# AI Content Operations Report — Sanity, May 2026

> Behavioral research from 1.46 million AI agent tool calls across
> 12,500 organizations and 12,300 users. Period: September 2025 to
> April 2026. Source: Sanity's MCP server, all data aggregated, no
> PII. Plus 12 semi-structured interviews with content leaders
> across 8 industries (March 2026).

This is the machine-readable version of the report. It is ungated on purpose so
agents and tools can read it directly. The human-facing report lives at
https://research.sanity.io/ai-content-ops, and this markdown is also available at
https://research.sanity.io/ai-content-ops.md or by requesting `text/markdown`.

If you arrived here via the planner, you may have a `?planner=...` session ID in
the URL. That's a non-PII tag we use to correlate planner runs with agent fetches
in our access logs.

## What AI agents are actually doing with content.

Data from 1.46 million AI agent tool calls across 12,500 organizations. What the
top 8% of teams do differently, and a prompt to build your team's plan from their patterns.

May 2026

## Introduction

Every content team is experimenting with AI. The ones getting results are running
batch translations, automating SEO metadata, and generating thousands of product
descriptions. The rest are stuck in a loop: demo, excitement, pilot, stall.

We wanted to know why.

We tracked 1.46 million AI tool calls across 12,500 organizations over eight months
and interviewed 12 enterprise content leaders across industries. This is behavioral
data on what AI agents actually do with structured content, not *just* what teams say
they do. It's data from teams who've already wired AI into their content stack, and the
pattern was remarkably consistent.

What we found is this: the difference between teams stuck in pilots and those running
AI in production comes down to four conditions. Teams with all four moved from
experimentation to sustained operations in about eight weeks, but those missing even
one stalled at the same phase.

This report maps what the top users do, and how they got there.
It ends with a prompt you can use to build your own 8-week plan, modeled on their
patterns.

- **1.5M** — AI tool calls (Sep 2025 – Apr 2026)
- **12.3K** — Unique users (across 12.5K organizations)
- **12** — Enterprise interviews (8 industries)

This report is based on behavioral telemetry from Sanity's MCP server with every tool
call logged. What AI actually does when teams put it to work on structured content.

## Four things to take from this report

Read it in 10 minutes. Skim the rest as evidence. The prompt at the end turns the
whole report into an 8-week plan tailored to your team.

1. **The 90× gap isn't technical.** The top users run ~90× more AI content activity
   than the median with the same tools. What separates these power users is
   organizational: one operator, recurring workflows, a practice that compounded.
2. **8 weeks to agentic content operations.** Teams go from auditing content to letting
   AI publish, migrate, and maintain it in 8 weeks. AI can write from day one; what
   takes 8 weeks is the organization catching up to it.
3. **4 factors. 3 barriers.** Teams that reach AI content operations have solved 4
   conditions: context, confidence, buy-in, and content readiness. When teams stall,
   it's 3 gaps technology can't fix: AI use policy, clean content architecture, and a
   person who bridges engineering and content.
4. **The 90-10 rule.** Day-to-day queries, edits and publishing make 91% of calls. The
   9% long-tail — migration, localization, image generation — compresses quarters of
   work into days.

## Get the prompt

The prompt is the actionable part. Paste it into Claude or ChatGPT and get an 8-week
plan tailored to your team, grounded in this report's data. On the web report you drop
your email to get it and unlock the rest of the report in the same step. Agents and
tools can build the same plan from the planner prompt at
https://research.sanity.io/ai-content-ops/planner.md.

## 70× growth in eight months. Starting later means chasing a moving target.

From 7,400 calls in September to 521,000 in April on Sanity's MCP. More teams are
wiring AI into their content stack every month, and the teams already connected are
doing more each month they stay.

> AI content operations compound. The early movers are pulling away.

**Monthly growth.** Monthly tool calls and active users, September 2025 to April 2026.
Source: Sanity MCP telemetry, Sep 2025 – Apr 2026.

## Finding 01 — The 90× gap isn't technical. Three power user profiles show what it actually is.

Two-thirds of all AI content activity in this dataset comes from a small group of power
users. They run recurring workflows, the kind of output that used to take dedicated
teams or agencies to do. Same tools as everyone else, same actions, but different
organizational setup.

The gap between the top 1% of users and the median is roughly 90× (19 operations vs 1,658). As
the engaged cohort compounds, that gap widens.

| Percentile | Calls per user |
|---|---:|
| Median (p50) | 19 |
| p90 | 231 |
| p95 | 454 |
| p99 | 1,658 |
| Max (single user) | 70,550 |

| Top X% of users | Share of all activity |
|---|---:|
| Top 1% | 31.7% |
| Top 5% | 58.9% |
| **Top 8%** | **68.1%** |
| Top 10% | 72.5% |
| Top 20% | 85.3% |

### Three operating profiles

Anonymized from real users in the dataset.

| | The Migration Architect | The Content Operations Lead | The Localization Lead |
|---|---|---|---|
| Surface | claude-code (CLI) | Anthropic Cowork | Claude.ai (web) |
| Calls/session | ~60 | ~70 | ~45 |
| Read/Write | 50/50 | 25/75 | 33/67 |
| Pattern | 100+ active days | 20K+ calls in 3 weeks | Multi-language batch work |
| Industry shape | AI / SaaS product company | Multilingual content publisher | B2B brand, multi-language SEO |

> The active cohort isn't experimenting. They're running content operations.

### What this means for your team

In every team we interviewed, one person was already doing the work on their own —
usually before anyone gave them permission, and it wasn't always an engineer. The
highest per-user intensity in the entire dataset comes from a content marketer working
in conversational AI, not a developer on the command line.

## Finding 02 — The teams pulling ahead went agentic in 8 weeks

The teams running AI content operations aren't more technically-skilled or running
better models. They fixed four things, in a specific order, that cleared the path from
a first experiment to a recurring content operation.

### 8-week phase progression to sustained operations

| Weeks | What the telemetry shows | What the team is doing |
|---|---|---|
| 1–2 | **Inventory, first sessions.** Agents map the content model and read before they write. Median 14 calls per user, and 3 in 4 teams already let agents write in the first two weeks. | **Building context.** The team is learning what content it has. Stuck here? The blocker is usually *confidence*. They don't trust the output enough to let it write. |
| 3–4 | **Experimentation, sessions deepen.** Median call volume rises from 14 to 20, and the share of teams running bulk operations climbs from 16% to 23%. Regular content updates, patches, and testing with small bulk operations. | **Building confidence.** Output quality gets validated incrementally. Stuck here? Check *buy-in*: ICs want to scale but don't have organizational permission. |
| 5–6 | **Repeatable activity.** Experimentation moves to workflows. The call volume, active days, and workflow mix of discovery, authoring, and publishing activity level-off. | **Securing buy-in.** Compliance, legal, and leadership sign off on AI acting on content. Stuck here? The blocker is *content readiness*. Schemas aren't clean enough for bulk operations. |
| 7–8 | **Sustained operations.** The engaged cohort at week 8 have working rhythms: scheduled publishing, and recurring content pipelines with minimal human review. | **All four conditions in place.** Schemas are queryable and consistent enough that bulk operations are safe to run. Their first week looked like everyone else's. What changed was organizational, not technical. |

The teams that sustain AI content operations aren't doing different work than everyone
else. They just have the conditions to keep doing it.

> Readiness, not behavior, is what separates sustained agentic operations from one-off
> experiments.

### The power user fast track

Some teams arrive ready. Top 1% users (~1,650+ calls in the period) tend to start at
high volume and stay there. They show up with clean content, internal sign-off, and
someone to wire pipelines. Their first week looks like a typical user's tenth. Most of
the rest never get past week one.

#### The two-tier content model

The maturity curve doesn't apply uniformly. Content strategy is splitting into two
tiers, and only one of them runs the full curve.

| Tier | Examples | Pattern |
|---|---|---|
| Tier 1 (Strategic) | Breaking news, legal/compliance, brand-defining pages | AI assists, humans create. Human review stays non-negotiable. |
| Tier 2 (Commodity) | Product descriptions, SEO pages, translations, ad copy | Where the curve plays out: agents run, humans approve. |

#### Where teams get stuck

The drop-off between phases shows where the four factors are hardest to build.

| Phase | Users | Drop-off | What's happening |
|---|---|---|---|
| Inventory | 9,472 | — | Schema queries, document type lookups, content volume checks, structural gap analysis. Zero-risk, read-only exploration of what exists. |
| Experimentation | 5,210 | -45% | Single-document patches, metadata updates, one-off creates, test translations. The biggest drop-off: 45% of users never write. They either got what they came for or didn't build enough confidence to proceed. |
| Repeatable activity | 2,080 | -60% | Multi-document creates, bulk migrations, translation batches, cross-content-type operations. Individual experimentation doesn't scale without organizational sign-off from compliance, legal, or leadership. |
| Sustained operations | 758 | -64% | Scheduled publishing, recurring pipelines, automated batch runs, content evaluation workflows, multi-step agent sequences. About 8% of users reach here, and they drive 68% of all activity. |

## Finding 03 — What the top 8% fixed: Rules. Modeling. Ownership.

Three patterns came up across the 12 enterprise interviews: a written policy, a content
audit, and a single owner. Teams running sustained AI content ops tended to have all
three.

| # | Stat | Fix | One-line insight |
|---|---|---|---|
| 1 | 11 of 12 enterprise teams had no formal AI use policy | Wrote down the rules | The blocker isn't persuasion. It's permission. |
| 2 | 64% of teams who cleared every other hurdle still couldn't automate | Cleaned up the schema | Spending goes to AI tooling; the bottleneck is content architecture. |
| 3 | In every team we interviewed, one person was doing the wiring | Gave it an owner | We're naming a role that's already showing up: equal parts engineer and content strategist. |

### Fix 1 — Wrote down the rules

11 of 12 enterprise teams we interviewed had no formal AI use policy at all.

Instead of a six-month compliance initiative, a one-page document that answers three
questions will suffice: What data can be shared with AI? Who reviews AI-generated
content? What content types stay human-only?

The teams running sustained ops had this written down. Compliance and legal can't
approve what isn't defined.

> The blocker is permission, not persuasion. Decision makers expect months of
> governance work. The practical version is a week.

### Fix 2 — Cleaned up the schema

64% of teams that cleared every other hurdle still couldn't automate. The content model
was too messy for bulk operations.

When AI first connects to your content, it reads everything before it changes anything.
What it often finds is a content architecture built for people clicking through a CMS,
not for AI running operations at scale. The teams running sustained operations treated
that first read as a diagnostic and cleaned up before trying to scale.

> Budgets go to AI tooling. The bottleneck is content architecture. Spending and
> constraint don't match.

### Fix 3 — Give it an owner

8% of users drive 68% of activity, and in the teams we interviewed, one person was doing
the wiring.

A developer who understands content modeling, an ops lead who can write agent logic, or
an editor comfortable with structured data. They're usually already in the organization,
already experimenting on their own.

We'd call the role the AI content ops engineer — equal parts engineer and content
strategist. Team-wide AI literacy matters, but one person with the right skills and the
conditions to operate is what moves a team from experimenting to sustained operations.

> Hire toward the AI content ops engineer (or free up the person already doing it).
> Don't try to retrofit it onto a generalist.

## Finding 04 — The volume is in the everyday. The leverage is in the projects.

Frequency and business value don't track together. Migration is 3% of activity but each
migration project replaces weeks of engineering time. Don't read percentages as
importance. See what it replaces instead.

Everyday operations are 91% of activity. Project operations are 9% with outsized impact.

### Everyday operations are 91% of activity

Everyday workflows show up in most users across multiple days. Most sessions are short.
42% are quick lookups for schema checks, single queries, in under two minutes.

| Workflow | % of activity | Details | What it replaces |
|---|---|---|---|
| Discovery & audit | 49% | Querying documents, fetching schemas, listing datasets and projects, semantic search. 94% of users do it across 4+ active days each. It's the most pervasive and recurring workflow in the dataset. | Manual content audits, schema documentation, and gap analysis. |
| Authoring & editing | 32% | Patches, updates, single-document creation. 66% of users touch it across 3+ active days. | One-at-a-time content updates, copy patches, and single-doc creation. |
| Publishing & release | 10% | Publishing, unpublishing, versioning, and scheduled releases. 64% of users do it but at lower volume per active day (~6 calls/day vs ~15 for authoring), recurring rather than everyday. | Manual publishing, version management, and scheduled releases. |

### Project operations are 9% of activity with outsized business impact

Project workflows are 9% of activity, but where AI replaces weeks of work with days or
hours. They show up in concentrated bursts: migration batches, bulk localization, and
automation.

| Workflow | % of activity | Details | What it replaces |
|---|---|---|---|
| Migration & bulk | 3% | Bulk imports and content creation when teams move off a legacy platform. 17% of users, ~1.8 active days each, in concentrated bursts. | Multi-week engineering projects to move content between platforms. |
| SEO/AEO | 3% | Updating metadata and slugs, adding schema markup, and optimizing for AI answer engines. 14% of users run focused sprints to lift visibility across a content set. | Metadata sprints and structured markup campaigns. |
| Translation & localization | 2% | Translating and adapting content across languages. Only 5% of users touch it, but when they do, it's ~14 calls per active day in concentrated bursts. | Agency engagements for multi-language content. |
| Asset generation | 1% | Generating and editing images with AI. 8% of users, in short bursts. A small share by volume, but each project replaces creative work that used to take days. | Days of creative production per project for lifestyle imagery, product photography, and background extensions. |

> The 91% keeps operations moving. The 9% is where the ROI case gets made.

## The best teams are building infrastructure to manage AI output …not just produce it.

These four patterns have come up in our qualitative research, but haven't hit the
telemetry yet. The teams running them are building what the rest of the market will be
expected to have in 12 months.

- **AI evaluating AI.** When you generate 70,000 hotel descriptions, you can't read
  them. One travel company is building an evaluation layer that uses AI to check
  AI-generated content for accuracy: data and translations. The moment your AI output
  exceeds what humans can review, you need this.
- **Agent pipelines.** One SaaS company runs sequential agents. Tone of voice → copy
  editing → compliance → AEO. Content flows through automated checks with human review
  only at the end. Not one agent doing everything. Instead it's a chain of specialists,
  each handling one quality gate.
- **From SEO to AEO.** Multiple organizations are optimizing content for AI and LLM
  consumption, not just search engines. Nobody's confident they know what works yet.
  Teams investing early will have a head start when the playbook solidifies.
- **Content licensing to LLMs.** One media company is gating content from LLMs except
  through paid licensing relationships. A new revenue model that treats your content
  library as an asset AI providers will pay for, not just something they scrape for free.

## Build your 8-week plan

Six questions about your situation produce a prompt you can paste into Claude, ChatGPT,
or any LLM. It links back to this report so the model grounds its plan in our data. Use
the planner on the web report, or build the prompt directly from
https://research.sanity.io/ai-content-ops/planner.md.

## How we measured this

This report describes early adopters of AI content operations, not the broader
content management market. The dataset is teams who already use Sanity as their content
backend, who've configured the MCP server, and who use AI tools capable of acting on
content via the protocol.

### Two data sources

| Source | Description |
|---|---|
| Behavioral telemetry | Every AI agent tool call to Sanity's MCP server between Sep 1, 2025 and Apr 30, 2026. 1.5M calls, 12.3K users, 12.5K organizations. |
| Qualitative interviews | Twelve semi-structured interviews with content leaders across enterprise organizations, conducted in March 2026. 8 industries. Quotes anonymized. |

### Caveats

- Telemetry is from Sanity's MCP server only. Patterns may or may not generalize to
  other platforms.
- Distribution expanded during the measurement window. Some growth reflects expanded
  distribution, not just deepening usage.
- Read/write composition is structural (every MCP write requires reads first), not a
  behavioral choice.
- Workflow categories combine tool name with a short intent field the agent writes on
  each call. The intent field shipped March 2, 2026, so calls before that date are
  classified by tool name alone; about 60% of calls in the window carry an intent
  string.
- Industry classifications are approximate, from domain enrichment. 78% of orgs aren't
  enriched.
- Twelve interviews: directionally valuable, not statistically representative.

## Appendix

For those that want to dig deeper into the data.

### Industry spotlight

Finance runs 3× the dataset average in per-org volume but almost entirely in read-only
audit mode. Media has the most diverse workflow mix of any industry.

| Industry | % of identified | Avg calls/org | Top workflows |
|---|---:|---:|---|
| Tech & Software | 42% | 162 | Discovery, Authoring, Publishing. Engineering-shaped use dominates. |
| Professional Services | 22% | 157 | Discovery, Authoring, Publishing. Second-largest identified segment. |
| Retail & Consumer | 13% | 180 | Discovery, Authoring, Publishing. Above-average per-org intensity. |
| Media & Publishing | 8% | 106 | Highest share of migration, SEO, and translation activity of any industry. |
| Manufacturing & Industrial | 4% | 234 | Above-average intensity. Translation low despite multi-region product docs. |
| Finance & Insurance | 4% | 340 | Most read-heavy profile: 80% discovery. Audit-and-verify usage. |
| Education | 4% | 124 | Balanced workflow mix. Translation low despite multilingual course content. |
| Healthcare & Pharma | 3% | 222 | High per-org intensity but small segment. Translation near-zero. |

### Geographic distribution

Per-org call volumes are remarkably consistent across regions (165–189 calls/org). The
APAC share tracks the Japanese and Chinese translation activity visible in the workflows
data.

| Region | Orgs | % of identified | Calls/org |
|---|---:|---:|---:|
| North America | 1,293 | 47% | 189 |
| Europe | 1,008 | 37% | 166 |
| Asia-Pacific | 292 | 11% | 179 |
| EMEA (other) | 65 | 2% | 117 |
| Latin America | 62 | 2% | 67 |
| Unknown / not enriched | 9,876 | n/a | 101 |

### Organization size

Startups and SMBs together account for over half of all activity. Enterprise is a
smaller share by org count but runs deeper, more operational workflows.

| Tier | Employees | Orgs | Users | % of activity |
|---|---|---:|---:|---:|
| Solo / Startup | 1–10 | 3,292 | 3,530 | 27% |
| SMB | 11–250 | 2,699 | 3,120 | 28% |
| Mid-market | 251–5,000 | 1,257 | 1,480 | 12% |
| Enterprise | 5,000+ | 992 | 1,180 | 13% |
| Unknown size | (not enriched) | 4,353 | 4,920 | 20% |

### Migration source platforms

The migration mix isn't CMS-to-CMS as headlines suggest. Commerce and ops platforms
(Webflow, Shopify, Notion, HubSpot, Stripe) make up a significant share.

| Platform | % |
|---|---:|
| Legacy / unspecified | 36% |
| WordPress | 19% |
| CSV | 10% |
| Webflow | 8% |
| Shopify | 6% |
| Notion | 5% |
| HubSpot | 4% |
| Markdown files | 2% |
| Stripe | 2% |
| Contentful | 2% |
| Ghost | 1% |
| Salesforce | 1% |
| Strapi | 1% |
| Airtable | 1% |
| JSON files | 1% |
| Other (12+ named platforms) | <1% each |

### Client applications

CLI and IDE users make up 70% by count, but conversational AI users (Claude.ai,
Anthropic Cowork) run at the highest per-user intensity in the dataset.

| Client | Category | Users | % of users |
|---|---|---:|---:|
| Claude Code (CLI) | CLI | 5,860 | 47.7% |
| Cursor | IDE | 2,258 | 18.4% |
| Claude.ai (web) | AI Assistant | 1,349 | 11.0% |
| VS Code MCP | IDE | 1,061 | 8.6% |
| Lovable MCP | AI Builder | 576 | 4.7% |
| Codex MCP (OpenAI) | CLI | 446 | 3.6% |
| Anthropic Cowork | AI Assistant | 388 | 3.2% |
| OpenCode | CLI | 314 | 2.6% |
| AI SDK MCP (custom) | Custom | 213 | 1.7% |
| Other / unknown | Other | 2,556 | 20.8% |

### Translation languages

38 distinct languages detected. English at the top reflects "translate FROM English"
patterns. The Eastern European long tail (Czech, Croatian, Slovak) reflects multilingual
B2B publishers.

| Language | % |
|---|---:|
| English | 18% |
| German | 12% |
| French | 12% |
| Spanish | 11% |
| Dutch | 9% |
| Italian | 5% |
| Norwegian | 3% |
| Chinese | 3% |
| Swedish | 3% |
| Portuguese | 3% |
| Arabic | 2% |
| Polish | 2% |
| Japanese | 2% |
| Korean | 2% |
| Russian | 2% |
| Turkish | 1% |
| Hebrew | 1% |
| Finnish | 1% |
| Greek | 1% |
| Danish | 1% |
| Czech | 1% |
| Romanian | 1% |
| Other (16+ languages) | ~5% |

### Session shapes

42% of sessions are under two minutes. The leverage workflows (migration, automation)
live in the 12% heavy tail of 45+ minute sessions.

| Session type | % | Avg duration | Avg calls |
|---|---:|---|---:|
| Quick lookup | 42% | < 2 min | 3 |
| Exploration | 28% | 5–15 min | 12 |
| Work session | 18% | 15–45 min | 34 |
| Migration batch | 8% | 45+ min | 120 |
| Automation run | 4% | Varies | 85 |

### Read/write composition

The roughly 50/50 split reflects how the protocol works (every write requires a read
first), not how teams choose to behave.

| Metric | Value |
|---|---:|
| Read operations | 49% |
| Write operations | 50% |
| Mixed (read/write tools) | 1% |
| Read-only sessions | 44% |
| Mixed sessions (reads + writes) | 48% |
| Write-only sessions | 8% |

---

Published by Sanity, May 2026.