It's a missing pipe.
A lot of organisations think they have an AI problem. They have an integration problem with an AI sticker on it.
Hat tip to Jamie Claret of Autonomate, who used the framing "AI is everywhere, but why doesn't it feel joined up?" at the Vistage AI for Growth Forum and made it the cleanest one-line description of this whole pattern I've heard.
Once you see the difference between joined-up and joined-on, you can't really see it any other way.
Joined-on
The shape goes like this. The CRM gets a copilot. The service desk gets summarisation. The finance system grows an AI assistant. Each one is locally impressive. Each one is locally correct. Each one looks great in the vendor demo.
And each one is being asked to give useful answers about a business it can only see a third of.
The CRM copilot doesn't know what was just invoiced. The service desk summary doesn't know which contract the customer is on. The finance assistant doesn't know which deal the cash relates to. Each AI is doing its job perfectly inside its own silo, and the staff using them are doing the integration the AI can't — by hand, in their head, between tabs.
That's joined-on AI. Intelligence bolted onto a system without changing the plumbing around it. It looks like progress for about three months. Then quietly, everyone goes back to copying things between tabs, because the answers from the AI are right but only kind of useful.
Why it keeps happening
This isn't a stupidity problem. It's the natural result of three things lining up at once.
Vendors ship features, not estates. Microsoft ships Copilot inside its products. Salesforce ships Einstein inside its products. Each vendor's incentive is to make their AI great inside their box. None of them is incentivised to fix what happens between the boxes — that's somebody else's job, and historically that somebody else has been the customer's IT team. The AI features arrive faster than the integration work needed to make them collectively useful.
Boards see demos, not data flows. A demo of an AI assistant inside a CRM is a five-minute wow. A discussion about whether the CRM and the finance system know about each other in real time is a forty-minute conversation with the IT director that involves the words middleware and idempotency. Guess which one wins the room. The shiny thing always wins, and it's nobody's fault that it does.
Integration is unfashionable. API plumbing has been the unsexy bit of IT for twenty years. It's the work that keeps a business running and almost never the work that gets credited. Adding AI to an integration project doesn't suddenly make it interesting; it just makes it more expensive.
So the path of least resistance is: turn on the AI inside whichever system you're already paying for, hope the answers are good enough, and quietly absorb the rekeying as a cost of doing business.
Joined-up
The other version isn't more complicated. It's just a different ordering.
You decide what the actual end-to-end work is — quote arrives, gets processed, customer onboarded, work scheduled, invoice raised, payment chased, ticket logged when something goes wrong — and you make the systems talk to each other across that path. Identity once. Customer record once. Two-way sync of the things that change. Events fired when state moves. Audit trail across the whole thing.
Then you put AI on top.
When you do it that way, the AI has something real to work with. The CRM copilot can reference the invoice that was raised yesterday. The service desk summary can pull the contract terms. The finance assistant can see which deal the cash matches. The intelligence is the same intelligence — but it's now answering questions about a business, not a database table.
This is the bit our missing pipe, not a missing brain line was reaching for. The bottleneck isn't intelligence. The bottleneck is the place where information stops moving.
What this looks like as actual work
Less mystical than it sounds. Concretely, when we run a joined-up engagement, the substance is mostly:
- A clear picture of the end-to-end work — the actual customer, employee or operational journey, not the org chart
- Identity that flows: one person, one access path, the right things visible at the right moments
- Two-way sync between the systems that need to know about each other — CRM and finance, service desk and ticketing, identity and onboarding, comms and audit
- Events fired when state changes, so downstream systems can react in real time rather than via overnight batches
- Ownership of the pipes — somebody, somewhere, has the runbook, the monitoring, and the responsibility when one of them goes wrong
- Then the AI: copilots, summarisers, assistants, agents — sitting on top of an estate that already moves information properly
That's the API and integration work that makes the rest of it stick. It is genuinely the bit that determines whether the AI investment ends up earning its keep or quietly becoming a line item nobody can quite explain.
The same thinking sits underneath Continuous Compliance. The reason that service exists is the same reason this one matters — the value the customer keeps long-term sits below the model, in the connective tissue. The model is hot-swappable. The pipes aren't.
Close
A clever model on top of a disconnected estate is a confident assistant who can only see one window of your business at a time. It will give you good answers, fluently, about the third of the picture it has access to.
A modest model on top of an estate that's properly joined up will quietly outperform that. Every time. Because it's answering questions about the business, and the other one is answering questions about a database.
The work, almost always, is the plumbing.
Then the cherry on top.