AI Transformation
on the Claude Platform

A 10-slide brief for executives sponsoring AI bets

When Claude wins, what it costs, how to roll it out in 90 days, and how to govern it.

As of 2026-05. Verify pricing/models at anthropic.com.

The shift

Foundation models are the new platform primitive

The build/buy question has changed. You no longer build the intelligence — you build on top of it.

Wave	What you used to build	What you build now
Mainframe → PC	The OS	Apps on the OS
On-prem → Cloud	The data center	Services on AWS/GCP/Azure
Custom ML → Foundation models	The model	Workflows on the model

Organizations that try to recreate the model layer in 2026 are repeating the "build our own data center" mistake of 2010.

The platform

Claude in 60 seconds

3 model tiers

Haiku 4.5 — fast, cheap, triage and high-volume
Sonnet 4.6 — balanced, the workhorse
Opus 4.7 — deep reasoning, hard problems

6 capabilities that change the math

Prompt caching — 90% off cached input
Extended thinking — model reasons before answering
Tool use — model calls your APIs
Computer use 2.0 — model drives a browser/desktop
Files API — upload docs, model reads them
Batch API — 50% off, async jobs

Plus: Skills (custom domain expertise), MCP (connect external systems), Agent SDK (build autonomous workflows), Plugins (bundle and distribute), Memory tool (cross-session state), Claude Code (engineering CLI).

The decision

When Claude wins (and when it doesn't)

Anchor heuristic: build where you have a moat, buy where you don't. Claude wins when the use case is a differentiator your competitors can't buy off the shelf — proprietary data, workflow, expertise. For pure-commodity use cases, packaged SaaS is usually the right call.

Claude wins when…

Long context matters (200K+ tokens)
Careful reasoning beats speed
Code-heavy workflows
Agentic / multi-step automation
Regulated data (no-train guarantees, BAA paths)
Hard-to-jailbreak production deployments

Pick something else when…

You need image/video generation (Claude is text + vision input only)
You need real-time voice (use a voice-native vendor)
Sub-100ms latency is the binding constraint
You're locked into another platform's stack and switching cost > benefit
The use case is so simple a templated SaaS already solves it

Cost

The cost levers most teams miss

60–80%

Achievable cost reduction vs. naive use (all-Opus, no caching, no batch).

Prompt caching

~90% discount on cached input tokens. Game-changer for RAG, system prompts, repeated context.

Batch API

50% discount on async work. Use for enrichment, classification, summarization at scale.

Model mix

Triage with Haiku, route hard cases to Sonnet/Opus. 5–15× cost gap between tiers.

→ Run the numbers in cost-calculator.html for your volume.

Speed

Weeks, not quarters

Platform features collapse traditional ML build timelines.

Capability	DIY timeline	On Claude
Domain-specific assistant	3–6 months (fine-tune + eval + serve)	1–2 weeks (Skills + system prompt + eval)
RAG over internal docs	2–4 months (embed + index + retrieval + serving)	1–3 weeks (Files API or vector store + prompt caching)
Agentic workflow	4–8 months (orchestration + tool wiring + recovery)	2–6 weeks (Agent SDK + tool use)
Connect to internal systems	Custom adapters per system	MCP server (1 adapter, all clients use it)

The platform compresses build time. The hard work shifts to evaluation, governance, and change management — which is where it always belonged.

Governance

What the risk reviewer needs to hear

Built-in

No training on API data by default
Data residency options (US, EU)
BAA available for HIPAA workloads
SOC 2 Type II, ISO 27001
Zero data retention options for sensitive flows

You add

Prompt versioning (treat prompts as code)
Eval suite per use case
Cost guardrails (per-team budgets, alerts)
Audit log (request, prompt hash, response, model version)
Multi-model abstraction for vendor concentration risk

→ Full mapping to NIST AI RMF + EU AI Act in governance-overlay.md.

Rollout

The 90-day rollout

Weeks 1–4 — Pilot

1 use case, 1 team
Executive sponsor named
BAA signed if needed
Success metric defined upfront
Eval suite v0

Weeks 5–8 — Guardrails

Prompt versioning live
Cost dashboard per team
Red-team pass
Audit log piped to SIEM
Internal docs + runbook

Weeks 9–13 — Scale

2nd and 3rd use case
Center of excellence stood up
Reusable Skills library
MCP servers for shared systems
Quarterly review cadence

→ Operational detail in adoption-playbook.md.

Risks

Five risks worth naming

Risk	Mitigation
Vendor concentration — single-vendor model dependency	Multi-model abstraction layer; quarterly evaluation of alternatives
Model deprecation — models get retired	Pin to model family in code; eval pipeline detects regression on switch
Cost surprises — unexpected bills	Per-team budgets + hard caps + per-request cost telemetry
Prompt drift — silent quality regressions	Prompt versioning + canary evals + alerts on metric drops
Shadow AI — staff using consumer Claude with company data	Sanctioned API access with logging; policy + training; SSO-gated

None are show-stoppers. All are predictable. Plan for them in week 1, not month 6.

Stale factual priors about Claude (context window, caching ROI, sandbox denyRead, refusal calibration) quietly distort all five mitigations. See claude-misconceptions.md before final architecture sign-off.

Decide

What to ask of your team this week

Pick the pilot use case. Bounded, measurable, executive sponsor named, fail-safe if it doesn't work.
Run the cost calculator for your projected volume. Decide whether the economics work before signing anything.
Score build vs. buy on the worksheet. Default to platform; override only with reason.
Stand up the eval suite in week 1. No eval = no go.
Name the governance owner. Not the AI team. Risk, legal, or security.

Companion artifacts: cost calculator · feature matrix · adoption playbook · governance overlay · build-vs-buy worksheet · reference architectures · Claude Code rollout

AI Transformationon the Claude Platform