Your stack isn’t short of AI. It’s short of tools that talk to each other.
Everything already has AI in it
Count the tools your Digital Marketing team opened this week. An SEO tool. A social listening platform. Two analytics platforms, because someone (me) standardised twice. A CMS. A design tool. A DAM. A translation service. A social publishing tool. A survey platform nobody’s looked at since the last wave closed.
Almost all of them shipped AI features this year. Summarise this, draft that, auto-tag the other. It’s in the release notes. It’s definitely on the renewal quote.
And your team still copies things between them by hand.
That’s the frustrating part. Nobody’s short of AI. We’re short of context that travels. Someone pulls a chart out of the listening tool, pastes it next to some search data, writes a brief in a doc, emails it to a designer, and files the finished asset in the DAM three weeks later (if we’re lucky). The marketer is the integration. That’s the actual cost, and in-platform AI doesn’t touch it.
What that actually costs you
Four things, and I’d bet you recognise all four.
Your analysts don’t analyse. The clever part of the job is the last twenty minutes. The first three hours are downloading and reconciling.
Reports land too late to act on. Data in three systems, narrative written by hand in a spreadsheet. By the time it circulates it’s history, not a decision.
Brand voice drifts. Every tool applies the guidelines separately, so every tool applies them slightly differently, so what you’ve really got is whoever happened to review it that day.
Customer insight sits there. Not because anyone doubts it’s valuable. Because linking a survey finding to a content decision takes effort, and effort loses to deadlines. Every time.
None of that is an AI problem. It’s plumbing. AI just happens to be very good at sitting on top of plumbing.
The thing that changed isn’t the models
Everyone’s fixated on model quality in the rush to AGI. That’s not what has shifted, imho. AI can be a productivity tool/enhancer/amplifier, if harnessed in the right way.
What shifted is that there’s now a standard way for AI tools to plug into enterprise systems, and vendors picked it up much faster than anyone expected. Most of the big marketing platforms publish their own connector now. You point your AI at it, log in as yourself, and it can read exactly what you were already allowed to read. Nothing more.
It’s called Model Context Protocol, if you want to sound informed in a meeting. The spec matters less than the economics.
Connecting two enterprise systems used to be a project. Build queue, architecture review, a line in someone’s capital budget, six months of polite chasing. If the vendor hosts the connector, a marketing team can switch it on over a long lunch.
That changes who gets to start. It doesn’t change who owns the hard stuff later. Worth keeping those two apart in your head.
Where I’d point it, and where I wouldn’t
Start with the messy, wordy work
AI is best at language. It’s literally a Large Language Model (LLM). Summarising listening data. Drafting and reworking copy. Finding an image by describing it instead of guessing the filename. Reading a thousand open-text survey comments so nobody has to.
Handily, that’s also the least sensitive corner of your stack. Search tools, listening, design, asset management. Almost no personal data in any of it.
Best capability and lowest risk in the same place doesn’t happen often. Take the gift and don’t look the horse in the mouth!
Leave the numbers alone
Analytics is the last place I’d start, and not just for privacy reasons.
These models are probabilistic. Reporting isn’t. A number that’s roughly right is a number that’s wrong, and it’s wrong in the sneakiest possible way, because it looks fine until someone spends money on it. One made-up figure in a board pack burns more trust than a year of time savings earns back.
So BI stays the source of truth for numbers. When AI does eventually reach your analytics, it should fetch the figure and cite where it came from, never work it out itself. Build that rule in early. Retrofitting it after everyone’s got used to asking the chatbot for the CTR is a miserable conversation.
“Orchestration” is two jobs wearing one coat
This is where I see the most muddled thinking, and vendors are not helping.
Two completely different things get sold under the same word.
One is a person working across tools in plain language. Asking, drafting, deciding, approving. Human in the loop the whole way. An enterprise AI platform with connectors does this nicely.
The other is machines talking to machines. Scheduled syncs, event triggers, data unification, all running at 3am with nobody watching. That’s an integration platform or a CDP. Different tool, different budget, different owner.
An AI assistant is not an integration backbone. An integration platform is not a reasoning layer. You’ll probably end up with both. Just don’t buy one expecting it to do the other’s job, because you’ll find out six months in, after you’ve built process on top of it.
Prove it small before you ask for money
The instinct in a big company is to go and get a mandate first. Platform decision, budget line, integration programme, steering committee, kickoff deck. It’s the slowest route available and it loads all the risk onto a business case you can’t yet evidence.
Try the other way. A small, capped trial. Only the tools that need no build. No personal data anywhere near it. Get security to sign off the boundaries rather than the architecture. Fund it from what you’ve already got. Keep it short enough that you finish before the market moves.
Then go back with numbers.
Approvers say yes to small, bounded, reversible things far more easily than to open commitments. And the second conversation goes very differently when you’re holding your own results instead of a vendor’s slide.
Measure it like you expect to be challenged
This is where I think most AI pilots quietly fall over.
“The team reckons it saves about 30% of their time” is not a result. It’s a feeling with a percentage stapled to it. Self-reported time savings are always generous, impossible to disprove, and everyone in the approval chain knows it.
Pick three or four jobs you do the same way every month. The listening report. Researching and writing a campaign brief. The competitor and search-trend scan. Time them properly before you start, and note what you’re paying an agency to do. Do it again after. Publish the difference, including the ones that didn’t budge.
Marketers have earned the eye-roll our productivity claims get. Only way out is to measure like finance would.
Where I think this goes next
Three bets.
Governance becomes the moat, not model access. Everyone gets good models. The teams allowed to point them at real company data will be miles ahead of the teams still in a working group about it.
You’ll pick models per job. One for images, one for long copy, one for analysis. Standardising on a single model for everything will age about as well as standardising on a single font.
The boring work pays. If you spent the last few years tagging assets properly, tidying content, and sorting out permissions, you can wire up an AI layer in weeks. If you didn’t, you’re about to do all of it anyway, just faster and with someone senior watching.
None of this needs a moonshot. Start where the risk is low, measure honestly, and stay clear-eyed about which problems AI actually fixes.
If you’re doing something similar, I’d genuinely like to compare notes. Especially on measurement. That’s the part almost nobody’s cracked.