Ten AI case studies. Not one of them is really about AI.
Every single one starts with a database nobody trusted.
We went looking for the best example of an AI stack in B2B marketing operations. Something recent, real, and worth copying. We found ten: Figma, ElevenLabs, Rippling, Anthropic, Sendoso, Oyster, Hex, Pendo, Exit Five, and Mission Cloud. Published between 2024 and this year, all with names and numbers attached.
Here's what we noticed reading all of them together: the AI is the least interesting part of every one. And that's genuinely good news, because the interesting part is the bit you can start on tomorrow.
Read the openings rather than the results and the pattern is hard to miss.
Figma's marketing operations lead ran an audit and found a chunk of their Salesforce contacts no longer worked at the companies they were filed under. Years of records, never updated since the day they arrived.
Anthropic couldn't tell who was signing up, because people register with personal Gmail addresses and their employer disappears with them.
Oyster had intent data pouring in from six different tools and nothing that acted on any of it.
ElevenLabs had signals from form fills, product usage and website visits, and no layer connecting them.
Exit Five had 27,000 contacts their COO described as essentially a black box. Mostly just email addresses.
Not one of these is an AI problem. They're the same problem wearing different clothes: the organisation didn't know who was in its own database. The AI arrives at the end, after someone has done the unglamorous work of making the data true. If it's the workflows themselves you're after rather than the pattern beneath them, we've collected those separately in AI in Marketing Examples: 10 Real B2B SaaS Workflows in Action.
Mission Cloud, an AWS partner now owned by CDW, ran a campaign targeting companies stuck on a legacy cloud platform after a licensing change pushed them to look elsewhere.
Their first attempt used intent data. It didn't work especially well. Plenty of companies were researching migration, but researching isn't the same as being contractually able to move. High intent, no permission to buy.
What fixed it wasn't a better model. It was renewal dates. Once they knew when each contract came up, they could tell the difference between someone reading and someone shopping. They reported 524 qualified accounts, $17m in pipeline and $8.2m in launched annual recurring revenue over six months.
The unlock was a spreadsheet column. That's the whole story, and it might be the most useful thing in the entire set.
Worth knowing before you copy any of this: nine of the ten were published by Clay, the B2B data tool, and the tenth by 6sense.
That isn't a criticism. Vendors have the access, the customers and the reason to write these up, and largely nobody else does. It's why the set exists at all, and we're glad it does.
It does mean they're marketing assets rather than research, and numbers drift accordingly. As one example: 6sense's written case study on Mission Cloud reports $17m in pipeline, while their own recap of the same campaign elsewhere puts it at $44m. Both are theirs. We don't know which to use, and that's a useful thing to notice rather than a scandal.
Pricing drifts too. Clay published a template for the Exit Five build, and its final step writes enriched data back into HubSpot. That's a CRM integration, which sits on Clay's Growth plan at roughly £330 a month ($446 annual, $495 monthly) rather than the cheaper tiers. The case study is from early 2025 and the pricing has been restructured since. Nothing untoward. The build just costs more than the story implies. For the realistic budget picture across this whole category, we've mapped it in The B2B Marketer's AI Stack: £0, £40 and £150 a Month.
None of that means ignore them. It means read them differently. Four questions get most of the value out:
Who published it, and what do they sell? Not to dismiss the result, but to know what's been left out. Nobody publishes the rollouts that quietly died.
What was broken before they started? This is the genuinely useful part, and it's almost always in the first two paragraphs. If their starting state doesn't look like yours, their result won't either.
Has anything changed since? Check the publication date against the tool's current pricing and plan structure. A build that made sense in early 2025 may sit behind a different tier now.
Which number is missing? Most report a hit rate, a lift, or a pipeline figure. Few report what it cost, how long it took, or how much of the output was usable.
That last one is where most of the disappointment lives.
Two things push back on all of this, and both have merit.
The first: the tools genuinely work. Waterfall enrichment, which just means asking one data provider after another until one returns a verified answer instead of giving up at the first miss, is a real improvement over single-source lookup. Exit Five enriched roughly half of 27,000 contacts and could finally describe their audience to sponsors. That's a real outcome from a real tool.
The second is sharper. "Fix your data first" is exactly the kind of advice that sounds responsible and quietly stops people ever starting. Data is never clean. Some of it can't be cleaned at all. If someone signed up in 2022 with a personal address and has changed jobs twice since, no tool is recovering that, and no amount of patience will.
So this isn't an argument for a cleanup project you have to finish before you're allowed to touch AI. That version becomes an eighteen-month data governance programme and no campaign.
The middle ground is narrower and more useful. You can't clean data in the abstract, only against a use. Pick one question you want answered about your list, something like which of these people work at companies big enough to buy from us, and fix only the data that answers it. Everything else stays broken on purpose until something needs it.
And when the data won't stretch to the question, that's an answer too. Scope the ambition to what you've actually got, rather than buying a tool to paper over what you haven't.
Not with a purchase.
Take 50 records off your list, matched to the buyers you actually care about, and run them through a free tier. Most tools in this category have one. Then check every result by hand against LinkedIn. Not the hit rate the tool reports, but how many you'd genuinely be willing to email: still at that company, not a catch-all domain, actually your buyer.
That number is the only one that matters, and no vendor will calculate it for you, because it's different for every list.
One of the case studies we read has a marketer paying for 2,000 enriched contacts and ending up with 347 they could use. The tool wasn't broken. They just measured the wrong thing until it was too late to matter.
And if you want the wider sequence for getting AI into your marketing without stopping for a cleanup project first, that's what our 90-day playbook is for.
Ten case studies, one lesson, and it's a cheaper lesson than it looks. The AI was never the hard part.
SaaStrix members get the 50-record test as a step-by-step walkthrough in the Operations Zone, and the full build in AI Labs.
What do the best AI marketing operations case studies have in common?
Every one starts with a data problem, not an AI problem. Figma found stale Salesforce contacts, Anthropic could not tell who was signing up behind personal Gmail addresses, and Exit Five had 27,000 contacts described as a black box. The AI arrives at the end, after someone has done the unglamorous work of making the data true.
Should you clean your data before using AI in marketing operations?
Not as an abstract cleanup project; that version becomes an eighteen-month governance programme and no campaign. You can only clean data against a use, so pick one question you want answered about your list and fix only the data that answers it. Everything else stays broken on purpose until something needs it.
How should marketers read vendor-published AI case studies?
Four questions get most of the value out: who published it and what do they sell; what was broken before they started; has anything changed since, especially pricing; and which number is missing. Most case studies report a hit rate or a pipeline figure, and few report what it cost, how long it took, or how much of the output was usable.
What is a good first step before buying an AI data tool?
The 50-record test. Take 50 records off your list, matched to the buyers you care about, run them through a free tier, then check every result by hand against LinkedIn. The number that matters is how many you would genuinely be willing to email, and no vendor will calculate it for you because it is different for every list.
SaaStrix is where B2B marketers become agentic marketing leaders, and you can try it free for five days.
.svg){kind=icon}