Nobody can tell you when “AI-Ready” Is Done
Most AI budgets fund what happened. Returns come from what you decide next. The Compounding Company Brain Series.
Part 1 of The Compounding Company Brain, a three-part series on why AI-readiness programs struggle to show ROI, and what to build instead.
Every conversation now starts with the same two letters. #AI
Board meetings. Earnings calls. Vendor pitches. The standup that was supposed to be about something else. AI, and then a version of the same question: is our data ready, is our company ready, are our workflows ready.
Nobody argues with the question. It sounds like the responsible thing to ask.
The trouble starts when you ask what “ready” means, because that is where the answers stop agreeing.
Many names, no finish line
Count what your own company has called this in the last eighteen months. AI-native tech stack. Making your data AI-ready. Enterprise second brain. Context layer. Agent-ready foundation… so on.
Different labels, one destination and the destination is never drawn. Something past search, past retrieval, past context. Everyone points at it. Nobody has specified it.
It now appears in almost every enterprise technology roadmap. Almost none says what completion/ key milestone looks like.
I watched someone finally ask about it in a steering meeting. Not aggressively. Just: what does AI-ready look like when it’s finished?
The answer was a roadmap. More sources connected, more coverage, more of the same thing. Which is not what was asked.
The program was already in its third budget cycle.
A question everyone is asking, and an answer nobody can give. That distance is what this piece is about, and what it is quietly costing you.
An unbounded precondition is the perfect thing to sell
Eventually someone asks these programs for a return. When they can’t produce one, the usual explanation is that the pilots failed.
Mostly they didn’t. They worked roughly as demonstrated.
The real problem is simpler. You can compute a return on infrastructure. You cannot compute a return on ready not until someone says which decision it was meant to improve.
Ask which decision got better and the room goes quiet. Not because nobody worked hard. Because nothing in the program was ever scoped to a decision.
That is a category error, not an execution failure. Readiness is a prerequisite being funded as an outcome.
And a project with no completion criteria cannot be finished which means it cannot be failed, which means it cannot be measured. That asymmetry does not favor the person paying for it.
No bad faith is required. Integrators sell phases because that is how services get scoped. Platforms sell consumption because that is how infrastructure gets priced. Every incentive is rational on its own. Together they produce a program designed to continue rather than to finish.
The readiness gap
Here is the more useful way to see it. Three layers, and most AI-readiness spending stays concentrated in the first.
Memory tells AI what happened.
Context tells it what matters.
A decision layer tells it what to do next.
Readiness programs overwhelmingly fund memory. Warehouses, catalogs, connectors, coverage all necessary, all real, and all of it returns something. But the return executives are asking about the business outcome only appears at the decision layer, where something is actually decided differently.
Some companies do fund pieces of the middle and far layers a rules engine here, a personalization program there but rarely as one governed system that anything else can reuse.
That distance is the readiness gap. Nearly every program I see is funding one end of it and being asked to justify itself at the other.
Which is why the honest answer to "what did eighteen months buy us" is often: an excellent memory, and no change in what anyone decides on Monday.
Programs report progress in the left column while executives expect returns from the right one.
The gap won’t close with a better model. It exists because context and decisioning rarely have a durable owner, a shared budget line, or a reusable operating model — while memory has had all three for a decade.
Which is why the honest answer to “what did eighteen months buy us” is often: an excellent memory, and no change in what anyone decides on Monday.
Jack Dorsey and Roelof Botha proposed a useful test earlier this year in From Hierarchy to Intelligence: what does your company understand that is genuinely hard to understand, and is that understanding getting deeper every day?
That points at the alternative, and it’s worth naming here because the rest of this series builds on it.
A compounding company is one where the context, rationale, and outcome behind one decision become reusable inputs to the next.
Memory accumulates. It does not automatically become organizational learning. Decisions compound only when their context, rationale, and outcomes are captured for reuse.
The judgment nobody captures
There is also a source of margin hiding in plain view, and it has nothing to do with your data estate.
Take one decision, and keep it in mind this series comes back to it three times.
Should this customer or account get this offer, right now, through this channel?
Two lifecycle leads make that call a hundred times a week. Same platform, same budget, same eligibility rules, same approval checklist. One consistently delivers better margin.
Every operator has watched this happen. Nobody has a line item for it.
The difference isn’t the platform, the budget, or the formal rules. It’s accumulated judgment that was never written down:
Which segment got burned out last quarter and needs to be left alone.
Which market punishes a discount for reasons no model contains.
Which exception exists because of a supplier agreement signed three years ago that nobody has read since.
Externalizing expertise is not a new ambition. What has changed is the cost of not doing it.
Judgment doesn’t scale the way everything else you fund does. Capital scales through allocation. Software scales through replication. Judgment does neither. It sits with one person, walks out at five o’clock, and eventually leaves permanently taking the reason for the exception with it.
The company already paid to develop that judgment. The readiness program was supposed to make it reusable. Mostly it hasn’t.
Unlike readiness, judgment attaches to a decision and a decision has a frequency, a dollar value, and an outcome. For the first time you have something to put a number against.
The reader is no longer human
One more change, and it’s what puts a clock on all of this.
Most enterprise knowledge systems were designed with a person at the end of them. A person retrieved the document and could recognize when it looked incomplete, stale, or wrong.
That assumption is now gone in both directions at once.
Inside your company, agents act on what they retrieve, with none of that reflex. Meta’s engineering team documented this on a 4,100-file pipeline: agents performed poorly until the tribal knowledge was pre-computed into context files, and the system was built to re-validate itself periodically because that context decays.
The implication generalizes beyond code. Missing context produces an unhelpful answer, which you catch. Stale context can produce a confident one, which you may not.
Outside your company, something is evaluating you without a human ever reading your homepage. A procurement agent scoring your documentation. A research assistant summarizing your terms for a buyer who will never visit your site. A shopping agent checking price, availability, and eligibility.
This is further along than it looks. Reviewing go-to-market tooling recently, I found more than thirty published MCP servers machine-readable interfaces that let agents query, and sometimes act through, business systems. Few executive teams can say who approved these interfaces, what data they expose, or which actions they permit.
Your agents read your context. Increasingly, so do the agents evaluating you. Only one side is within your control.
That’s the deadline a readiness program doesn’t have. Readiness can slip a quarter forever. The other side of that conversation isn’t waiting.
Start with one decision that pays
You can’t close the readiness gap by understanding your enterprise. Enterprises are too large, the coverage question never terminates, and you’ll be back in the same steering meeting next year with better memory and the same silence.
You close it one decision at a time. Start where decisions recur often, carry visible money, and return a result fast enough to learn from.
For most companies, that points to the customer.
In Part 2, I’ll rebuild that offer decision as an Audience Brain Loop, and show how to measure whether the organization is actually learning.
Because even the best learning loop fails when nobody owns the judgment behind it. That’s Part 3.
If this changed how you’re reading your own AI budget, forward it to one person who needs it. Reply if you want to argue the pieces get sharper when readers push back.
I agree on the notion of what Decision is being impacted. Decision Intelligence is the next frontier. Check out Cloverpop.com. I am on the board AND they get it.