Most firms find this out in the wrong order. You buy the tool, wire it into what you already have, watch it work in the demo, and watch it produce something mediocre in production.
It's the layer underneath every AI tool you will ever buy.
McKinsey's April 2026 report on reimagining infrastructure for agentic AI puts a number on it: less than 10% of agentic programs reach meaningful scale.
What "inadequate infrastructure" looks like inside a real firm
The phrase sounds like a technology problem. In a studio or a practice you'd recognize it as a set of small frictions you have lived with for years.
Your discovery call notes are in an inbox or a transcription account. Your proposal is in Drive, your signed contract is in HelloSign, your project record is in Notion or Asana. You keep the invoices in QuickBooks and the client emails in Gmail.
None of those systems talk to each other. You hold the thread between them in your head, which works because you're good at it and because you've been doing it for years.
Then you ask an agent to draft a proposal from a discovery call. It has to read from one system, look up context in three more, and write to a fifth. There's no single record for it to pull from, no consistent shape across the records that do exist, and no clean handoff between systems. It does the best it can with what it can reach, and you read the output and go back to doing it yourself.
That result is what your data layer permits. Any tool you buy will hit the same ceiling.
Why agentic AI exposes this and chatbots never did
For two years the data layer stayed invisible.
Assistive AI runs on whatever you paste into it. You are the data layer. You select the relevant facts, you supply the context, and the model never touches anything you didn't hand it.
Agentic AI runs on the data layer directly, reading and writing across systems on its own. Scatter your records and you get incoherent output. Consolidate them and you get your firm's own intelligence, amplified.
McKinsey stops short of naming a single cause. We think the data layer is most of it, and we would say so as our reading and not as their finding.
The small firm version, which is editorial work
In the firms we have done this with, the consolidation pass takes a few focused weeks, and almost none of that time is engineering. At enterprise scale you can't buy that timeline at any price, which is why the enterprise framing makes this sound harder than it is for you.
One. Pick one source of truth. Notion, a CRM, a project tool, something purpose built. One, not three. You'll hold the canonical version of every client, every project, every decision there. It may not make sense to pull all of your business into one source, so consider keeping your financial data separate from your project management data.
Two. Decide what lives where. Put the discovery summary in this field, the agreed pricing in that one, the decision log here. Write the decisions down, walk your team through them, and treat the schema as something you'll revise.
Three. Backfill enough recent projects that an agent has something real to read on its first runs. Not all of them. Enough.
Four. Wire the things that produce work into it, so records update as work happens and nobody has to remember.
Most of that is making decisions you've been carrying unwritten and putting them on paper. You can do it while the business keeps running.
Related questions
What is AI data infrastructure readiness?
The condition where an agent can read what it needs, in a consistent shape, from one place, and write back without breaking other systems. A firm that glues five systems together with somebody's memory isn't there yet.
Why do firms only discover this after launching AI?
Assistive AI runs on whatever a human pastes in, which keeps the data layer invisible. Agentic AI runs on the layer directly, so its problems surface immediately. Most firms formed their habits in the assistive era.
Does a small firm need an enterprise data platform to run agents?
No. The small version is one source of truth plus a documented schema for what lives where, and the work is editorial.
Should we pause an AI pilot to consolidate first?
Usually yes. A few focused weeks of consolidation costs less than finishing a pilot on a broken data layer and then consolidating anyway.
What does McKinsey report on agentic AI at scale?
That less than 10% of agentic programs reach meaningful scale, in its April 2026 report on reimagining infrastructure for agentic AI.
How we read this
A firm with a real source of truth is faster on its own terms, and that has nothing to do with AI. You stop re-explaining decisions. Your team stops asking the same question twice. A new hire gets useful sooner. A client asks about something from March and you give them the actual project history instead of a version reconstructed from memory.
So we'd sequence it this way regardless of what you decide about agents. Do the consolidation and then choose never to add AI, and you've still made the firm materially better. Add agents without it and you've bought amplification for your own confusion. You have one project here, not two, because the work that supports agents is the work that supports the firm running well. AI is what finally makes it urgent enough to do.
It's unglamorous, which is why you have left it, and it's among the few things in this whole conversation that compound. Every tool you add afterward starts from a higher floor.
What you can do this week
Write down where six things actually live right now. Not where they're supposed to live. Where they are.
The last discovery call. The current version of a proposal. The signed contract. The decision you made in week three of a project that changed the scope. The vendor lead time you learned the hard way. The reason a past client left.
If those six answers name more than two systems, you've found your first project, and it isn't a software purchase.
Working together
Flow State Found works with a limited number of businesses to make their best work their baseline. Most of the firms we meet have the answer written down somewhere, across four tools, in three formats, none of them current. We do the consolidation pass first, because looking is how you find out whether it is a week of work or the whole engagement.
We take on limited engagements, so it starts with a conversation.
Start a conversation