Four patterns and fifteen practices from 25 Companies that rolled out AI across their business beyond simple LLM chat.

How this was built. We read 25 Claude customer stories and interviews published on claude.com/customers between June and August 2026. A practice made the list only if it appeared at three or more unrelated companies and a ten-person business or team could run it without engineers.

Read it with three caveats. Every source is vendor-published, so results are self-reported and the sample contains no failures. Three sources are engineering interviews; we kept only what transfers to office work. And the companies range from a two-person fabrication shop to 15,000-person firms, so the practices are the signal and the numbers are directional.

Four Patterns

1. Buy fluency with leaders' time, not employees' courses. No company here credits a training curriculum. They credit executives who used the tool before anyone was asked to, and first sessions engineered so every participant left with a finished output. Advantage Solutions trained its leadership team before any employee; Epic gave up a full leadership meeting to hands-on use; PwC walked 400 people through three real tasks in an hour. Fluency is bought top-down and hands-on, and the first metric is leadership hours in the tool.

2. Make usage visible, including the failures. Adoption spread through a bazaar, not a mandate. Champions were found by watching who shared unprompted. Written records of wins, failures, and wasted budget were the learning channel that scaled. League's COO posted overnight results in Slack and access requests followed; Spotify runs a prototype store its co-CEOs contribute to. Visibility also carried the message that mattered most: hours returned to people, named task by task, not headcount removed.

3. Set the fence, then get out of the field. The companies that moved fastest did not micromanage use cases. They built a fence with three rails: red lines on where AI never goes and who owns the final call, access scoped by role before anything was connected, and a hard date. Inside the fence, teams had what Cox calls carte blanche. Cyera mapped its data first and then went company-wide in 17 days; Mercy Corps assesses privacy per use case because the risk lives there, not in the tool.

4. Start small, capture everything, compound. The first step is small: the task people dread, a build that takes twenty minutes. The tempo is not slow. What turns small steps into organizational capability is capture. The best performer's method becomes a shared skill, the context becomes a file the AI reads every time, the failed idea is retried on the next model, and the personal automation graduates to team ownership. Brainlabs authored 400 skills in four weeks; YMCA South Australia's skills now outlast staff turnover. Without capture, week one feels magical and week five disappoints.

The carousel below gives fifteen concrete practices behind these four patterns, with the evidence and sources for each.

The 15 practices, grouped by pattern
FluencyPractice 1 of 15

Adoption follows leadership hours in the tool, not leadership endorsement of it

What to do
  • Owner and every manager complete one real task in the tool before rollout
  • Block a full leadership meeting for hands-on use, not a demo
  • Track leadership hours logged as the first adoption metric
Evidence
  • Advantage Solutions: executives trained first, then senior leaders in a 3-hour session; 98% of use cases surfaced after
  • Epic: one leadership meeting converted to hands-on use; over half of usage now non-developers
  • Quantium: CEO built his own advisory-board tool; 91% weekly usage across 1,200 staff
Sources: claude.com/customers/advantage-solutions, /epic-systems, /quantium-qa, /cox-communications-qa
FluencyPractice 2 of 15

First sessions convert when every participant leaves with a finished output

What to do
  • Remove setup friction before the session (accounts, connectors, installs)
  • Run a "day in the life": two or three tasks people actually do
  • Build alongside the trainer; no watch-only demos
  • Follow with weekly office hours
Evidence
  • PwC: 400 people produced a proposal, a spec, and an app in one hour
  • Cyera: 20 department sessions, twice-weekly office hours, 18-hour livestream; 88% weekly usage in a month
  • Blank Metal: 700+ trained using a build-along companion
  • HubSpot: a connected dashboard as the first-win exercise
Sources: claude.com/customers/pwc-qa, /cyera-qa, /blank-metal-qa, /hubspot-qa, /advantage-solutions
FluencyPractice 3 of 15

Output quality tracks the context loaded before the first prompt

What to do
  • Write a one-page context file: who you are, goals, house style, definitions
  • Connect the tools you already live in before asking for anything
  • Feed messy real inputs (photos, spreadsheets, notes), not clean questions
Evidence
  • HubSpot: "week one magical, week five disappointing" without scaffolding; first drafts 60–80% usable with it
  • Miro: a Project holding strategy, OKRs, and per-channel writing style
  • Epilepsy Foundation: six months of context turned a grant into a one-day job
  • Shy Bird: photos plus sales data produced a station-level diagnosis
Sources: claude.com/customers/hubspot-qa, /miro-qa, /epilepsy-foundation-qa, /shy-bird, /bla-studios
VisibilityPractice 4 of 15

Broad access outperforms hand-picked pilots because champions cannot be predicted

What to do
  • Give every employee access and a usage budget on day one
  • Watch who shares unprompted; name one champion per department
  • Maintain a visible list of who uses AI for what
Evidence
  • DoorDash: tokens for all 4,000 staff; champions self-identified by enthusiasm
  • Epic: internal directory of AI users accelerated copying across teams
  • Advantage Solutions: 50+ champions emerged from a 150-person pilot
  • Cyera: shadow AI treated as a signal of innovators; sandbox provided
Sources: claude.com/customers/doordash-boris-office-hours, /epic-systems, /advantage-solutions, /cyera-qa, /miro-qa
VisibilityPractice 5 of 15

Teams adopt when AI is positioned as reclaimed capacity and the hours are named

What to do
  • State in writing what AI is for: freeing people for the work they were hired to do
  • Track hours redirected on named tasks, not blended percentages
  • Publish the task, the before, and the after
Evidence
  • Advantage Solutions: 70,000 hours a year from paperwork to training and floor presence
  • Mercy Corps: 540 hours saved in one month across 29 tracked tasks; staff's loudest complaint was fear of losing the tool
  • RAINN: headcount framing named as the fastest way to lose the team
Sources: claude.com/customers/advantage-solutions, /mercy-corps, /rainn, /cyera, /ymca-south-australia
VisibilityPractice 6 of 15

Written records of wins and failures distribute learning faster than demos or all-hands

What to do
  • One shared channel: what worked, what failed, what wasted budget
  • Leaders post their own automations and results
  • Keep a place to share working prototypes, not slides
Evidence
  • DoorDash: written artifacts named the most scalable learning channel; failures included
  • League: COO's overnight tasks posted in Slack; access requests followed
  • Spotify: internal prototype store; co-CEOs contribute
  • HubSpot: PM channel filled with demos after rollout
Sources: claude.com/customers/doordash-boris-office-hours, /league, /spotify-boris-office-hours, /hubspot
The fencePractice 7 of 15

Standards enforced inside the instruction scale; standards enforced by end-of-line review do not

What to do
  • Require a linked source for every factual claim
  • Add a completeness gate before work enters any queue
  • Run a quality pass on each skill before team-wide release
Evidence
  • Miro: claims must trace to a Slack message, doc, or page; skills checked for sourcing and data boundaries
  • Cyera: gatekeeper agents catch missing files before tickets submit; 12 hours a week saved
  • Epic: early summaries linked to the record built trust for later autonomy
Sources: claude.com/customers/miro-qa, /cyera-qa, /epic-systems, /epilepsy-foundation-qa
The fencePractice 8 of 15

Explicit red lines and a named human owner are prerequisites for trust, not constraints on it

What to do
  • Write the categories AI never handles
  • Assign a person to approve anything external or irreversible
  • Reframe sensitive tasks from "write it" to "review it and coach me"
Evidence
  • RAINN: AI never in front of the survivor; its job is speed to a human
  • NDWA: immigration questions routed to trusted humans; worker councils set red lines
  • YMCA South Australia: risk-assessment skill returns coaching questions, not the assessment
  • Mercy Corps: AI proposes a grade; staff confirm and own it
Sources: claude.com/customers/rainn, /national-domestic-workers-alliance-qa, /ymca-south-australia, /mercy-corps, /miro-qa
The fencePractice 9 of 15

Scoped permissions done first remove the fear that otherwise delays connected workflows

What to do
  • Map which systems AI will touch and what sensitive data sits there
  • Scope access by role; add shared definitions for key metrics
  • Short privacy assessment per use case, not per tool
  • Then connect; scoped access beats workarounds
Evidence
  • Cyera: data mapping and role-level security before a 17-day rollout to 1,500 staff
  • Mercy Corps: privacy impact assessment per use case; one provider shortened review
  • YMCA South Australia: personal use, then SSO, then skills and an acceptable-use policy
  • Miro: fear of connecting email lifted once permissions were understood
Sources: claude.com/customers/cyera, /mercy-corps-qa, /ymca-south-australia, /miro-qa, /quantium-qa
The fencePractice 10 of 15

A hard constraint and a fixed date surface the real blockers that open-ended pilots hide

What to do
  • Pick a number (3x faster, live in two weeks) and a date
  • Run a 48-hour event on real backlog items
  • Give sponsorship to remove process blockers teams raise
Evidence
  • League: signed Friday, live Monday; 48-hour event reset expected pace; cycle time halved
  • DoorDash: 3–5x faster challenge; a quarter-long migration done by one person in three weeks
  • Brainlabs: one-week target, four-week rollout, ~400 skills
  • Cyera: company-wide in 17 days
Sources: claude.com/customers/league, /doordash-boris-office-hours, /brainlabs, /cyera
CompoundingPractice 11 of 15

The first automation should target the most resented recurring task, not the most visible one

What to do
  • List tasks done more than three times a month; rank by resentment
  • Automate the top one; time-box the first build to under an hour
  • Defer anything requiring judgment or creative taste
Evidence
  • Miro: influencer-content review; on-time delivery rose from 64% to 99%
  • Brainlabs: "more than three times" rule produced ~400 skills in four weeks
  • Blank Metal: lead processor built in 20 minutes; saves hours daily across four people
  • Shy Bird: slow Sunday brunch diagnosed and fixed the same week
Sources: claude.com/customers/miro-qa, /brainlabs, /blank-metal-qa, /shy-bird, /mercy-corps-qa
CompoundingPractice 12 of 15

Encoding the best performer's method as a skill makes quality independent of who is on shift

What to do
  • Pick one process with variable quality across people
  • Interview the best performer; write the steps and checks as a skill
  • Name it by the exact task it covers so it gets found
Evidence
  • Cyera: outreach skill built from top performers' best intros
  • Brainlabs: SEO audit skill refined over thousands of iterations; quality no longer strategist-dependent
  • YMCA South Australia: 20+ skills provide continuity through staff turnover
  • Vercel: a non-coder marketer authored top-100 public skills
Sources: claude.com/customers/cyera, /brainlabs, /ymca-south-australia, /vercel-qa, /blank-metal-qa
CompoundingPractice 13 of 15

Skill libraries hold value only with open access, named owners, and routine pruning

What to do
  • One shared library anyone can read
  • Every skill has an owner who approves changes
  • A suggest-a-skill path for any employee
  • Quarterly audit for duplicates, contradictions, and bloat
Evidence
  • Brainlabs: auditor skill built after versions 2–4 of the library became unwieldy
  • Miro: overbuilt, unmaintained skills reduce reliability; keep each piece simple, then chain
  • Cox: "before you build, talk to the team that built it"; ten valuable use cases over a thousand
Sources: claude.com/customers/brainlabs, /miro-qa, /hubspot-qa, /cox-communications-qa
CompoundingPractice 14 of 15

Twenty real cases scored against human decisions reveal both model gaps and process gaps

What to do
  • Pull 20 past cases with known human outcomes
  • Run them; score by how much human correction was needed
  • Investigate every disagreement before trusting the workflow
  • Monitor edit rates: 0% and 100% acceptance are both red flags
Evidence
  • Mercy Corps: 75% agreement; disagreements exposed rubric ambiguity and some wrong human grades
  • HubSpot: tools benchmarked on real tasks by least intervention to finish
  • League (counterpoint): skipped the bake-off; committed and went deep
Sources: claude.com/customers/mercy-corps, /hubspot, /epic-systems, /league
CompoundingPractice 15 of 15

Retrying failed ideas on new models and graduating personal tools to team ownership compounds returns

What to do
  • Keep a list of attempts that failed; re-run on each major release
  • When a personal automation becomes load-bearing, move it to a shared, owned version
  • Feed the same workflow feedback weekly; it improves each run
Evidence
  • DoorDash: ideas that failed months earlier worked on the next model
  • Cyera: formal "graduation" from one person's agent to a business-owned one
  • Epilepsy Foundation: weekly report improves with weekly feedback
  • Blank Metal: failing task rebuilt as a project on the tool's own suggestion
Sources: claude.com/customers/doordash-boris-office-hours, /cyera-qa, /epilepsy-foundation-qa, /blank-metal-qa

Sources

All 25, September 2026.

1 Miro, Q&A claude.com/customers/miro-qa
2 Cyera, Q&A claude.com/customers/cyera-qa
3 Cyera, case study claude.com/customers/cyera
4 Mercy Corps, Q&A claude.com/customers/mercy-corps-qa
5 Mercy Corps, case study claude.com/customers/mercy-corps
6 Epilepsy Foundation, Q&A claude.com/customers/epilepsy-foundation-qa
7 National Domestic Workers Alliance, Q&A claude.com/customers/national-domestic-workers-alliance-qa
8 League, case study claude.com/customers/league
9 League, Q&A claude.com/customers/league-qa
10 Epic Systems, Q&A claude.com/customers/epic-systems
11 DoorDash, Office Hours transcript claude.com/customers/doordash-boris-office-hours
12 Spotify, Office Hours transcript claude.com/customers/spotify-boris-office-hours
13 Blank Metal, Q&A claude.com/customers/blank-metal-qa
14 PwC, Q&A claude.com/customers/pwc-qa
15 Brainlabs, case study claude.com/customers/brainlabs
16 Cox Communications, Q&A claude.com/customers/cox-communications-qa
17 Advantage Solutions, case study claude.com/customers/advantage-solutions
18 Shy Bird, small-business story claude.com/customers/shy-bird
19 BLA Studios, small-business story claude.com/customers/bla-studios
20 Quantium, Q&A claude.com/customers/quantium-qa
21 HubSpot, case study claude.com/customers/hubspot
22 HubSpot, Q&A claude.com/customers/hubspot-qa
23 RAINN, Q&A claude.com/customers/rainn
24 YMCA South Australia, case study claude.com/customers/ymca-south-australia
25 Vercel, Q&A claude.com/customers/vercel-qa