AMOTION
Resources

Read this before we talk.

No gated PDFs. No webinar signups. Two things worth your time: the ten-minute pre-read we send every prospect before a first call, and the public research behind every problem number on this site.

The pre-read · 10 minutes

Are we seeing this right?

Five problems we think AI coding created. No pitch — we want you to tell us where we’re wrong. Every number is sourced below; if one is wrong, that’s exactly the kind of thing we want to hear.

This is the note we send before a first call, reproduced in full. Each problem ends with the question we’d ask you about it.

Problem 1

Your team ships PRs faster than anyone can review them

AI made writing code fast. It did nothing for reviewing it. Where teams went heavy on AI, median review duration rose 441% and the per-developer defect rate went from 9% to 54% where review collapsed. Nearly half of agent-written fix PRs — 46% — get closed without merging. And in the only controlled trial, experienced developers believed they were 20% faster with AI; measured, they were 19% slower(the study’s own 2026 follow-up is inconclusive, so we cite it as a caution, not a law). The reviewer’s real question — what else does this change break? — is answered by nobody: not the AI that wrote it, and often not the reviewer either.

Faros AI, 22,000 devs / 4,000 teams · via Osmani, Jun 2026 — arXiv:2606.13468 — METR RCT, arXiv:2507.09089, Jul 2025

Last week, how many PRs sat in your queue more than 2 days? Has a “looked fine” AI change broken production?

Problem 2

On your legacy system, AI agents make things worse, not just less helpful

On greenfield, agents shine. On a 10-year-old codebase, the rules that matter live in old postmortems, Slack threads, and the head of someone who resigned. Benchmarks confirm it: frontier models solve 65%+ of curated bug tasks (SWE-Bench Verified) but only ~23%on the harder, contamination-resistant SWE-Bench Pro built from real professional repos. Teams report having “to diff through the code base the old way” to confidently accept or reject what the agent claimed.

SWE-Bench Pro, Scale AI · arXiv:2509.16941

Have you let agents touch your core or legacy system? What happened?

Problem 3

Nobody can say how much of the codebase the team still actually understands

51.9% of merged code is now AI-authored — nearly doubled in a year (self-reported; as a floor, Google said back in early 2025 that more than 30% of its new code was AI-generated,). Code churn has roughly doubledsince AI assistants arrived (3.3% → 7.1%), and copy-pasted code now exceeds refactored code for the first time in that dataset. Call it comprehension debt: the system moves, tests pass, and fewer people can explain how it works. In a regulated industry, “the AI wrote it and we didn’t fully review it” will not survive a post-incident report.

DX, Q2 2026 · 400+ companies · self-reported — GitClear, 200M+ changed lines · vendor analysis

If your best engineer left tomorrow, which modules would nobody understand? Is that list growing?

Problem 4

Your AI spend has no ceiling and no link to outcomes

Uber burned its entire 2026 AI budget in about four months and imposed a $1,500/tool/month cap . Its COO, asked whether the spend connects to shipped value: “That link is not there yet.” 73% of enterprises exceeded their AI budgets in a year when token prices fell 67%. Gartner predicts AI coding costs will surpass an average developer’s salary by 2028— and most teams can’t say today whether they’re a heavy user or a typical one.

Fortune, May 26 2026 · TechCrunch, Jun 2 2026 — EMA 2026 (industry survey) — Gartner press release, Jun 24 2026

Do you know your AI spend per team last month? Could you connect it to anything that shipped?

Problem 5

Individual adoption is high; the company hasn't changed

The median company-level throughput gain from AI is +7.8%— while individual output multiplies. MIT’s 2025 enterprise study found 95%of GenAI pilots showed no measurable P&L impact (preliminary, 153-response sample, not peer-reviewed — we cite it with that caveat), and DORA’s ~5,000-respondent 2025 report found the same shape: individual throughput up, company-level delivery flat. Roughly 90%of employees use AI daily — mostly personal tools, around governance, not through it. Boards are now asking what the spend bought, and engineering leaders don’t have a number.

DX “The AI efficiency plateau” · May 2026 · 500+ companies — MIT NANDA 2025 · preliminary, 153-sample — DORA 2025 · ~5,000 respondents

If your board asked “what did our AI investment change?” — what would you show them?

What we do about it

Timo checks every AI change against what you asked for before it merges, writes memory and docs your agents can actually use, and gives leadership a scoreboard with frozen baselines. Free to start, everything in your repos. The full story is on the product page— but on a first call we don’t demo. We ask the four questions below.

Tell us where we’re wrong.

These four questions are the point of the call. Your answers shape what we do next — including “nothing.”

Q1

Which of these 5 is a real, current, painful problem for you — and which is overblown?

Q2

Which one would you spend money to fix this quarter?

Ranking, not politeness.
Q3

What problem did we missthat’s bigger than all five?

Q4

What are you doing about the painful one today — tool, process, person, or nothing?

— Mahesh · founders@amotion.ai

Sources

The research behind every problem number.

One canonical figure per claim — never rounded, never blended. Where a source is company-published, it carries a visible vendor flag and we treat it accordingly. Last checked 2026-07-13.

DORA, State of AI-Assisted Software Developmentsurvey
2025
~5,000 respondents. Individual throughput up, company-level delivery flat — and AI adoption shows a measured negative relationship with delivery stability.
DX, “The AI efficiency plateau”
May 2026
500+ companies. Median company-level throughput gain from AI: +7.8% — while individual output multiplies.
DX, AI-authored-code measurement
Q2 2026
400+ companies, self-reported: 51.9% of merged code is AI-authored — nearly doubled in a year.
Faros AI telemetry, via Addy Osmani
Jun 2026
22,000 devs / 4,000 teams: median review duration +441%; per-developer defect rate 9% → 54% where review capacity collapses; zero-review merges up 31.3%.
Stack Overflow Developer Surveysurvey
2025
n≈49k: 66% name AI output that’s “almost right, but not quite” as the top frustration; 45.2% say debugging AI code takes longer; 84% use or plan to use AI, only 33% trust its accuracy.
GitClear, code-churn analysisvendor
2025
200M+ changed lines: churn roughly doubled since AI assistants arrived (3.3% → 7.1%); copy-pasted code exceeds refactored code for the first time in their data.
GitLab surveyvendor
2026
85% say the bottleneck moved from writing code to review and validation.
EMA, enterprise AI-budget study
2026
73% of enterprises exceeded their AI budgets in a year when token prices fell 67%.
Gartner press release
Jun 24 2026
AI coding costs will surpass an average developer’s salary by 2028; nearly a quarter of tech leaders already spend $200–500 per developer per month.
Jellyfish, State of Engineering Managementvendor
2026
600+ leaders: only 46% of companies track any AI-specific metrics at all.
arXiv:2606.05647
Jun 2026
94% of developers failed to detect agent-inserted sabotage in code review; 56% accepted it even after a monitor flagged it.
arXiv:2604.22750
Apr 2026
Agent-driven coding consumes ~1000× the tokens of single-shot use, with 30× variance on the same task.
arXiv:2601.14470
Jan 2026
59.4% of tokens go to the agent reviewing its own work.
arXiv:2606.13468
Jun 2026
AIDev dataset: 46% of agent-written fix PRs are closed without merging.
arXiv:2509.16941
Sep 2025
SWE-Bench Pro (Scale AI): frontier models solve 65%+ of curated bug tasks but ~23% of tasks built from real professional repos.
arXiv:2605.17526
May 2026
SaaSBench: >95% of agent failures on enterprise systems occur before the agent even reaches business logic.
arXiv:2606.01772
Jun 2026
Industrial study: “tool integration, not tool capability, is the binding constraint” — benefits accrue to individuals while team processes stay limiting.
The other side

We also cite the studies that cut against us.

The METR trial — and its inconclusive follow-up

The only randomized controlled trial of AI coding tools found that experienced developers who felt 20% faster were measured 19% slower. That result argues for caution about every speed claim in this industry — including ours. METR’s own February 2026 follow-up is inconclusive, and we say so every time we cite it. Same discipline for MIT’s “95% of pilots showed no P&L impact”: preliminary, 153-response sample, not peer-reviewed — it never appears here without that caveat. Reading the counter-evidence is how we’d want to be sold to.

METR RCT, arXiv:2507.09089, Jul 2025 · Feb 2026 follow-up inconclusive — MIT NANDA 2025 · preliminary, 153-sample
What to read next

Done reading? Three places to go.

Get started

The next thing to read is your own repo.

Five minutes to your first Readiness Score. Read-only, nothing leaves your network.