# How to measure the ROI of AI investment in a portfolio company How to measure the ROI of AI investment in a portfolio company — a practical framework for PE operating partners covering baselines, attribution and exit narrative. Published: 2025-02-13 Author: Rodan Analytics Most AI investment in portfolio companies gets measured the wrong way, or not measured at all. Operating partners approve a budget line, a system gets deployed, and six months later someone is asked whether it "worked". By then the baseline is gone, attribution is impossible and the answer is usually a shrug dressed up as a case study. The mistake most PE firms make at this stage is treating AI investment like IT infrastructure — necessary, broadly beneficial, hard to quantify. That framing lets mediocre implementations survive and good ones go unrecognised. It also makes it nearly impossible to build a credible AI value creation thesis across a portfolio. This article gives you a practical framework for measuring AI ROI in a portfolio company from the point of investment decision through to exit. It covers how to set the right baseline, which metrics actually matter, how to separate AI contribution from other value drivers, and what good looks like when you are presenting to an IC or preparing an asset for sale. --- ## Why standard ROI frameworks fail for AI Traditional capex ROI works when outputs are discrete and causally linked to the investment. You buy a machine, it produces units, you count units. AI does not work like that. AI systems affect decision quality, throughput, error rates and staff capacity — often simultaneously, often indirectly. A demand forecasting model might reduce working capital, lower warehouse costs and improve customer fill rates at the same time. Which number do you report? The other problem is the counterfactual. When a manufacturing business in your portfolio reduces procurement costs by 8% in the year after deploying an AI sourcing tool, how much of that is the AI, how much is commodity price movement and how much is the new CPO you hired at the same time? Without a structured approach, you cannot say. There is also the timing problem. Some AI benefits compound over time as models improve and adoption deepens. Early measurement will understate value. Measure too late and you have lost the baseline entirely. The solution is not to wait for a perfect attribution model. It is to instrument the investment before you deploy it. --- ## Set the baseline before the project starts This is the most consistently skipped step and the most consequential one. Before any AI system goes live, document the current-state metrics in the specific areas the AI is intended to affect. Do not rely on management accounts — they aggregate too much. You need operational metrics at the process level. A practical baseline pack for a portfolio company deploying AI in a commercial function should include: - The volume of the task or decision the AI will touch (transactions per week, quotes processed, tickets handled) - The current cost to complete that task (fully loaded, including management time) - The error rate or quality measure (forecast accuracy, pricing variance, customer complaint rate) - The cycle time (how long it takes to complete the process end-to-end) - The revenue or margin outcome associated with that process where measurable For example: a B2B distribution business at £650m revenue is deploying an AI-assisted pricing tool. The baseline pack captures that the pricing team currently processes 1,200 requote requests per month, taking an average of 22 minutes each, with a win rate of 34% and an average discount of 11.4% against list. That is your baseline. In six months you measure the same five numbers. The delta is your value case. This level of specificity is uncomfortable for management teams who are used to narrative reporting. Push through it. If you cannot define the metric before deployment, you almost certainly cannot measure it after. --- ## Separate value creation from value enablement Not all AI value lands on the P&L directly. Some of it is structural — it makes the business capable of doing something it could not do before, or it removes a constraint that was limiting growth. For measurement purposes, separate three categories: **Direct P&L impact.** Revenue increase, cost reduction or margin improvement directly attributable to the AI system. This is the number your IC wants. Isolate it, document it, be conservative on attribution. **Capacity creation.** Hours freed from manual tasks that have been redeployed to higher-value work. This is real value, but only if you can show where the capacity went. Headcount reduction is the simplest proof point. Redeployment to growth activity is harder to measure but worth documenting with specific role-level evidence. **Structural capability uplift.** The business can now do something it could not before — personalise at scale, forecast at SKU level, respond to market movements in hours rather than weeks. This is a valuation argument, not a trading argument. Quantify it at exit through revenue quality, margin resilience and competitive positioning, not through a monthly dashboard. A consumer goods portfolio company running at £900m revenue might show £2.1m in direct procurement savings, 4.2 FTE of commercial capacity redeployed to account management, and a newly built capability to run dynamic promotional pricing across 40,000 SKUs. Each of those is real. Each requires different evidence to substantiate. Do not collapse them into a single "AI delivered £X" number — that obscures more than it reveals. --- ## Build the measurement infrastructure into the deployment ROI measurement is not a finance exercise you run after implementation. It is an architectural decision you make at the start. When commissioning an AI deployment in a portfolio company, the technical specification should require that the system logs the decisions it makes and the outcomes those decisions produced. An AI that cannot produce an audit trail of its own performance is not enterprise-ready — and in a portfolio context, it is not investment-ready either. Practically, this means: - Define the KPIs before vendor selection, not after - Require that the platform produces decision-level logging, not just aggregate dashboards - Build a control group or holdout set where operationally feasible — even partial controls improve attribution significantly - Integrate AI performance reporting into the monthly management information pack from go-live, not retrospectively - Assign ownership — someone in the business is responsible for the number, not just for the tool The third point is consistently underused. A logistics business deploying AI route optimisation across a fleet of 200 vehicles can run the AI on 150 of them and keep 50 on the existing system for 90 days. The difference in fuel cost, delivery completion rate and driver utilisation is as clean an attribution as you will get in a real business. It requires operational will to hold the control group, but it is worth it. --- ## What the exit story looks like When you are 18 to 24 months from exit, the AI investment narrative needs to shift from operational metrics to enterprise value arguments. Buyers and their advisers will not pay a multiple on a cost saving. They will pay a multiple on a structural capability that generates durable margin and reduces operational risk. The exit pack for an AI-enabled business should contain: - A clear before-and-after on the two or three metrics that matter most to the acquirer's investment thesis - Evidence that the capability is embedded (adoption rates, management dependency, process integration) not just deployed - A roadmap of unrealised value — what this capability will do in the next owner's hands with further investment - Clean documentation of the data assets the AI relies on, because a sophisticated buyer will run technical due diligence on this A technology-enabled services business at £1.1bn revenue with an AI-driven client insight layer has a fundamentally different quality-of-earnings story than one without it — if the documentation exists to prove it. If it does not, the buyer's advisers will discount it or ignore it entirely. --- ## The cost of getting this wrong Operating partners who do not build measurement infrastructure before deployment will spend the back half of the hold period trying to reconstruct value that cannot be reconstructed. The baseline is gone. The control group never existed. The system has been iterated several times. You are left with management assertions and a chart that went up and to the right, which any decent buy-side adviser will interrogate. The firms that do this well treat AI measurement as part of the value creation plan from day one — the same rigour they apply to commercial due diligence or operational improvement programmes. If you are at the point of approving AI investment in a portfolio company and you do not yet have a measurement framework in place, that is the right place to start. Rodan runs focused diagnostic engagements — typically completed in two to three weeks — that establish the baseline, define the KPI architecture and set the reporting infrastructure before deployment begins. It is a modest investment that protects a much larger one. [Book a diagnostic with the Rodan team.](https://rodan.io) --- **Meta description:** How to measure the ROI of AI investment in a portfolio company — a practical framework for PE operating partners covering baselines, attribution and exit narrative. HTML: https://rodan.io/insights/how-to-measure-the-roi-of-ai-investment-in-a-portfolio-company