Aventary Field Guide

The Operator's AI CanonThe sources that actually explain where enterprise AI is going — for the person who has to make it work inside Salesforce.

Everyone has an AI take. Almost none of it is written for the person who has to make it work inside a Salesforce org, next to a pipeline that has to close this quarter. So I keep a running map of the sources that moved how I think — the papers, the launches, the reversals — grouped around ten themes that keep resurfacing. Read the ten picks first. The full list underneath is the receipts.

Mendy Ezagui · Aventary · 70+ sources · updated Aug 2026
The through-line

Ten themes, one argument

Read enough of this and the same ten ideas keep showing up. They're the lens the whole list is organized around.

  • 01Being a good human. AI's impact on roles, jobs, and what's worth doing yourself.
  • 02Multi-model orchestration and routing. No single model wins; picking the right one per task is the skill.
  • 03Embedding AI around the system of record. Agents on top of your CRM, not a rip-and-replace.
  • 04Hallucination, verification, AI checking AI. Trust nothing you can't verify.
  • 05The disappearing interface. Headless platforms and agents running under the workflow.
  • 06Build vs. buy in the age of cheap building. When making it beats buying it.
  • 07Memory, context, and MCP. Agents stop being islands when they can reach your data.
  • 08Automated verification and governance. Feedback loops and control, not vibes.
  • 09The adoption gap. Turning AI on is easy; changing how teams operate is the job.
  • 10Learning that compounds. Skill with these tools builds over months — it's a capability, not a switch.
Start here

If you only read ten

01 · The mental model

The Bitter Lesson — Rich Sutton (2019)

The idea sitting under everything else on this list. General methods that ride compute beat hand-crafted logic, every time. If you're deciding what to build versus buy, start here — it's why the cheap, general approach keeps winning.

Themes 6 · 10
02 · The plumbing

Introducing the Model Context Protocol — Anthropic (Nov 2024)

The most consequential piece of infrastructure for your world. MCP is what turns "agents as islands" into agents that can actually reach your system of record. Everything downstream — Salesforce adopting it, the registry, the foundation, the enterprise-hardened spec — traces back here.

Themes 3 · 7
03 · The moment it landed in your stack

MCP Support Across Salesforce — Salesforce (Jun 2025)

The moment your specific ecosystem stopped fighting the open standard and embraced it. Concrete proof that "embed AI around the system of record" beat "rip and replace." Read it next to Nadella's "SaaS collapse" clip below — this is the answer to it.

Themes 7 · 3 · 2
04 · The threat model

Satya Nadella on BG2, "SaaS collapse" (Dec 2024)

The sharpest version of the threat model to a RevOps job: agents dissolve the business-logic tier and orchestrate across records directly. You need a point of view on this. The Klarna reversal is the counter-evidence.

Themes 3 · 5 · 6
05 · The cautionary tale + the fix

Klarna Reverses Its AI Push (May 2025) + McKinsey's Agentic Adoption Gap (Aug 2026)

Read them as a pair. Klarna turned AI on without changing how the operation ran — quality dropped, they rehired humans. McKinsey's 1:3:5 (a dollar of tech to three of process to five of capability-building) is the prescription. This is the adoption gap with both the failure and the cure.

Themes 9 · 1 · 4 · 8
06 · The playbook

Building Effective Agents — Anthropic (Dec 2024)

The most useful playbook here: simple composable workflows beat heavyweight frameworks, and you add autonomy only when it earns its keep. The antidote to agent hype, and it maps directly to how you'd actually ship.

Themes 5 · 7 · 2
07 · The current face of routing

Opus 5's "effort dial" (Jul 2026) + Willison's "Which AI to Use" (Jul 2026)

Routing is no longer just an infrastructure choice between models — it's inside the model now (an effort dial, mid-task switching), and picking the right one per task has become the core practical skill. The freshest "no single tool wins" evidence.

Themes 2 · 6 · 10
08 · The hard data on skill

Anthropic Economic Index: Learning Curves (Mar 2026)

Rare hard data that skill with these tools compounds over months, and experienced users route their highest-value work to the strongest models. Your evidence that AI fluency is a durable capability, not a one-time switch you flip.

Themes 10 · 1 · 2
09 · How AI really changes work

Navigating the Jagged Technological Frontier — Dell'Acqua, Mollick et al. / BCG (Sep 2023)

The empirical bedrock: big gains inside the frontier of what AI is good at, confidently wrong outside it. The study every serious adoption-and-verification argument cites — foundational for how you scope where AI helps and where it hurts.

Themes 9 · 1 · 4
10 · Why verification is a governance problem

Why Language Models Hallucinate — OpenAI (Sep 2025)

The best current source on verification: hallucination sticks around because evals reward confident guessing over "I don't know." That reframes verification as an incentive and governance problem — which is exactly the lever an operator can actually pull.

Themes 4 · 8
The full map

The complete list

Sorted oldest to newest, so you can watch the arguments build on each other. Every item links to the primary source. Where a number is contested — the MIT "95% of pilots fail" stat especially — I flag it and treat it as directional, not gospel.

The map is free. The build is the hard part.

This is the reading. Turning it into a working operating system inside your Salesforce org — where the agents actually reach your records, the handoffs hold, and the metrics tell you what to build next — is the work Aventary does.

Start with the free Operating Systems Diagnostic, or book a working session and we'll map your weakest link.

← More Insights