Scuttlebutt

Scuttlebutt

DevRev's memory layer beat Claude Code by 50% on accuracy using the same model — its CTO

7 insights: why only 28% of CIOs got the AI value they expected, why enterprise data needs converting again, and how DevRev cut token use by 75–80% against Claude Code

Scuttlebutt's avatar
Scuttlebutt
Nov 14, 2026
∙ Paid

Eye On A.I. · Nov 13, 2026 · 37m · 29m saved

With Ahmed Bashir, Chief Technology Officer of DevRev

DevRev's memory layer beat Claude Code by 50% on accuracy using the same model — its CTO

Ahmed Bashir, chief technology officer of DevRev, told Craig S. Smith on Eye On A.I. that DevRev's agent, Computer, scored 50% higher accuracy than Claude Code when both ran the same Opus-family model on DevRev's own benchmark.

The usual explanation for weak returns on enterprise AI is that the model is not good enough. Ahmed puts the gap somewhere else: in the layer between the model and a company's scattered data, which he calls shared memory.

"A more sophisticated model cannot always make up for not having a sophisticated memory layer." — Ahmed Bashir

Ahmed studied databases academically and then spent roughly 12 to 13 years at Apple, where he was the only backend engineer on iCloud backup and restore in the year iMessage and iCloud launched. DevRev, he said, has spent close to a quarter billion dollars building its memory layer.

The full interview is covered here so you can skip it. 37 minutes of audio, 8 minutes of reading.

Here are the 7 insights that matter.


This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.


Key Takeaways

  • Enterprise AI fails after the demo because of governance, orchestration and data meaning, not the model

    • Ahmed cited reports in which only 28% of CIOs saw the value they expected

  • Each new generation of computing has forced companies to convert their data into a new format, and AI is the next one

  • DevRev's agent was 50% more accurate than Claude Code running the same Opus-family model

  • The same agent used 75–80% fewer tokens than Claude Code on the same tasks

  • DevRev keeps ordinary databases for anything that must be exact and saves the language model for judgment calls

  • DevRev has spent close to a quarter billion dollars on its memory layer over about five and a half years

1. Why AI Demos Disappoint

Ahmed's starting point is a pattern he has heard from many companies: AI looks impressive in a demo and falls short once it is deployed.

Enterprise AI disappoints once it meets a real company's complexity

"The problem, and I've heard this from a lot of people, is that enterprise AI demos work amazingly, but once you get them into the enterprise and the complexity of the enterprise, everyone's disappointed." — Ahmed Bashir

He pointed to survey reports from MIT and IDC on the gap between spending and results.

Only 28% of CIOs saw the value they were expecting from AI

"We've seen reports where CIOs are saying that only 28% of them saw the value that they were expecting." — Ahmed Bashir

Nearly every vendor claimed "industry-leading AI," and Ahmed called that a "reductive and misleading premise." Buyers assumed that a vendor using a leading model had leading AI overall. He said the real shortfall is in governance, orchestration and an understanding of what the data means at scale.

2. AI Needs a New Data Format

User's avatar

Continue reading this post for free, courtesy of Scuttlebutt.

Or purchase a paid subscription.
© 2026 The Golden Age LLC · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.