AI value creation requires a measured baseline before any spend is approved

An operating partner who approves an AI roadmap in a portfolio company before anyone has established a baseline is funding activity against a vision. The pattern repeats across the first 100 days: a management team arrives at a board meeting with a slide listing copilots, model pilots and a data warehouse migration, and a capital ask attached. The CFO signs because the deal thesis mentioned AI, and six months later the forecast is unchanged and nobody can say which line item moved. AI value creation for portfolio companies fails at this exact point, when the spend is approved against a vision instead of against a measured starting position and a named owner.

This guide is written for the CFO, FP&A lead or data owner inside a portfolio company who has to decide what to fund, in what order, and how the result will show up in the numbers the board already tracks. It assumes budget and a thesis. It does not define standard terms.

1. Name the EBITDA line the AI work is supposed to move

Every AI initiative in a portfolio company should trace to a line in the operating model that a deal partner can point to: gross margin, sales and marketing as a percent of revenue, headcount per unit of output, days sales outstanding, or churn. If the initiative cannot be tied to one of those, it is a research project, and research projects do not belong in a value creation plan with a hold-period clock running.

Bain’s annual private equity report has tracked how the industry’s return expectations have shifted toward operational improvement rather than multiple expansion, which is the context in which AI spend gets judged. You can read the methodology and the yearly data in the Bain & Company Global Private Equity Report. McKinsey’s private capital research makes the same point from the operating side, that value creation plans now carry the weight the entry price used to; their work is collected at McKinsey.

The discipline here is the same one that governs any technology investment inside a private equity holding. Classify the expected impact before you fund it: realized, run-rate, forecast, or risk avoided. A customer-support copilot that cuts handle time is run-rate once it is live and measured. A demand-forecasting model that might improve inventory turns is forecast until the turns actually move. Do not let a forecast sit in the plan as though it were realized.

Tie Each AI Initiative to One Operating Line | table with columns Initiative, EBITDA line it moves, Impact class (realiz

2. Establish the baseline before you fund anything

You cannot claim an AI result without a number that existed before the work started. In most mid-market portfolio companies that number is hard to produce because the data lives in disconnected systems and the real reporting happens in Excel. That gap is itself the first finding, and it usually explains why the previous AI pilot could not prove anything.

Before the roadmap, require three things: a read-only view into the source systems, a single agreed definition of each metric the AI work claims to affect, and the current value of that metric with the method that produced it written down. Our team pulls Semrush keyword and competitor data for the 60+ domains we track when the metric in question is organic demand, but the principle holds for any metric: the baseline has to be reproducible by someone other than the person who built the model.

The sequence here mirrors how a readiness assessment works before capital is committed. The deeper version of this step is laid out in how to score a portfolio company’s AI readiness before you fund the roadmap, which is the right artifact to run during the first 100 days rather than after the budget is spent.

3. Decide the capability question honestly

The second decision is who runs the work. A portfolio company CFO rarely needs to hire a full data science team to get the first results, and hiring one before the foundation exists tends to produce expensive pilots that never reach production. The options are narrow: hire a permanent lead, contract a specialist, or rent an embedded team that runs the work and hands it back.

The judgment is the same one that governs any fractional technology engagement. The failure mode, documented in fractional CTO retainers fail without defined enterprise-value work, is paying for capacity without a defined piece of enterprise value attached to it. If you are weighing whether to rent the capability, how to buy fractional CTO services for PE-backed companies and judge whether they work sets out the judging criteria, and what a fractional CTO should own maps it to the operating model.

BCG’s work on principal investors covers how operating capability gets structured across a portfolio rather than rebuilt in each company, which is relevant when the same AI foundation can be reused across several holdings; their research hub is at BCG.

4. Sequence the work so the foundation comes first

Most AI value in a mid-market company sits below the model layer, in whether the CRM is governed, whether data is captured consistently, and whether the reports leadership looks at are the right ones. Advanced automation pitched before that foundation exists is the most common way AI budget gets wasted, because the model inherits the data problem and amplifies it.

The Order AI Value Actually Lands | 4-tier pyramid, bottom to top: Tier 1 Data visibility (read-only connections, unifie

Tier 1 and 2 are where the early return lives

Connecting ERP, CRM and martech into one model and shipping the first dashboards usually produces more decision value in weeks than a model would in months, because leadership can finally see the number the AI work was supposed to improve. PitchBook’s research and data on value creation levers is useful for benchmarking where peers in the same sector have put their effort; it is at PitchBook, and S&P Global Market Intelligence covers the same ground from the credit and operating-metric side at S&P Global.

Keep two output formats from day one

Finance keeps Excel on the same underlying data; leadership and the board get dashboards with insights flagged. Forcing one format on both audiences is how reporting automation breaks at the first board cycle. The mechanics of surviving that cycle are covered in portfolio company reporting automation that survives a board cycle.

5. Price and scope it against AI-era build costs

Entry-level AI and data work in 2026 should cost tens of thousands and deliver in days. A quote priced like a traditional time-and-materials data project is usually built on an older cost base, and the portfolio company CFO should benchmark it accordingly. The AICPA and CIMA have published guidance on how finance functions should evaluate AI and automation spend, available at AICPA & CIMA.

Scope the first engagement as a low-risk proof the champion can take to the deal team: read-only access, an NDA and MSA, a small fixed fee, anonymised proof of the finding, and a staged plan for what comes after. That structure is what lets an operating partner approve the next tranche against evidence rather than a pitch.

6. Judge the result against the baseline, not the activity

The number of tickets closed, features shipped or prompts tuned tells you the team was busy. It does not tell you the EBITDA line moved. The judging standard is the baseline from Step 2 compared against the current value, with the same method, plus a classification of whether the delta is realized, run-rate or still forecast.

This is also where governance matters, because AI work touches data that diligence and the board will eventually question. The Harvard Law School Forum on Corporate Governance has tracked board-level AI oversight expectations at Harvard Law School Forum on Corporate Governance, and Harvard Business Review’s M&A coverage addresses how these capabilities are valued at the next transaction, collected at Harvard Business Review. Preqin’s data on how LPs now scrutinise value creation claims is at Preqin, and the trade reporting in Private Equity International, Buyouts and PE Hub shows how these results read to the market at exit.

Where the AI work touches security or regulated data, the diligence standard is higher. Cyber due diligence that changes what you pay and the broader technology due diligence process both intersect with any model that handles customer data, and the SEC’s disclosure guidance applies once those risks become material, published at the U.S. Securities and Exchange Commission.

7. The decision checklist before you sign

  • Every initiative names one EBITDA line it moves and one accountable owner.
  • A reproducible baseline exists for each metric, with the method written down, before funding.
  • The capability decision (hire, contract, or embed) is tied to defined enterprise-value work.
  • Data visibility and reporting hygiene are funded before applied AI.
  • Finance keeps Excel; leadership and the board get dashboards with insights flagged.
  • The first engagement is priced against AI-era build costs and scoped as low-risk proof.
  • Results are judged against the baseline and classified as realized, run-rate or forecast.

Choosing the monitoring layer that holds all of this together is its own decision; how to judge a private equity portfolio monitoring solution before you sign covers it, and technology leadership that actually moves EBITDA covers who should own the agenda.

8. Implementation note

The practical move for a portfolio company CFO is to run the baseline and capability assessment as a small, fixed-scope diagnostic before the roadmap budget is committed, so the first dollar of AI spend is approved against a measured number and a named owner rather than a vision. That diagnostic is also the artifact that lets the operating partner defend the spend at the next board meeting and the finding that survives to the next transaction.

If you want that baseline and value creation diagnostic run inside a portfolio company, route the engagement to the DevriX private equity data practice, which runs the read-only assessment, the metric baselines and the staged plan as an embedded retainer.