The 90% Rule: AI Agents Are Coming for Ops, Not Code

For the first three years of the generative AI wave, the loudest use case was coding. GitHub Copilot, Cursor, Claude Code, and Anthropic’s own developer tooling defined the category. The implicit market thesis was: engineers adopt first, then everyone else follows.

That thesis just broke, publicly, in the vendor’s own numbers.

In its July 7, 2026 expansion of Claude Cowork to mobile and web, Anthropic disclosed usage data that reframes the entire market: roughly 90% of Cowork sessions are non-developer work. The 1.2 million weekly sessions across 600,000+ organizations are running finance workflows, marketing operations, customer support triage, executive research, HR automation. Engineering is a rounding error.

This is not a marketing anecdote from a vendor blog. It is the shape of the actual market, disclosed by the model provider with the most complete visibility into it. If you are a solo founder, an SMB owner, or a mid-market ops lead trying to figure out where to spend your first €10K on AI agents, this number should change your answer.

What Anthropic actually disclosed

The specifics from the July 2026 release notes and Anthropic’s enterprise agents report:

• 1.2 million weekly Claude Cowork sessions across the platform (up from ~340K at Q1 launch).
• 600,000+ organizations with at least one active Cowork session.
• ~90% of sessions are non-code work. Anthropic’s breakdown of the top use categories: research and analysis, document drafting and review, customer/prospect communication, data extraction and summarization, workflow orchestration across SaaS tools, project management.
• 57% of organizations are running agents on multi-stage workflows (not single-turn tasks).
• 80% of enterprise deployments report measurable ROI within 6 months.

Two things to notice about that 90% number.

First, it is directionally consistent with independent surveys. Upwork’s 2026 State of AI in SMBs found that among small businesses actively deploying AI beyond ChatGPT-style chatbots, the leading use cases are decision support (41%), information retrieval (36%), and workflow automation (34%) — none of which are coding.

Second, it maps to what Anthropic charges for. Claude Enterprise pricing is per-seat, and the seats are being bought by ops, finance, and marketing budgets — not engineering budgets. That is a much larger addressable market and Anthropic knows it.

Why the code-first assumption was wrong

Three reasons the industry got the shape of the market wrong for two years:

1. Developers were the loudest early adopters, not the most valuable.
Engineers try new tools fast, tweet about them, write blog posts. That created the illusion of a coding-dominated market. Meanwhile, a finance analyst quietly automating monthly reporting was invisible on X but was generating €40K/year of reclaimed time.

2. Coding is a legible ROI story. “Ship features faster” is easy to sell to a VP of Engineering with a budget. “Get 6 hours a week back for our controller” is a harder story to tell to a CFO who has never bought AI. So the coding tools were priced, packaged, and pitched first — but the demand curve was always larger on the ops side.

3. Agent tooling reached “good enough for ops” later than “good enough for code.” Code has fast, cheap feedback loops (tests, compilers, human review). Ops workflows require reliable tool use, memory, and coordination that only became stable in the Claude 3.5 → 4.x → 5 window. Once the reliability arrived, ops adoption exploded and coding adoption plateaued (relatively).

If you are still buying AI agents on the assumption that this is a developer-first category, you are budgeting against the world of 2024, not 2026.

Three implications for how SMBs should buy

Implication 1: Your first agent is almost certainly not in your engineering department.
Even if you have engineers, they are the wrong first team to build for. The ROI density is higher elsewhere: finance closing the books, sales research briefs, support triage, HR onboarding, legal contract review. Find the function with the most repetitive knowledge work per head and start there.

Implication 2: Stop budgeting AI as an IT line item.
If your AI budget lives inside IT, it is under-scoped. IT budgets are sized for infrastructure, not for buying back human time across departments. Move the AI agent budget to the operating budget of the function that will use it. Rebill IT for the infrastructure component (API costs, security review). This unlocks 3-5x the spend that was actually justified.

Implication 3: Choose vendors and agencies by ops fluency, not engineering pedigree.
An engineering-heavy AI shop will build you a technically clean agent that solves the wrong problem. An ops-fluent shop will spend the first two weeks embedded with the team that will use the agent, then build a technically simpler thing that actually gets adopted. In 2026, adoption is the binding constraint, not model capability. Pick accordingly.

How to pick your first non-code agent

If the data says your first agent should be non-code, the practical question is: which non-code workflow? A four-part test that maps well to the Workforce Starter build pattern:

Criterion	Green	Yellow	Red
Repetition	Runs daily or weekly, same shape	Runs monthly	Runs quarterly or ad hoc
Volume	≥4 hours/week per person	1–4 hours/week	Under 1 hour/week
Data source stability	Same 2–3 systems (Gmail, HubSpot, Notion)	4–6 systems	7+ systems or custom formats
Judgment complexity	Rules-based with occasional exceptions	Rules with frequent exceptions	Mostly judgment, no rules

Four greens = ship this quarter. Two greens and two yellows = ship after a light discovery. Any red = pick a different workflow first.

The workflows most SMBs land on when they run this test: inbound lead research briefs, invoice extraction and coding, first-line support triage with routing, weekly reporting drafts, meeting note structuring and action-item extraction. All non-code. All in the 90%.

The one case where code-first still wins

Two conditions where investing in code-agents first still makes sense:

1. You are a software company selling to developers. Your engineering velocity is your product. Coding agents (Claude Code, Cursor) are direct product-quality investments, not ops overhead. The 90% rule does not apply to you the same way.

2. You have a legacy modernization backlog that is the strategic bottleneck. If your business runs on a codebase that no one can maintain and code-agents can materially accelerate the modernization, that is a legitimate first spend. Rare, but real.

Everyone else should follow the 90%.

What to do this quarter

Three concrete steps if you are convinced by the data:

1. Audit where your knowledge workers spend their time. Two weeks, informal, not a McKinsey diary study. Ask each department lead: “What is the workflow that eats the most of your team’s time and produces the most similar-looking outputs?” The answer is your first agent candidate.

2. Skip the AI strategy deck. With this much market data available, another 12-week strategy engagement is not what you are missing. Deploy one narrow agent, learn from the actual deployment, and let the strategy sharpen against real data. See AI Agent vs AI Consultant for the sequencing argument.

3. Ask your finalist vendors to show you a non-code case study. If their strongest reference is a code deployment, they are calibrated for the 10%. You are the 90%. Pick differently.

---

The market has been telling you where it is going for two years, but the signal was drowned out by the volume of the engineering conversation. Anthropic just published the actual shape. The winners in 2027 will be the operators who read it and rebalanced their AI budget in Q3 2026.

If you want help pressure-testing which non-code workflow in your business is the highest-value first agent, the 30-minute discovery call is designed for exactly that question.

---

Key terms in this post: AI agent · agentic AI · Workforce Starter · tool use · Claude Cowork