// FIELD_BRIEFING
The adoption question is settled. The advantage question isn't. What separates companies getting durable returns from AI and companies stuck at the demo isn't the model — it's what they teach it.
Reading Anthropic's "Building AI Agents for the Enterprise" // Kerry Morrison's interpretation // June 2026
THE_SETUP
40%
of US employees use AI at work (Sept 2025)
— up from 20% in 2023
So the question is no longer whether to use AI.
It's whether all that usage becomes a lasting advantage — or fizzles into incremental gains that plateau by the end of the quarter.
The answer depends entirely on scope.
THE_TRAP
A chatbot here, a summarizer there, a pilot that impresses as a demo but never scales beyond the team that built it.
This kind of adoption can be genuinely useful. But it doesn't change how the organization operates. The tool sits to the side of the work instead of inside it.
THE_DIVIDE
One side
Answers a single question in isolation. Responds, then waits. Effective for FAQs and lookups — but it doesn't plan, act, or follow through.
The other side
Plans, makes decisions, and uses tools across multiple steps to finish a task — applying your domain expertise, not just answering questions about it.
One responds. The other operates.
THE_PATTERN
Not "where do we add AI?" but "how do we rebuild around it?" — across three pillars, simultaneously.
People
Hand off the repetitive, information-dense work. Free judgment for where it matters.
Process
Compress information-dense work from months to minutes — without losing quality.
Product
Ship capabilities that weren't feasible before — and generate net-new revenue.
PILLAR_01 // PEOPLE
L'Oréal built an internal platform that routes plain-English questions through 15+ specialized agents and returns sourced answers with visualizations — so anyone can query data directly instead of waiting on a custom dashboard.
Source: Anthropic customer story
99.9%
conversational analytics accuracy
(up from 90% with previous GenAI)
44,000 monthly users · 2.5M messages/mo · 15,000 daily active
PILLAR_02 // PROCESS
−87%
customer support resolution time
+30% decision accuracy · 30–40 min waits → seconds
Lyft put Claude behind customer support. The agent greets customers by name, investigates the specific situation, and resolves it — routing to humans (with AI-written summaries) only when judgment is required.
Source: Anthropic customer story
THE_FLYWHEEL
Normal process
Every project starts from scratch. The same review effort is required every single time. Nothing accumulates.
Compounding system
Every expert correction becomes the new baseline for everyone. Tribal knowledge becomes institutional infrastructure.
So this is a head-start game. The company that starts narrow today isn't ahead by one project — it's on a different curve.
PILLAR_03 // PRODUCT
Rakuten offloaded its agent infrastructure to a managed harness, freeing engineers to build the experience instead of maintaining the plumbing underneath it.
Source: Anthropic customer story
Every 2 weeks
major releases — was once a quarter
−97% critical errors (pilot) · −30% agent cost & latency · specialist agents shipped in a week
THE_CONSTRAINT
In regulated industries, security and compliance aren't features — they're prerequisites. An AI product that needs sensitive data to leave the security perimeter triggers reviews that take months and often end in rejection.
Any AI product that operates outside your trust boundary is a product that cannot ship. Solve the trust architecture first and you can move fast. Treat it as an afterthought and your best ideas stall in compliance review indefinitely.
THE_REAL_LESSON
Two companies using the same model get dramatically different results depending on how much of their own context they encoded into it.
The model is a commodity you rent by the token. The advantage is the layer on top — your standards, your methodology, your compliance framework, your brand voice, your chart of accounts.
That's the part a competitor can't buy.
THE_SCALE_MECHANISM
L'Oréal, Lyft, and Rakuten each built a platform. Most companies can't — and now don't have to. Claude Cowork gives non-technical teams the same agent capabilities, and plugins are how expertise scales.
What a plugin is
A package of skills, context, and connectors that gives Claude one team's role-specific expertise — their CRM, their templates, their risk framework.
Why it matters
Build once, share across every team. The institutional knowledge in your best people's heads becomes available to every new hire on day one.
THE_NON_NEGOTIABLE
Admin controls, audit trails, and approved-tool catalogs are prerequisites for rollout — not features you bolt on after things sprawl.
| Requirement | What it prevents |
|---|---|
| Admin marketplace | Shadow AI, unsanctioned tool sprawl |
| Local execution | Data leaving the security perimeter |
| OpenTelemetry audit | "How was AI involved in this decision?" being unanswerable |
| Role-based access | Wrong capability in the wrong hands |
THE_PLAYBOOK
Weeks 1–3
Pick 2–3 teams with clear pain. Install or build plugins. Define success before anyone starts.
Months 2–3
Real production workflows, not sandboxes. Measure adoption weekly. Prove value.
Months 4–6
Deploy marketplace controls and approval workflows. Each new wave moves faster than the last.
THE_PRINCIPLES
Start with specificity, not scale.
Generic output on day one and people never come back. Give it your real context first.
Pick pilots with a measurable finish line.
"Improve productivity" is easy to dismiss. "Contract review from five days to one" isn't.
Build for reuse from the beginning.
Encode the knowledge once. Marginal cost of sharing is zero; marginal value is enormous.
Don't underestimate governance.
Skip it early and you'll spend more cleaning up sprawl than you saved moving fast.
THE_REFRAME // KERRY'S TAKE
Most businesses can't stand up a 15-agent orchestration layer — and read these case studies as proof the whole thing is for a different budget.
It isn't. The mechanism scales down cleanly. You don't need the platform; you need one process with an obvious finish line, enough of your own context to make the output shippable, and the honesty to measure what happens.
The divide was never between companies with AI and companies without it. Almost everyone has it now. It's between companies that bolted it on and companies that built it in.
START_NARROW
You need a specific starting point, quantifiable success criteria, and the willingness to learn from what happens next.
Attribution
Framework, case studies, and metrics are from Anthropic's "Building AI Agents for the Enterprise" (April 2026). The mid-market reframe and editorial interpretation are Kerry Morrison's.
Find your starting point → betterstory.co/contact