An AI value creation plan for a private equity portfolio company is the part of the value creation plan (VCP) that names the operating loops where AI will change the economics of the business, the baseline each loop is measured against, and the order in which the work happens. It is a sequence of closed loops, not a list of pilots.
The distinction matters because the industry has plenty of pilots. In early 2025, Bain reported that among private investors representing $3.2 trillion in assets, a majority of portfolio companies were in some phase of generative AI testing and development, while nearly 20% had operationalized generative AI use cases and were seeing concrete results. In the 2026 StepStone/Bain survey of 103 GP professionals, 39% said they did not expect AI to have any material financial impact on portfolio companies in 2026. Activity is everywhere. Expected value is not.
The broader corporate data says the same thing. BCG's The Widening AI Value Gap found that 5% of firms worldwide are generating value from AI at scale, while 60% report minimal revenue and cost gains despite substantial investment. A portfolio is a sample of that distribution. The operating partner's job is to move companies out of the 60%, on a hold-period clock.
§ 01 / THE PROBLEMHow a sponsor learns a company is off plan
Start with how a sponsor learns that a company is off plan. In most portfolios the answer is the board meeting. Books close weeks after quarter end, someone assembles a deck, and the deck is presented long after the decisions it describes. We made this argument in a short field note, The quarterly board deck is an open loop: a quarterly deck is not feedback, it is a record of a quarter that can no longer be changed.
That matters for AI specifically, for two reasons. First, every AI initiative in a VCP needs a baseline and a way to read its effect. If the only instrument is the quarterly deck, an initiative launched in month two cannot be evaluated until month six, and its effect is blended with everything else that happened that quarter. Second, the Closed Loops pillar of our methodology holds that the operating physics of an AI-Native company depend on outcomes flowing back to the decisions that produced them. A firm that sees its companies at quarter resolution cannot close any loop that matters at week resolution.
A use case you cannot measure is an opinion. The reporting loop is what turns it into evidence.
§ 02 / THE LEVERSWhich value-creation levers AI actually moves
Be conservative here, because the evidence is. In the same GP survey, 46% of respondents expected cost savings or efficiency to be the primary AI outcome in portfolio companies in 2026, and 10% expected revenue growth. Bain's 2026 midyear report describes cost reduction as "often an early focus," while noting AI is also improving sales and customer acquisition and enhancing pricing sophistication. Four levers are defensible in a VCP today. None of them comes with a promised EBITDA number, and a plan that leads with one is guessing.
- Cost-to-serve in document- and decision-heavy loops. Order entry from emailed purchase orders, invoice matching, claims intake, quote preparation, credit applications. These loops have high volume, a definable correct answer, and a human doing transcription or routing. That is where models earn their keep first.
- Working capital. Collections prioritization, cash application, dispute resolution, and demand signals feeding inventory decisions. The lever is not a model; it is a decision about which invoice to chase or which SKU to reorder, made daily rather than at month end.
- Pricing. Quote-level price decisions, discount approvals, and floor enforcement. Where pricing runs through email chains and deal-desk queues, moving the routine cases into a logged, rule- or model-driven decision is a workflow change before it is a data science problem.
- Sales coverage. Which accounts get attention this week, which lapsed customers get called, which quotes get followed up. Coverage is a scheduling problem over a ranked list, and many companies have the data to rank but no loop to act on it.
What AI does not do is substitute for the plan. Bain's September 2026 piece on the AI value paradox in private equity puts it directly: start with the VCP, then isolate the three to five key opportunities in it and map those workflows. The levers above are a menu. The VCP decides which one is on the plate.
§ 03 / THE FIRST LOOPClose the reporting loop first
Before any model touches a decision, the company's operating numbers need to flow continuously into one queryable layer and be read weekly by the people who can act. This is what real-time portfolio monitoring should mean in practice: not a prettier dashboard on the quarterly pack, but order, quote, invoice, and cash data landing in one place at the cadence the business runs.
In methodology terms this is the first slice of the AI Factory's publish-subscribe data layer, and the start of the Intelligence Layer. The work itself is plumbing: source systems (ERP, CRM, billing, the bank feed) publish into a shared store under documented definitions, and a small set of operating metrics is computed from it, the same way every week.
This first loop is worth doing for three reasons that have nothing to do with AI. The board sees drift earlier. Management stops reconciling competing spreadsheets. And every later initiative gets a baseline, which is the difference between a value creation claim and a value creation fact.
§ 04 / THE SEQUENCEA 100-day sequence for one portfolio company
The sequence below is a framework, not a promise of results by a date. It is the hold-period form of the 90-day pilot-to-production roadmap, stretched to 100 days to fit a VCP's first board cycle. Its purpose is to make sure that by day 100 the company has one working closed loop and a measured answer about whether to build the next one.
Days 1 to 30: diagnose and baseline. Score the company on a consistent maturity instrument. Map the three to five VCP initiatives to candidate operating loops. Pick the weekly metrics each loop will be judged on and start capturing them from source systems, even if the first version is crude. Name an owner inside the company for each candidate loop.
Days 31 to 70: close one decision loop. Choose the single loop with the clearest baseline and the highest volume. Rebuild it end to end: inputs arrive under a defined contract, the routine decision is made by a rule or model that logs its reasoning, exceptions route to a named human, and the outcome of every decision is recorded against the case. In the parts distributor above, that might be quotes below list: routine cases priced and approved in seconds, unusual ones escalated, every outcome (won, lost, margin realized) captured.
Days 71 to 100: measure and decide. Compare the loop against its baseline. Check the exception rate, the cycle time, and the business metric the VCP cares about. Then make an explicit call: scale the loop, change it, or stop it. A stopped loop with a clean measurement is a better outcome than a running pilot nobody can evaluate.
By day 100 the deliverable is not a model. It is one loop with an owner, a contract, and a metric.
This is the same pattern we describe in our anatomy of Ant Financial's origination platform, at a far smaller scale: rebuild the data supply, move the routine decision into a system that logs its reasoning, and relocate the human from the line to the edges of it.
§ 05 / THE PORTFOLIORunning the plan across a portfolio
A single company can be run as a project. A portfolio needs a method. Three things make the plan repeatable.
A repeatable assessment. Every company is scored on the same instrument, so differences are not artifacts of different frameworks. Our Maturity Diagnostic is a fast, self-scored starting point: eight questions produce a radar across six axes (Data, Decisions, Workflow, Architecture, Talent and Strategy) and place the company in one of three bands, AI-Enabled, AI-First or AI-Native. What each axis measures is covered in a separate essay.
Shared architecture patterns. The reporting layer, the decision-loop pattern, and the evaluation harness should be designed once and reused. Portfolio companies differ in what they sell, much less in how an invoice, a quote, or a purchase order moves through a system. A pattern proven in one distributor should transfer to the next with configuration, not a new program.
Comparable maturity scores. With one instrument across the portfolio, the operating team can see which companies are capped by their data, which by decision rights, and which by talent. That changes where operating-partner time goes. A company weak on Data needs the first loop before anything else. A company strong on Data but weak on Decisions is ready for a decision loop now.
§ 06 / THE ASKWhat operating partners should ask for
Whether the work is done internally or with outside help, the questions are the same.
- Which VCP initiative does this serve? Every AI initiative should trace to a named line in the plan. If it does not, it is a pilot.
- What is the baseline, and when was it last measured? If the answer is "last quarter's board pack," the reporting loop is not closed yet.
- Who owns the loop inside the company? A loop owned by an outside team stops when the team leaves.
- What happens to the exceptions? A loop with no defined exception path is either unsafe or quietly still manual.
- How is the outcome of each decision recorded? Without it, the loop cannot learn and the sponsor cannot audit.
- What does the company keep? The working infrastructure, and the internal capability to extend it, should outlast the engagement.
Those questions map onto how we work. Our engagements start with a four-week, fixed-fee assessment that produces a written report naming the three highest-leverage interventions and a 90-day execution roadmap. An architectural engagement then rebuilds one operating system to AI-Native standard over 90 to 180 days: typically a data pipeline, a decision system, or an experimentation platform, delivered as working infrastructure plus the internal capability to extend it.
The fastest way to see where a portfolio stands is to run the same instrument across it. Have each company's leadership take the Maturity Diagnostic, compare the radar shapes, and bring the results to an engagement conversation. The axis most companies share as their weakest is the natural place to look for the first loop worth closing.
END
§ FAQ / READER QUESTIONSQuestions readers ask
What is an AI value creation plan in private equity?
An AI value creation plan ties AI work in a portfolio company directly to the initiatives in its value creation plan. It names the few operating loops where AI can change the economics, such as cost-to-serve, working capital, pricing, or sales coverage, sets a measurable baseline for each, and sequences the work so the company first sees its own operating numbers weekly and then automates specific decisions against them.
How should operating partners use AI across portfolio companies?
Operating partners get the most leverage by standardizing the method rather than the tools. That means one repeatable assessment for every company, a small set of shared architecture patterns for data, decisions, and feedback, and a comparable maturity score so the firm can see where each company stands. Each company then picks its own highest-value loop from its value creation plan.
What does a portfolio company AI assessment measure?
A useful assessment measures the operating machine, not the list of AI tools in use. It asks where operational data goes, who or what makes recurring decisions, what triggers work and how drift is caught, how load-bearing AI is to daily operations, who can ship a change to a model, and where the roadmap's center of gravity sits. Scored the same way everywhere, the results are comparable across companies.
What should a 100-day AI plan for a portfolio company include?
A credible 100-day AI plan has three phases. First, baseline the business and put weekly operating numbers into one queryable layer. Second, pick one recurring, high-volume decision tied to the value creation plan and rebuild it as a closed loop with an owner and a metric. Third, measure the result against the baseline and decide whether to scale, change, or stop.
Why does real-time portfolio monitoring come before AI use cases?
Without continuous operating data, a firm learns a company is off plan at the quarterly board meeting, weeks after the quarter closed. AI use cases built on that foundation cannot be measured either, because there is no baseline to compare against. Putting weekly operating numbers into one queryable layer gives the board earlier signal and gives every later AI initiative a way to prove its value.