Why AI rollouts stall, and where your company’s brain should live
Eight problems that stall AI rollouts, from licences that change nothing to pilots stopped by security. Most come down to process, data and ownership, including where the company’s knowledge should live.
- Customer focus
- Process and ownership
- Shared data
- AI-supported work
A RevOps consultant asked me where his client’s company knowledge should live. That question sits behind most of the AI rollouts I see stall. Here are the eight problems I run into most, why a different model fixes none of them, and what I told him about owning the company’s brain.
- Most AI rollouts stall on process, data and ownership. Changing the model fixes none of that.
- Keep the company’s knowledge in a store the company owns, with agreed definitions on top, so every assistant gives the same answer.
- Give the work to a cross-functional team rather than to IT alone, and change one process at a time.
Where does the brain live?
I recently spoke with a RevOps consultant about one of his clients, an insurance broker with about a hundred people. The company already had AI tools. What it did not have was an answer to a simpler question: where does the brain live? Where should all of the company’s knowledge go, from call recordings and support tickets to sales emails, so that every person and every assistant works from the same picture?
He also asked whether Five Dots could be that place. It is not, and that is on purpose. Five Dots works inside the AI setup a company already has, and no business data is ever sent to us. The knowledge has to live somewhere the company controls.
The question stayed with me, because it sits behind most of the AI rollouts I see stall. When I list the problems companies run into while deploying AI, almost none of them is about the model. Switching from one assistant to another would not fix a single item on the list. They come down to process, data and ownership, which is another way of saying change management.
Eight problems I keep seeing
They fall into three groups, and they rarely turn up alone. A company with a messy CRM usually also has process documents nobody has updated in years, and both make the AI less reliable. That is how a team ends up checking every answer.
Eight reasons AI rollouts stall · tap the ones you recognise
Process
Data and instructions
Trust and ownership
Tap the problems you recognise to see where to start reading.
The process never changed
The most common pattern is also the simplest. A company buys licences for everyone and waits for the gains. Individuals do get quicker at drafting and summarising, and that is real. Meanwhile the monthly bill climbs, and the process around those people is the one they had last year. There is nothing anyone can point to and say: this now takes half as long.
The second is the demo that never ships. Someone builds an impressive agent and shows it around, and then nothing happens. It has no owner, it was never connected to live data, and no team changed how it works to make room for it. A demo shows that something is possible. The process only changes when someone decides that it will.
The third is easier to miss. In many of the processes I look at, the work itself takes a small share of the elapsed time. Most of it is waiting for an approval, a handoff or a reply, so making the work faster changes little. Try it on the made-up quote below.
Made-up example · one quote, from draft to send
Someone workingWaiting
A copilot on top of a mess
Companies want an assistant that can answer questions about their customers, while the data underneath sits in two CRMs, full of duplicates, next to an acquired business that was never merged. The assistant cannot know which record is right. It will answer anyway, and it will sound sure of itself.
Connecting agents to everything makes this worse. An agent with access to a dozen tools spends its tokens searching, and two runs of the same question can reach different sources and come back with different answers. Agents do better with a small set of tools and one place to get their data from.
Then there are the instructions. When IT builds an automation, it often starts from the standard operating procedures. Those are frequently out of date, or they describe how the process was meant to work years ago. An agent built on the wrong instructions does the wrong thing quickly and consistently. Before you automate a process, sit down with the people who run it and write down how it should work today.
All three problems lead back to the consultant’s question. Agents need one place to get their knowledge from, with clean data and agreed definitions.
Own the brain
My answer to him was that the company should own it, rather than handing it to a third-party tool, however convenient that tool is. Young AI companies come and go, and you do not want the centre of your business to disappear with one of them. You also want to stay free to change AI providers. If you move from ChatGPT to Claude next year, or use both, your knowledge should not have to move with you.
The best home is the data warehouse or database the company already runs, such as BigQuery or Supabase. The real work is getting everything into it, and agreeing the rules on top: who may see which data, and a staging area to test changes before they reach everyone.
On top sits a semantic layer, where the company writes down what its data means. Annual recurring revenue (ARR) is calculated one way, so whether someone asks in Claude, in HubSpot or in a dashboard, they get the same number.
You can start in Notion or Google Docs while you learn what people ask. Documents are hard to query and hard to keep current, though, so for anything robust, invest in the data layer. Our guide to running RevOps on Claude walks through the full setup.
- CRM
- Support tickets
- Emails
- Call transcripts
- Finance
- The vendor’s assistant
- Claudeon the vendor’s terms
- ChatGPTon the vendor’s terms
- Dashboardson the vendor’s terms
The knowledge sits in someone else’s product. Change AI provider, or lose the vendor, and it has to move.
Not a job for IT alone
The last group of problems is about trust. When the AI is wrong just often enough, someone checks every line, and the time saved goes straight back into checking. A better prompt rarely fixes that. Fixing the data and the definitions does, and then the checking can match the stakes.
Security and hosting reviews are the other trust problem. Pilots with a clear business case stop there because nobody asked where the data would go until the very end. Bring security in during the first week, not the last.
Which brings me back to the brain. I suggested that the consultant tell his client’s leadership not to let IT lead this alone. The company’s knowledge is the centre of everything, and too important to get wrong. It needs a small cross-functional team that builds it together, keeps it maintained, and settles most of the trust questions along the way.
Give everything the full check and the time saved goes into checking.
Who owns the company’s knowledge
- IT or the data teamThe platform and who has access
- RevOpsCommercial data and the definitions
- Subject leadsTheir own area: finance, legal, support
- SecurityInvolved from the first week
Where to begin
If you recognise a few of these, you are in good company. Most of the fixes are not technical. They are decisions about process, data and ownership, and they are easier to make one process at a time. Here is what to have in place before you put AI into one.
Before you roll AI into a process · tick what you have
0 of 6Start at the top: one process, mapped honestly.
Sources & further reading
Put the method to work.
Run the free GTM Audit in the Claude setup you already have.