The 5 Levels of AI in Revenue Operations

Anthropic shipped five rungs across April 2026. Most teams should be on rungs 1 to 2. Here is the decision framework and the mistakes that cost teams their AI programmes.

18 min read Strategic AI Strategy James Killick

The Anthropic stack is the five-rung ladder of AI work surfaces Anthropic shipped across April 2026: Cowork (shared workspace), Scheduled Tasks (recurring jobs), Dispatch (mobile initiation), Routines (cloud triggers) and Managed Agents (hosted runtime). Each rung unlocks a different class of revenue work and builds the operational discipline the next rung requires.

Why the ladder exists

Anthropic did not ship five products and leave it to you to connect them. The release cadence across April 2026 was deliberate. Cowork reached GA on 9 April. The Managed Agents API shipped in beta on 7-9 April. Routines opened up on 14 April. Together they describe a complete path from "Claude helps you in a single conversation" to "Claude runs stateful autonomous work for hours".

The reason this matters for revenue teams: the rungs are not optional, but nor are they all necessary. The ladder exists so you know what the next rung unlocks, not so you race to the top.

Level 1 - Cowork: the shared workspace

Tool

Claude Cowork (GA 9 April 2026)

Who uses it

Sales, marketing, ops, finance - anyone working with context

Typical starter use case

Call prep, proposal drafting, meeting follow-up

Cowork is Claude as a deskside collaborator inside a shared web workspace. Sessions persist. Shared skills and knowledge carry across conversations. Teammates can join. This is the rung every revenue team should start on because the discipline built here - brand voice, shared context, approval patterns - is exactly what higher rungs require to run safely.

GA on 9 April 2026 added RBAC, OpenTelemetry observability and a usage analytics dashboard, closing the enterprise governance gap that had blocked deployment at most $10M+ companies. See our Cowork GA breakdown for the procurement conversation.

Level 2 - Scheduled tasks: supervised recurring work

Tool

Claude Code scheduled tasks (local machine)

Who uses it

Ops or dev team builds, broader team reviews output

Typical starter use case

Overnight pipeline hygiene, Monday competitor digest

Level 2 is the first step away from live collaboration. Work runs on a schedule without a human initiating each run, but a human still reviews the output before it affects a customer. This is the safest place to learn operational patterns like structured output, approval flows and failure modes.

The tradeoff at this level: scheduled tasks in Claude Code run on your local machine. Close your laptop and the job stops. That limitation is what Level 4 resolves.

Level 3 - Dispatch: mobile-initiated work

Tool

Claude Dispatch (mobile)

Who uses it

Founders, executives, travelling sellers

Typical starter use case

Contract review, board deck drafting, fast research

Dispatch is about where the work starts. A founder in an Uber asks Claude to read the inbound MSA and flag anything that diverges from the standard terms. A CRO in an airport asks for a three-bullet prep for tomorrow's board meeting based on this week's pipeline.

Dispatch is a workflow unlock for senior time-starved people. It is less an architecture choice than a distribution channel for Levels 1 and 2.

Checkpoint

Levels 1 to 3 are where most teams should live in their first 6-12 months with AI. If your team is still producing off-brand output or failing to ship reliably at Level 2, do not move to Level 4. Get the fundamentals right first.

Level 4 - Routines: cloud-native triggers

Tool

Claude Code Routines (research preview 14 April 2026)

Who uses it

Ops or dev team builds, triggered by events in your stack

Typical starter use case

Event-driven inbound lead triage, post-demo follow-up

Level 4 is the jump from "Claude runs on my schedule" to "Claude runs when something happens in the world". Three trigger types: cron schedule, API webhook, GitHub event. The work runs in Anthropic's cloud, independent of your laptop.

For a revenue team, the practical shift is that Claude starts participating in the same event-driven plumbing that your CRM and marketing automation already run on. See our three starter routines for sales teams for concrete workflows.

Level 5 - Managed Agents: the production runtime

Tool

Claude Managed Agents API (beta 7-9 April 2026)

Who uses it

Dev team builds, runs autonomously

Typical starter use case

Proposal-drafting agent, overnight competitive intel

Level 5 is a full hosted agent runtime. Anthropic handles container provisioning, tool execution, session persistence, streaming and retry logic. You describe the agent configuration once and the loop runs server-side for hours.

Pricing is $0.08 per session-hour plus token costs, metered to the millisecond. A one-hour Opus session with 50k input and 15k output tokens costs roughly $0.70 total. See our Managed Agents vendor conversation post for how this reshapes AI vendor quotes.

The killer combination: Routines + Managed Agents

The two releases that matter most together are Level 4 and Level 5. Routines handles the trigger. Managed Agents handles the execution. An inbound lead hits your CRM, a Routine fires, a Managed Agent session picks up the work, runs qualification, drafts a response and leaves it in your SDR's queue.

Before April 2026, this pattern required stitching together cron, webhook receivers, a custom agent harness and glue code. Anthropic collapsed the assembly into two official products designed to pair. This is why the ladder matters - it lets you see where to invest effort vs where the platform does the work for you.

Pick your rung: decision matrix

Situation Start at Why
No AI in revenue ops yetLevel 1Build brand voice and shared skills first
3+ months at Level 1 with stable usageLevel 2Introduce scheduled jobs with human review
Leaders are mobile and time-starvedLevel 3Reclaim transition time without new workflow
Stable Level 2 for 2+ quartersLevel 4Move to event-driven, always-on work
Clean Level 4 with observabilityLevel 5Ready for autonomous production workloads

Where Code and Cowork fit

Claude Code powers Levels 2, 4 and 5. Claude Cowork powers Level 1 and often serves as the review surface for output produced at Levels 2 and 4. For the full breakdown of who uses which tool, see our Code vs Cowork decision guide.

The skipped-rung failure pattern

The most common AI programme failure is "we jumped to Level 5". The pattern is always the same.

Week 1-2

Team stands up an autonomous agent. It works. Stakeholders are excited.

Week 3-4

Agent produces slightly off-brand output. Nobody catches it because there is no review loop. A customer sees it.

Week 5-6

Trust collapses internally. The programme gets paused for "review". It never restarts.

Levels 1 to 3 build the review loops and brand-voice discipline that stop this pattern. They are not bureaucratic steps to skip. They are the foundation.

90-day plan by starting point

Starting at zero

Days 1-30: Cowork + brand voice document. Days 31-60: first scheduled task at Level 2. Days 61-90: measure and refine.

Already at Level 1

Days 1-30: pick one recurring manual task, move to Level 2. Days 31-60: add observability. Days 61-90: plan first Level 4 routine.

Already at Level 2

Days 1-30: move one job to Level 4 Routines. Days 31-60: pair with a Managed Agent. Days 61-90: design approval flow for first autonomous use case.

Frequently Asked Questions

What are the 5 levels of the Anthropic AI stack?
Level 1 is Cowork (real-time shared workspace). Level 2 is Scheduled Tasks in Claude Code (recurring supervised jobs). Level 3 is Dispatch (mobile-initiated work). Level 4 is Routines (cloud-native event triggers). Level 5 is Managed Agents (hosted production runtime). Each rung unlocks a different class of work and builds the operational discipline required for the next.
Which level should a B2B sales team start on?
Most $5-50M ARR B2B sales teams should start at Level 1 or 2. Level 1 (Cowork) builds the shared brand voice and approval patterns that make higher rungs safe. Level 2 introduces scheduled tasks where a human still reviews output. Only move to Level 4 and 5 once Level 2 runs have been stable for at least a quarter.
Why not skip to Level 5 autonomous agents?
Skipping rungs is how AI programmes fail publicly. Autonomous agents need clear brand voice, approval flows, observability and trusted data access. Those disciplines get built at Levels 1 to 3. Teams that jump straight to Level 5 typically produce off-brand output, lose internal trust and end up pausing the project within a quarter.
Is this framework specific to Anthropic or does it apply to other AI vendors?
The ladder describes the work, not the vendor. OpenAI, Google Gemini and others offer similar tiers under different labels. We use Anthropic naming because most of our sales clients standardise on Claude. The decision framework - which rung fits which work - applies whichever provider you choose.
What is the killer combination in the stack?
Routines + Managed Agents (Levels 4 and 5). Routines handles the trigger - a schedule, a webhook, a GitHub event. Managed Agents handles the execution - a stateful session that can run for hours. Together they replace the DIY cron + webhook + agent harness assembly that teams cobbled together before April 2026. This is the "always-on Claude" pattern.
What tools sit at each level?
Level 1: Claude Cowork (shared web workspace, reached GA 9 April 2026). Level 2: Claude Code scheduled tasks (local-machine recurring jobs). Level 3: Claude Dispatch (mobile initiation). Level 4: Claude Code Routines (cloud-native triggers, research preview 14 April 2026). Level 5: Managed Agents API (hosted runtime, beta 7-9 April 2026).
How long does each rung take to set up?
Level 1 takes a week for the first team and 3-4 weeks to reach the point where brand voice and shared skills are stable. Level 2 takes 2-4 weeks once Level 1 is solid. Level 4 takes 2-6 weeks depending on integration complexity. Level 5 takes 4-8 weeks because it requires observability, approval flows and production-grade configuration. Skipping the foundational rungs looks faster but produces unreliable output.

About the Author

James Killick
James Killick

Co-founder at Njin. Building AI-powered sales systems for B2B businesses.

Ready to pick your rung?

Talk to our AI about where your revenue team should start on the Anthropic stack.