Let's get the obvious out of the way: somewhere between 85% and 91% of marketers now report using AI tools, depending on whose survey you trust. Two years ago it was barely half that. If you're not using AI in your content workflow, you're either lying or you've made a deliberate choice that we'd genuinely love to hear about.
But here's the thing nobody wants to admit: most teams using AI are doing it badly. They're treating ChatGPT like a magic content machine, copying and pasting outputs directly into their CMS, and wondering why their traffic is tanking and their audience is bouncing. Google deindexed 837 websites in March 2024, and every single one showed signs of AI-generated content.
We've spent the last year building AI into our actual workflows, not as a gimmick, but because our clients need to ship more content without hiring more people. Here's what's actually working and what we got wrong along the way.
MCP is the thing nobody's talking about enough
Model Context Protocol. If that means nothing to you, it will soon. Anthropic released it in late 2024, and now OpenAI and Google have adopted it too. Think of it as USB-C for AI, a standard way for AI models to talk to your tools.
Why does this matter? Because Sanity shipped an MCP server with 47 tools. Contentful's has 43 tools and AI Actions built in. WordPress.com announced MCP support in October 2025. This isn't "AI features" bolted onto a CMS. This is your AI assistant actually understanding your content schema, querying your database, and staging changes for review.
We renamed "Sanity Create" to "Sanity Canvas" across hundreds of documents last month. Not with find-and-replace. We described what we wanted in plain English, the AI found every instance, staged the changes, and created a Content Release for us to review. One conversation. Done.
That's not a parlour trick. That's just how content operations should work.
Contentful's MCP server does something similar but adds "AI Actions"—custom operations you define that the AI can invoke. Storyblok has community implementations with up to 155 tools. Even Drupal has production-ready MCP support now.
If your CMS can't talk to your AI tools natively, you're going to be copy-pasting between windows for the rest of your career. And frankly, you deserve better.
What we're actually doing with AI (not theory, actual work)
Enough theory. Here's what we actually ship.
Migrating 80 pharma documents into Sanity (the kind nobody wants to touch)
One of our clients had 80 pharmaceutical documents in Google Drive. Not marketing fluff—actual technical content with test results, clinical data, and the kind of information that gets embedded mid-sentence in ways that make structured extraction a nightmare.
You know the type: "The Phase II trial (n=340, double-blind, placebo-controlled, conducted across 12 sites in accordance with ICH-GCP guidelines) demonstrated a 23% reduction in primary endpoint markers (p < 0.001)…" That's one sentence. Try teaching a regex to parse that.
The old way to handle this: pay a specialist, watch them cry, hope they don't miss a decimal point that matters for compliance.
Our way: Claude Code + MCPs + careful sanitisation.
We built a pipeline that pulled documents from Google Drive, but the real work was teaching the AI to understand pharmaceutical content structure. Where does the methodology end and the results begin? What's a reference and what's inline data? Which numbers are critical and which are contextual?
Claude handled the extraction. We handled the verification. Every piece of clinical data got flagged for human review. Every test result got cross-checked against the source document. The AI was fast; we made sure it was accurate.
Was it fully automated? Absolutely not. Pharma content doesn't get to be fully automated—there's too much at stake. But what would have been months of specialist work became two days of supervised automation with a clear audit trail.
The content models we built were cleaner than anything that would have come from manual migration. Because we weren't just moving content—we were restructuring it with an AI that understood context, then sanitising the output like our client's compliance team was watching. Because they were.
Research that doesn't suck
Breaking news happens. A client needs a hot take. The old way: spend two hours reading everything, take notes, start writing. The new way: we have AI do deep research on a topic and output a structured markdown file with sources, key arguments, and gaps in coverage. We're not asking it to write the post. We're asking it to do the homework so we can write something informed.
This works especially well for technical topics where we need to understand what's out there before we have an opinion. The AI reads the boring stuff. We bring the perspective.
Traffic analysis in parity with content
We've hooked up PostHog's MCP with our Sanity content. This lets us ask questions like "which blog posts from Q4 have declining traffic?" and get answers that reference actual content in our CMS. We're building a plugin to streamline this further because right now it's a bit scrappy, but even scrappy it's saving us hours.
The goal is simple: know what's working, what's dying, and act on data instead of vibes.
Image generation everywhere
Every content workflow we build now has image generation baked in. Not as the final asset—please don't ship AI images to production without human review—but as a starting point. Need a hero image concept? Generate five options in 30 seconds. Need placeholder assets for a staging environment? Done. Need to visualize an abstract concept for a brief? Generate it, iterate, hand it to a real designer if it needs to be real.
We're not replacing designers. We're giving designers better briefs and giving content teams fewer blockers.
Cron jobs that actually do something useful
This is where it gets properly nerdy. We've built scheduled jobs that run against our content and surface problems before they become embarrassing.
Spelling and naming consistency checker
Every week, a cron job runs through all published articles and flags spelling errors, brand name inconsistencies, and terminology drift. Is it "Next.js" or "NextJS" or "Nextjs"? Is it "e-commerce" or "ecommerce"? Did someone spell Kubernetes wrong in a way that spellcheck missed because technically "Kubernates" is a valid word in some dictionary somewhere?
The job outputs a report. We review it. We fix the issues. It takes 20 minutes instead of the hours it would take to manually audit content we published six months ago.
This sounds small. It's not. Inconsistent terminology makes you look sloppy. Search engines notice it. Readers notice it. And nobody has time to manually review 200 articles every quarter.
Freshness monitor (our favourite)
Here's the one we're most proud of: an internal tool that tells us when content goes stale and what specifically needs updating.
It runs on a schedule, checks every article against a set of freshness criteria. Has a linked resource changed? Is there a date reference that's now in the past? Does the article mention a version number that's been superseded? Is there a stat that's more than 18 months old?
When it finds something, it doesn't just flag the article. It tells us what to update. "This article references Next.js 14, but Next.js 16 is now stable." "This stat about AI adoption is from 2023 and there's newer data." "The link to the Sanity docs in paragraph 4 returns a 404."
Fresh content ranks better. Everyone knows this. But keeping content fresh at scale is a nightmare without automation. We go deeper on how AI is changing content discovery in our post on AI crawling your content. We're not doing anything magical here—we're just checking things a human would check if they had infinite time and attention. They don't. So we built a robot to do it.
We build these for clients too. If your content operations should be running themselves, check out our agentic workflows service.
The economics are real but complicated
Let's talk numbers, because the numbers are wild and also misleading.
AI-assisted content costs about $131 per post on average. Human-written content runs about $611. That's a 4.7x difference. HubSpot says marketers save 3 hours per piece of content. CoSchedule's 2025 survey found 84% of respondents reported faster delivery.
Sounds great. Here's the catch.
MIT research shows a "J-curve" where AI adoption initially decreases productivity before improving it. The recovery typically takes 4+ years for established companies. A Danish study of 25,000 workers found "virtually no impact" on wages or employment two years into the ChatGPT era. And here's the stat that haunts me: while AI users complete 21% more tasks, review time increases 91%. The bottleneck just moves.
70-85% of AI initiatives fail to meet expected outcomes. That's not pessimism, that's MIT and RAND Corporation data.
So yes, the economics are real. But only if you build workflows that capture the efficiency gains without creating new bottlenecks. Most teams don't.
No spam, only good stuff
Subscribe, for more hot takes
The slop problem is worse than you think
"Slop" was Merriam-Webster's 2025 word of the year. Usage of the term increased 9x from 2024 to 2025. Pinterest added tools for users to filter out AI-generated images after being, in their words, "overrun."
Consumer trust is cratering. Enthusiasm for AI-generated creator content collapsed from 60% in 2023 to 26% in 2025. 70% of consumers familiar with generative AI say it makes content harder to trust. 98% say authentic images are important for establishing trust.
And Google is watching. The March 2024 update didn't just penalize low-quality content, it completely deindexed sites. 20 million monthly organic visits, gone. If you're pumping out AI slop, it's not a matter of if you get caught. It's when.
The stack we're actually using
For what it's worth, here's our current setup:
Sanity is the CMS. Its MCP integration, Content Agent, and schema-aware generation make everything else we do possible. The AI actually understands our data structures.
For research, we use Claude with deep research capabilities, outputting structured markdown. Sanity's embeddings index handles semantic search across our own content.
PostHog's MCP is connected to Sanity for analytics. We're still building the plugin to make this smoother.
Image generation goes everywhere we can fit it. DALL-E, Midjourney, whatever works for the use case. Always with human review before anything ships.
Sanity Content Agent handles bulk operations, Content Releases stage changes. Custom cron scripts check for spelling/naming consistency and content freshness, with reports piped to Slack so we actually see them.
This isn't the only way to do it. But this is what works for us right now.
The bottom line
Content generation isn't dying. But bad content generation is about to get very expensive. Google is penalizing slop. Consumers can spot it. And the efficiency gains everyone's chasing only materialize if you build actual workflows, not just throw prompts at problems.
The CMS platforms with real MCP integration, schema-aware generation, and batch operations through natural language are pulling ahead already. The ones bolting on a ChatGPT wrapper and calling it "AI-powered" are going to have a bad time.
If you're a content team trying to figure this out: start with research and first drafts. Build human review into everything. Measure what actually improves, not just what feels faster. And for the love of god, don't publish anything an AI wrote without reading it first.
We're moving fast, but we're not producing slop. That's the only strategy that actually works.
If you want help building AI into your content workflow without wrecking your brand or your search rankings, we're here to help. We've made most of the mistakes already so you don't have to.