06-reference

every executive guide implementing ai

2026-06-07·reference·source: Every·by Natalia Quintero
every-greatest-hitsai-adoptionorganizational-changeai-championsevalsphdataenterprise-agents

"An Executive's Guide to Implementing AI" - Natalia Quintero (Every Consulting)

Why this is in the vault

This is the clearest published version of the org-adoption playbook phData sells: a five-step loop and a 60-day plan from Every's consulting head, built on engagements at the New York Times, Ripple, Headway, Thumbtack and several investment firms. It was first seen in the vault as a one-line mention in the 2026-06-07 Sunday digest ([[2026-06-07-every-ai-ready-organizations-arent]]). This note is the full-text digest from the paid re-read.

The core argument

Quintero's claim: the bottleneck has moved from model capability to organizational capability. She cites McKinsey's "AI high performers" (more than 5% EBIT impact). Only 6% of roughly 2,000 surveyed organizations qualify, and they are nearly three times as likely to have redesigned workflows. She also uses METR's task-length curve: about 4 seconds of human work in 2022, about 6 minutes for GPT-4 in mid-2023, about an hour for o1-preview in late 2024, and 10+ hours for Claude Opus by late 2025.

Three adoption waves:

  1. License wave (late 2022 to early 2024): companies bought seats, and gains were uneven and individual.
  2. Prompt wave (2024 to mid-2025): prompt trainings, prompt libraries and custom GPTs. These had no owner and no evaluation, so they rarely stuck.
  3. Implementation wave (mid-2025 onward): Claude Code moved the frame to agents. "Prompt libraries are giving way to skills libraries," meaning reusable workflows with instructions, examples, reference material, scripts, eval criteria and named owners.

The loop: Get fluent → Assign AI champions → Pick one painful workflow → Build to 95 percent → Scale what works.

  1. Get fluent. Executives must build something themselves: a skill, an agent or an automation, not chat. One media/data company's reviewers had never built anything, and all of their initiatives had failed. Building exposes the real constraints: connectors, data access, and IT/security policy. Without this, leaders "misread" low adoption as employee reluctance when it is really an access problem. Fluency also tests whether a leader can define excellence. In the guide's case example, an executive wanted to automate a board-metrics pull from Snowflake but could not say what "excellent" looked like. Six questions to ask: What can AI see? What can it do? Where are the constraints? Which workflows are painful? Who owns what we build? How will we know it works?
  2. Assign AI champions. Choose for curiosity, people orientation, and authority plus time, not for technical depth. Champions should sit close to the painful workflow. The most common failure is making the role a side project; the guide asks for at least 2 protected days per month and a clear mandate. Champion charter: own 1-3 workflows, maintain their docs, build or manage eval sets, collect feedback, update skills when models change, and report time saved and errors reduced. Structure varies: an ambassador model for large companies, 1-2 per department for mid-size, and rotating "fellows" for PE and holding companies.
  3. Pick one painful workflow. Don't start with the board deck. Every's own first instinct was to automate consulting project management, which was too broad. Build "one artery," not "the whole body." Score candidates on seven criteria: frequency, pain, data availability, risk, ownership, evaluation clarity, maintenance burden.
  4. Build to 95 percent. "An automation that works 80 percent of the time is a demo." Getting from 60% to 95% takes gold-standard examples, structured evals, human review gates and a named owner who maintains the workflow across model updates. The section is headed "Automation is a lie": treat the agent like an employee you keep onboarding. The guide gives a per-workflow eval table: real test example, current output, expected output, errors, cause, prompt/skill change needed, retest result, human review needed, owner, review cadence.
  5. Scale what works. One visible win creates pull; a company-wide mandate does not. Most workflows are department-specific, so for each win decide: shared skill, team-specific, or kill. Pre-scale checklist: real pain solved, tested on real examples, named owner, review process, risks understood, team can explain it, feedback loop exists.

60-day plan: Weeks 1-2 fluency plus IT/security mapping. Weeks 3-4 champions and a workflow shortlist. Weeks 5-7 build and eval. Weeks 8-9 scale or kill. The promised outcome is modest on purpose: one reliable workflow, trained champions, and a repeatable process.

Cost disclosure: Quintero spent 100+ hours in January with Every's internal champion (a forward-deployed engineer). She now spends 10-15% of her time maintaining skills and giving feedback to agents. In return, Every has a skill library and an agent doing roughly one FTE of PM, sales-ops and delivery support (Claudie; see [[2026-08-14-every-ai-employee-security-framework]]).

Mapping against Ray Data Co

Step 4 is the copilot agent factory's design spec, described from the buyer's side. Ben's factory already has stateless skills, document-tracked state, eval plans and human review gates. That matches Quintero's "skills library with named owners and evals" almost line for line. The gap the guide exposes is ownership. The factory's 50+ skills have evals, but the guide's per-workflow table also assigns an owner and a review cadence per skill. A skill with no named maintainer is exactly how a skill library drifts back into a prompt library.

Why it matters for RDCO / The Denominator

⚠️ Sponsorship

House promo: this guide is written by the head of Every Consulting and functions as top-of-funnel for that practice. It ends with "no outside firm can implement AI for you" while showing its own client roster. Named outcomes (100+ agents at an investment firm, a week of variance analysis done by Opus, a PE firm hiring champions) are self-reported and unaudited. The method is sound and matches independent practice; treat the impact anecdotes as marketing. sponsor_entity: self.

Related