By Josh Miramant, Blue Orange Digital
Private equity has always been in the business of underwriting what a management team can execute. Revenue growth assumptions, EBITDA margin improvement targets, headcount efficiency ratios. All of it traces back to the same question: can this team actually do it?
So why are PE operating partners underwriting AI differently?
Right now, most firms approach AI capability the way they once approached technology modernization: as a function of what a vendor can build. They scan the CIM, call a few portfolio company CTOs, and ask whether the target has a data science team or a model deployment infrastructure. If the answer is yes, they mark the AI box checked. If no, they budget a few hundred thousand dollars for a "digital roadmap" in the first-100-days plan.
This is the wrong framework. And it's getting more expensive to be wrong.
The FTI 2026 PE Value Creation Index identified AI and M&A as the two top value drivers across the asset class. That is not a trend to watch. It is a mandate to act on. Portfolio companies that fail to operationalize AI in this hold period will enter their next transaction at a material valuation discount. Operating partners who cannot diagnose and accelerate AI adoption are leaving multiple turns of EBITDA on the table.
The correct framework for underwriting AI is not "what can this team build?" It is "what can this team redesign and run at cost?"
That shift in question unlocks everything else.
Why Build Capability Is the Wrong Lens
The build-capability lens made sense when AI was expensive to deploy and required specialized infrastructure. It does not apply anymore. The intelligence layer is now infrastructure. Large language models accessible via API have commoditized the core capability. Any reasonably technical team can call a model. The actual differentiator is whether a company can redesign its workflows around what those models can do, and whether it can sustain that redesign at a manageable operating cost.
Two cost dynamics matter before you can evaluate AI with any precision.
First: inference is the majority of AI spend. Across enterprise AI deployments, inference, meaning the ongoing cost of running models against live data, accounts for roughly 85% of total AI spend. Training and fine-tuning are relatively minor line items by comparison. This means that evaluating AI investment based on upfront build cost dramatically underweights the sustained operating cost that will show up in EBITDA for the duration of the hold.
Second: agentic workflows amplify token consumption in ways most operators do not anticipate. When a company moves from one-shot AI interactions (ask a question, get an answer) to agentic workflows (an AI system that reasons through multi-step tasks, calls tools, and iterates on outputs), token consumption can jump 5 to 30 times relative to simpler deployments. This is not a reason to avoid agentic AI. The productivity gains are real and significant. But it means a company that designed its AI architecture around pilot-scale economics is likely to hit a cost cliff when it scales.
The operating partner who understands these dynamics can do something valuable: benchmark the portfolio company's AI architecture against what it will cost to run at scale, not just what it cost to build.
Diligence: What to Evaluate Before You Close
AI diligence is still immature at most PE firms. The questions are too high-level ("do you use AI?"), too technology-focused ("which models are you running?"), or too input-oriented ("how many data scientists do you have?"). None of those questions predict value creation.
Here is a sharper diligence framework built around the right question: what can this team redesign and run at cost?
Workflow redesign depth, not pilot count
Ask the management team to describe their three most significant AI use cases. Then ask: was the underlying workflow redesigned to remove the human steps that AI replaced, or was AI added as a layer on top of the existing process? Companies that have truly operationalized AI have reduced headcount in the workflows AI now handles, or reallocated that headcount to higher-leverage work. Companies still in pilot mode have AI as an additive cost, not a substitution.
The tell is in the P&L. If a company claims meaningful AI adoption but headcount in the relevant functions has not changed and productivity per head has not improved, the AI is ornamental.
Inference cost awareness
Ask the CFO or CTO what the company spends on AI inference per month. If they do not know the number, or if they conflate it with licensing fees for AI tools, that is a signal that cost at scale has not been modeled. A company with mature AI operations tracks inference cost as a line item, can articulate the unit economics per workflow (cost per document processed, cost per customer interaction handled), and has some form of cost governance in place.
Team redesign vs. team augmentation
The companies that generate durable AI value have reorganized their teams around what humans need to own when AI handles the execution layer. This typically means fewer people doing rote tasks and more people doing review, exception handling, and judgment calls that require context AI does not have. Companies that have just added AI tools on top of unchanged teams have increased their software cost without reducing their labor cost. The math does not work at scale.
Data readiness: be specific
Data readiness is the most overused term in AI diligence and the least precisely evaluated. The right question is not "do you have clean data?" but "is your data organized in a way that supports the specific workflows where you plan to apply AI?" A company can have a perfectly maintained CRM and still be completely unready to use AI in its finance function because the relevant finance data lives in PDFs and spreadsheets. Evaluate data readiness workflow by workflow, not as a global asset.
Dependency on rotating AI talent
If the company's AI capability is concentrated in one or two individuals who came in from outside on a consulting arrangement, that is a concentration risk. The underlying knowledge of what workflows are being changed, why, and how, lives in those individuals' heads, not in the organization's operating model. Ask what happens to the AI roadmap if those people leave. If the answer involves rebuilding from scratch, the AI capability is not embedded yet.
The First 100 Days: From Assessment to Operating Model
The first 100 days serve two functions in an AI context: they generate the information you need to prioritize, and they establish the operating model that will carry you through the hold.
Days 1-30: Map the workflow landscape
Before identifying AI opportunities, you need a clear inventory of the company's workflows and a rough sense of their economics. This does not need to be exhaustive. You are looking for the highest-leverage targets, not a complete audit. Focus on workflows that are high-frequency (done many times per day or week), relatively structured (the inputs and outputs are predictable), and currently handled by people whose time is more valuable elsewhere.
Document these workflows at the level of: what triggers the work, what information is used, what decisions or outputs are produced, and who reviews or acts on the output. That documentation becomes the substrate for AI redesign.
Days 31-60: Select and sequence
Prioritize based on two dimensions: speed to production (how quickly can a working version be deployed?) and economic impact (what is the labor or time cost of the current workflow?). The combination creates a 2x2: quick wins that validate momentum, strategic bets that require more setup, efficiency plays in the middle, and things to deprioritize.
A critical decision in this phase is whether to use off-the-shelf AI tools that require minimal configuration, or to build more customized workflows using the company's own data. The right answer depends on the workflow. But the default should be to reach for the off-the-shelf tool first. Custom builds are not inherently better. They are more expensive to maintain and slower to ship.
Days 61-100: Deploy, instrument, and stabilize
The first production deployments in this window serve a second purpose beyond the direct value they generate: they surface the cost and governance issues you will need to manage at scale. What is the actual inference cost for this workflow? What is the error rate, and what is the review overhead required to catch errors? Are there latency issues that affect user adoption?
Instrument everything from day one. The companies that struggle with AI at scale are usually the ones that did not build measurement into their early deployments and are therefore flying blind when cost or quality issues emerge.
The second major output of the first 100 days is an embedded AI operating model. This is distinct from the AI roadmap. The roadmap is a list of what you plan to build. The operating model is the structure that will execute the roadmap, maintain what has been built, and keep the organization's AI capability current as the technology evolves. The roadmap without the operating model is just a slide deck.
Hold-Period Value Creation: The Embedded Team Advantage
The most common mistake PE firms make in AI value creation is treating AI capability as a consulting engagement. They bring in an AI firm or a functional AI operating partner, run a transformation program, declare success, and exit. The portfolio company has a set of AI tools it did not build and a team that does not fully understand them. Eighteen months later, half the workflows have drifted back to manual because the AI system needed maintenance no one knew how to do.
The alternative model, and the one that generates durable hold-period value, is embedded context.
An embedded AI team is not a rotating roster of generalist AI consultants who apply the same playbook at every portfolio company. It is a small team, often two to five people, who work deeply inside a specific portfolio company, know that company's data, workflows, and competitive context, and build AI solutions specifically calibrated to how that company operates.
The difference shows up in outcomes. An external AI operating partner typically spends significant time on discovery: learning the business, mapping the workflows, building the context that an embedded team already has. That discovery overhead is not trivial. On a six-month engagement, it can consume six to eight weeks. An embedded team with 12 months of context can deploy in weeks what an external team would spend months understanding.
There is also a knowledge retention dynamic. When an embedded team deploys an AI workflow, the reasoning behind the design lives in the team's institutional memory: why that data source was prioritized, what edge cases the model struggled with, how the review process was set up. That knowledge compounds over time. When a rotating external team deploys and rotates out, that knowledge leaves with them.
The economic model matters too. The total cost of an embedded AI team (salary, benefits, tools) is typically lower than the total cost of a sustained consulting engagement at comparable output levels. More importantly, the cost structure is predictable and can be explicitly planned against the hold-period operating model.
Reframing AI ROI for the Hold Period
AI return calculations for PE portfolios tend to focus on cost savings in specific workflows. That is legitimate, but it is a partial view. The more complete ROI framework captures three value pools.
Operating efficiency. Direct cost reduction from AI handling work that people previously did. This is the most legible and easiest to model. It shows up as headcount reduction or headcount growth avoidance in growing functions, measured against a labor cost baseline.
Revenue impact. AI-enabled workflows that accelerate or improve revenue generation. This includes AI-assisted sales processes that improve conversion rates, AI-powered customer service that reduces churn, and AI-driven pricing or demand forecasting that improves margin. These are harder to isolate but often larger in aggregate than the operating efficiency gains.
Organizational capability. A team that knows how to identify, deploy, and maintain AI-enabled workflows becomes more valuable as AI capability continues to compound. A portfolio company with a mature AI operating model will absorb new AI capabilities faster and more cheaply than one perpetually in pilot mode. At exit, this translates to a faster buyer diligence process and a defensible narrative around AI-driven competitive moat.
Underwriting all three pools together produces a more accurate picture of AI value than workflow-level ROI calculations in isolation.
The Playbook in Practice
Here is the framework condensed for operating partners who need a repeatable process across a portfolio.
At diligence: Test for workflow redesign depth (not pilot count), inference cost awareness, team reorganization evidence, and AI talent concentration risk. Evaluate data readiness per target workflow, not as a global asset.
In first-100-days planning: Prioritize by speed to production and economic impact, not by technological ambition. Deploy with instrumentation from day one. Establish an embedded AI operating model as an explicit output of the first 100 days. Not just an AI roadmap.
Through the hold: Resist the consulting engagement model for AI. Build or source an embedded team with deep portfolio-company context. Track inference cost as an operating line item from the first deployment. Measure value across operating efficiency, revenue impact, and organizational capability, not just workflow-level cost savings.
At exit: Frame the portfolio company's AI capability in terms of operating model maturity, not technology stack. Buyers who understand AI will look for evidence that the capability is embedded, scalable, and maintainable. Not that the company has impressive demos.
Why This Matters Now
The FTI index is not identifying AI as a future trend. It is documenting what is already separating the top performers in the asset class from everyone else. The operating partners who develop genuine fluency in AI value creation, not just familiarity with AI tools, but the ability to diagnose, design, and execute AI-enabled operating model changes, will have a structural advantage in portfolio value creation for this hold-period generation and the next.
The framework is not complicated: underwrite what a team can redesign and run at cost. Apply it at diligence, get the first 100 days right, and build the embedded model that compounds through the hold.
That is it. Everything else is noise.
Blue Orange Digital is an AI and data consultancy that builds embedded teams inside PE-backed portfolio companies. We specialize in AI strategy, workflow redesign, and the operating model infrastructure that turns AI pilots into durable value creation. If you are working through the AI diligence or first-100-days framework at a current or prospective portfolio company, we would welcome the conversation.