Note

The most revealing line
in Amazon's AI memo.

Published

Colin Whitfield · Founder and Operating Partner · 2 min read

Amazon fixed the AI inputs. Delivery is the next gap. A quote from Terk Terkowitz, VP and Distinguished Engineer at Amazon, on tokenmaxxing and lines of code; the input fix (AI usage is a gauge, not a target; engineers move closer to customers); and the output gap: 94% rate AI code higher at review, 82% had a production failure from it (New Relic, 2026).

Everyone is quoting Amazon's warning against "tokenmaxxing" this week.

But the most revealing line in Amazon's new engineering memo isn't about the engineers avoiding AI.

It's about the ones using it the most.

VP & Distinguished Engineer Terk Terkowitz tells managers to inspect the 90th percentile of AI adoption—the teams burning the most tokens and generating the most code changes—and check "whether they're seeing delivery increases to match."

Why wouldn't heavy AI usage automatically translate into faster delivery?

Because AI made generating a working version cheap. It did not make reviewing, hardening, and maintaining it in production cheap.

Look at the gap in New Relic's 2026 State of AI Coding survey (200 US tech decision-makers):

  • 94% say AI-generated code scores higher at initial code review.
  • 82% report production failures caused by AI code in the last six months.

When you give agentic AI to teams close to the customer, they can spin up a slick prototype in days. On screen, and in the review queue, it looks cleaner than it is.

If you don't put a hard architectural fence around those experiments, one-off prototypes get quietly promoted into the shared core product. Rework compounds, releases slow down, and the R&D budget quietly fills up with custom maintenance—a dynamic buyers punish in diligence (1–2.6x revenue for services-heavy software vs. 3–3.8x for vertical SaaS, per Aventis Advisors).

Amazon is right: token use is a gauge, not a goal.

But the real bottleneck in 2026 isn't how much AI your team uses to build a prototype. It's having a clear gate—and a named technical lead—to decide whether that code gets Adopted, Refactored, or Rebuilt before it ever touches the core.

Where does your team draw the line between a fast AI prototype and production-ready code?

Sources

The full mechanism — how absorbed custom work prices as a services business, and how the shape is resolved — is in the report →

Where this document sits

Where this note leads: the papers behind it, in reading order.

01The ThesisWhy the discount exists, and how to detect itRead the report →02The Scaling LadderWhere it forms, stage by stageRead the report →03Portfolio OSThe operating model and the mandateRequest the operating model →04The Modelled CaseThe arithmetic, entry to exitRequest the modelled case →

Send your numbers.

The six indicators and your gross churn. We return the modelled bridge, run on your P&L, within two weeks.

Send your numbers →