pull requests, and he is blunt about who is now asking. "Finance is not just in the room." Teams burned annual token budgets in two or three months, and the question flipped from whether developers had adopted AI, which is settled north of 90%, to what leverage each dollar bought. The benchmark itself is bimodal. Per LinearB's data, the top decile of AI users more than doubled merged output year over year while non-users came in flat, with the break landing in January.

The most useful number in the set is a social one. Human-owned pull requests merge at roughly 80 to 90%. Fully autonomous agent loops collapse to about 30%, and Beeri's read is not that the code is worse: "you have detached the human ownership part. You now see a dramatic decrease in yield rates all the way down to 30%." Nobody chases review for an agent's work, because it is nobody's job.

Customer success hit the same wall from the other side. Anika Zubair's case is that CS teams pointed AI at the same self-serving quarterly deck and simply built it faster: "they are using it to do the wrong job faster." Her line for it: "We end up taking a meeting that was already backwards and made it backwards in half the time." The first letter of her framework is B, for baseline.

Charity Majors, who says she stopped being an AI skeptic late last year, wants the reporting to run both ways. Her complaint is not that the wins are fake but that they get published alone. She points at Intercom's own disclosure as the model: "And they showed that for 18 months. Reliability and code quality went down." Couple every win with its cost, or the people carrying the pager keep concluding the wins are invented.

Which leaves distribution. Two weeks ago the labs were selling forward-deployed engineers as the bridge to production. This week Uber's answer, picked apart on Dev Interrupted's Friday show, runs that bridge in reverse: send the engineers who already have the gains backward into legal, marketing, and sales instead of hiring a transformation consultancy. The hosts were unimpressed by the branding. "We used to call it just like inner source."

Sources: Stephan Bajaio (VibeLogic), Metrics that Measure Up, "Web Presence Intelligence with Stephan Bajaio, Co-Founder & CEO, VibeLogic," Aug 12, 2026. Yishai Beeri (LinearB), Dev Interrupted, "The playbook to close your team's AI productivity gap | LinearB's Yishai Beeri," Aug 11, 2026. Anika Zubair (The Customer Success Pro), The Customer Success Pro, "AI For QBR Prep and How To Build A Board Ready Value Story," Aug 12, 2026. Charity Majors (Honeycomb), The Pragmatic Engineer, "Stop being skeptical about AI for development with Charity Majors," Aug 12, 2026. Ben Lloyd Pearson and Andrew Zigler (LinearB), Dev Interrupted, "Telling your agent 'no' is a moat now, rearward deployed engineers, and harnessing the context for your SDLC," Aug 14, 2026.