Every deal memo in circulation this year carries an AI value-creation thesis. Most of them have the same flaw: not in their conclusion, but in their premise. They assume that if a company is AI-ready, AI will compound value. If it is not, you budget a modernization sprint and proceed.
The problem is how "ready" gets defined. In diligence conversations, it defaults to a technical inventory: cloud infrastructure, data stack, model deployment capability. Some firms have formalized this into a maturity rubric, L1 through L5 or some equivalent scoring grid. The rubric produces a number. The number goes in the memo.
The number does not predict outcomes.
Here is what the data actually shows: McKinsey's State of AI 2025 report found that roughly two-thirds of organizations running AI initiatives have not moved those initiatives past the experimentation or pilot stage to enterprise scale, despite widespread tool adoption, stated AI strategies, and dedicated AI headcount. These are not companies that lack AI ambition. What they lack is a specific set of operational conditions that determine whether AI compounds into the business or cannibalizes margin while looking productive.
There are three of those conditions. Each is measurable. None requires a technical audit. And a deal team that reads them correctly before close will make materially better capital allocation decisions than one working from a maturity score.
Why Maturity Scores Miss the Point
A maturity rubric answers the wrong question. It asks: how sophisticated is this company's relationship with technology? The question that matters in a deal context is: will AI multiply or erode margin here?
Those are different questions with different answers. A company can maintain a world-class data platform and still fail to extract AI value if its core workflows are manual, fragmented, and unmonitored. Conversely, a company that scores modestly on infrastructure can be primed for AI compounding if its processes are digitized, its data is current, and it owns its outcome signal.
An AI readiness assessment built for PE diligence does not produce a score. It produces a signal read: three operational dimensions that predict whether AI will compound value internally or leak margin to vendors and manual workarounds. Here is how to measure each one.
Signal 1: Data Decision Latency
Definition:** How long does it take the business to answer a new operational question with data it already owns?
This is not about data quality in the abstract. It is about how fast data moves from the business into insight: whether the company can answer an unplanned question in days or in weeks. Ask the CFO or COO: when did you last need an answer your standard reports did not provide, and how long did it take from question to reliable number? Days is operational. Weeks is a structural blocker.
Proxy metric:** Time from novel business question to data-backed answer, measured in calendar days. Pull it from the last three or four operational decisions that required data the company did not have pre-packaged in a standard report. This question can be asked cold in any management session without infrastructure access.
The compound vs. cannibalize read:
Low latency (days) signals that data is instrumented, queryable, and owned. Pipelines run. There is a single source of truth for the metrics that drive decisions. AI layered into that environment has current, reliable inputs to work with. It accelerates decisions already being made with data. Value compounds.
High latency (weeks) signals manual data assembly: analysts downloading CSVs, reconciling spreadsheets across disconnected systems, waiting for monthly reporting cycles to close. AI applied to that environment does not accelerate decision-making. It adds a sophisticated inference layer to a broken data loop. Outputs look polished. They reflect stale or mismatched inputs. That is cannibalization: the AI spends inference cost while producing the appearance of insight without the operational substance.
What a low score costs post-close:** Every AI initiative that depends on current data (demand forecasting, churn prediction, dynamic pricing) will need the data loop rebuilt before it can function correctly. That rebuild is not a sprint. Fixing a fragmented reporting infrastructure takes three to six months of dedicated engineering work. Catch it at diligence and it can be sequenced correctly. Discover it at month 14 and it becomes an unplanned interruption in an AI roadmap already committed to LPs.
Signal 2: Process Digitization Depth
Definition:** What share of the core value-chain workflow runs through software today, versus email, spreadsheets, and manual handoffs?
AI compounds when applied to digitized processes. Applied to manual ones, it adds cost without adding lift. This is not a judgment about legacy companies. It is a mechanical constraint. AI systems need structured inputs and defined outputs. A workflow where inputs arrive in email threads and outputs are typed into cells is not something you augment with AI. You have to replace it. That is a different scope, a different timeline, and a different capital requirement.
Proxy metric:** Walk the primary revenue-generating or margin-critical workflow end to end. Count the software-mediated steps (logged, time-stamped, searchable in a system of record) against the human-mediated steps handled by email, document, or verbal handoff. Express this as a ratio. A workflow that is 80% software-mediated is extensible. One at 40% requires foundational digitization work before AI can touch it productively.
The compound vs. cannibalize read:
High digitization depth means the data for AI already exists as a byproduct of normal operations. The system logs what happened. AI can read that log, identify patterns, surface anomalies, and improve the workflow with each cycle. For deal teams sizing AI upside in the thesis, this is the signal that the projection is operational, not aspirational.
Low digitization depth means the workflow produces no data trail. AI cannot learn from what it cannot see. Deploying AI into this environment forces a choice: accept that AI will operate without feedback, producing unmonitored outputs; or digitize the workflow first, then apply AI. The second path moves the ROI timeline out 12 to 18 months before the first production deployment generates real return.
What a low score costs post-close:** Discovering required workflow digitization mid-hold carries a heavier price tag than any other pre-condition on this list. It is capital-intensive, operationally disruptive, and does not produce AI value during the build. Pricing it into the thesis at diligence and sequencing it into the 100-day plan is tractable. Discovering it at month 14 against a timeline that has already been presented to investors is a different problem entirely.
Signal 3: Feedback-Loop Ownership
Definition:** Does the company own its outcome data and the ability to improve its AI systems over time, or does that improvement accrue to a vendor?
This signal separates companies building a proprietary data flywheel from those renting generic model capability. Both approaches can produce results in the first 12 months. Only one compounds into durable competitive advantage by exit.
Proxy metric:** Ask the management team how AI outputs improve over time. Is there a process for capturing when an AI recommendation was accepted or rejected, and feeding that signal back into the system? Who controls that feedback data: the company or the vendor? If the answer involves a third-party SaaS AI tool with no export capability and no customization path, the company is renting inference. It is not building a flywheel.
The compound vs. cannibalize read:
Owned feedback loop: each production cycle generates training signal that stays inside the company. The system improves. The data advantage widens. Competitors running the same underlying model cannot replicate the accumulated learning, because the learning lives in proprietary data and usage history no one else has access to. Value compounds internally, and by exit, the AI capability reads as defensible IP rather than a software subscription.
Rented feedback loop: the vendor captures usage signals across its entire customer base, improves the generic model for everyone, and the company receives no differentiated benefit from being an early or heavy user. Switching costs accrete to the vendor over time. Margin leaks as dependence grows on a capability the company does not control. That is not a technical risk. It is a strategic one that shows up in the exit multiple when a sophisticated buyer prices the lack of proprietary data moat.
What a low score costs post-close:** Rented loops are a different failure mode than the other two signals. The first year looks fine. The AI works, outputs are reasonable, the use case delivers. The cost surfaces when the company tries to extend AI to adjacent workflows and discovers it needs a new vendor contract, a new integration, and outputs it cannot own or port. By exit, the AI story holds up in demos and falls apart in acquirer diligence when there is nothing proprietary underneath it.
Running the Three-Signal Read in Diligence
None of these signals require infrastructure access. Each can be read from a management session with five targeted questions:
Data decision latency:** "Walk me through the last time you needed an answer your standard reports didn't give you. When did you ask the question and when did you have a reliable number?"
Process digitization depth:** "Take your highest-volume workflow. What percentage of steps leave a system-of-record timestamp versus a note in someone's inbox or a row in a spreadsheet?"
Feedback-loop ownership:** "How do your AI tools get better over time, and who controls the data that makes them better: you or the vendor?"
The output is not a score. It is a blocker map: which signals are strong, which require remediation, and whether the remediation fits inside the hold timeline at reasonable capital cost. A deal team running this read at diligence has two things a maturity-rubric buyer does not: a realistic picture of what the AI thesis requires before it produces returns, and a 100-day plan that sequences the right infrastructure work before the first production deployment.
What Low Scores Cost at Exit
Deal teams that skip the signal read at entry tend to encounter the consequences at exit, not as a valuation event but as a set of questions from sophisticated buyers that require uncomfortable answers.
A buyer who asks "what is your data decision latency for your core demand signals?" and receives a blank look knows immediately that the AI capability in the CIM is aspirational. A buyer who sees high digitization depth and an owned feedback loop across two or three workflows has something to underwrite as operational. That difference in underwriting confidence does not produce a modest valuation gap. For AI-positioned businesses in the current market, it produces a meaningful multiple gap, particularly as acquiring PE firms and strategic buyers develop their own signal-reading capability.
Getting these three right during the hold is what earns the right to put AI in the exit narrative and have a buyer who can verify it believe you.
What Blueprint Measures
The three signals above describe what an AI readiness assessment built for PE diligence needs to measure to be actionable. That is the design specification behind Blueprint.
Blueprint is Blue Orange Digital's AI-readiness scan for PE portfolios. It runs in about ten minutes, requires no infrastructure access, and produces a blocker map tied directly to these three operational dimensions. The output is not a maturity score. It is a sequenced action plan: which signals are blocking the AI thesis, what each blocker costs to resolve, and what the 100-day path to first production deployment looks like from the current state.
If the AI value-creation story in a current or prospective deal is being underwritten on management confidence rather than a signal read, a Blueprint scan is the fastest path to knowing what you are actually working with, before the capital goes in.
Josh Miramant is founder and CEO of Blue Orange Digital, an AI and data engineering firm that has delivered more than 250 production AI deployments across PE-backed mid-market companies over the past decade.