Three common mistakes that waste millions across portfolio companies — and the operational approach that actually delivers measurable EBITDA improvement from AI.
Every PE firm we talk to has AI on its value creation agenda. The operating partners have attended the conferences. The portfolio company CEOs have been briefed. The board decks include "AI transformation" as a strategic priority. And yet, across the portfolio, the actual EBITDA impact of AI initiatives is negligible. Pilot projects that never scaled. Vendor contracts that delivered dashboards instead of decisions. AI teams that built impressive demos but couldn't explain how they'd move the P&L.
This is not because AI doesn't work. We have deployed AI agents across a £6.4M transformation programme at a major UK retailer, where seven AI agents run 24/7 across every department, directly enabling headcount optimization, pricing automation, and vendor replacement. AI works. But the way most PE firms deploy it does not.
Three mistakes account for the majority of failed AI value creation efforts.
The most common first move: hire a Head of AI (or "VP of Data and AI" or "Chief AI Officer") and give them a team. They recruit machine learning engineers. They set up an ML infrastructure stack. They start building models. Six months later, they have a churn prediction model with 87% accuracy that nobody in marketing knows how to use, a demand forecasting model that the supply chain team doesn't trust, and a recommendation engine that sits in a staging environment because the e-commerce platform can't integrate it.
The problem is not the team's technical ability. It is their total disconnection from the operational reality of the business.
An AI team that reports to the CTO builds what is technically interesting. An AI capability that reports to the operations of the business builds what is financially impactful. The difference is not subtle — it is the difference between a cost center and a value driver.
At the retailer, we did not start with AI. We started with a full operational audit. We mapped every cost line, every vendor contract, every process, and every headcount line. Then we identified where AI would deliver the highest EBITDA impact. The in-house CDP was built because the £45K vendor was replaceable. The pricing engine was built because the £75K pricing vendor was replaceable. The AI agents were deployed because specific human tasks could be automated. Every AI deployment was tied to a specific financial outcome before a line of code was written.
If your AI team cannot tell you, in pounds, what their next quarter of work will deliver to the P&L, you have a research lab, not a value creation function.
The second pattern: the portfolio company buys AI-powered SaaS tools. An AI-powered customer service platform. An AI-powered pricing tool. An AI-powered marketing automation system. Each one costs £40-150K per year. Each one solves one problem in one department. None of them talk to each other.
We see this constantly. A retailer paying £75K for a pricing vendor that runs elasticity models on their data, £45K for a CDP vendor that segments their customers, £30K for a visual search tool, and £25K for an SEO analysis platform. Total spend: £175K per year on "AI" that the business doesn't own, can't customize, and can't integrate.
The point solution approach fails for three reasons:
Data fragmentation. Each vendor has a partial view of the business. The pricing vendor sees transactions but not customer segments. The CDP vendor sees customer profiles but not pricing decisions. The recommendation engine sees browsing behavior but not margin data. The most valuable insights come from connecting these data sources, and the vendor model structurally prevents that.
Vendor lock-in and cost escalation. SaaS pricing is designed to grow with usage. As your customer base grows, the CDP costs more. As your product catalog grows, the pricing tool costs more. As your traffic grows, the search tool costs more. You're paying a margin on top of the compute cost, and that margin compounds with scale.
Lack of competitive differentiation. Every one of your competitors can buy the same tools. If your pricing advantage comes from a vendor that any competitor can also subscribe to, it is not an advantage. It is a commoditized capability that you're renting.
The alternative is to build core AI capabilities in-house, on your own data infrastructure, where they can be customized, integrated, and compounded. At the retailer, we replaced £175K in annual vendor costs with in-house builds that run on their existing Snowflake warehouse. The pricing engine uses their margin data, their inventory data, and their customer segments simultaneously — something no single vendor could do. The CDP uses churn prediction and discount suppression that identified £8.3M in margin protection — features no vendor CDP offered.
Building in-house is not always the right answer. It requires engineering capability and ongoing maintenance commitment. But for PE portfolio companies where the hold period is 3-7 years, the compounding value of owned AI infrastructure far exceeds the convenience of rented point solutions.
Curious what your margin opportunity looks like?
Free Tool
How much margin are you leaving on the table?
Answer 6 questions. Get a personalised margin estimate in under 2 minutes.
Take the Free Margin AuditThis is the deepest and most damaging mistake. The PE firm frames AI as something the technology team does. It sits under the CTO's budget. It is governed by the technology steering committee. It follows the IT project methodology: requirements, design, build, test, deploy. It is treated like a software implementation.
AI value creation is not a technology project. It is an operational transformation that uses technology as a tool.
The distinction matters because it changes who is accountable, how success is measured, and where the work happens.
A technology project is measured by: Was it delivered on time? Does it work? Is it stable? These are necessary but insufficient conditions for value creation. A perfectly delivered AI model that nobody uses has zero EBITDA impact.
An operational transformation is measured by: Did the P&L move? By how much? In which line items? These are the questions that matter to PE firms, and they can only be answered by people who understand the operations of the business, not just the technology.
At the retailer, our transformation tracker has 119 initiatives, each with an EBITDA value. The AI agents are not tracked as "technology projects" — they are tracked as business outcomes. The CX Agent is valued at the headcount it enables removing. The Pricing Agent is valued at the vendor it replaces plus the revenue lift it generates. The SEO Agent is valued at the external consulting spend it eliminates. There is no "AI budget." There are business outcomes that happen to use AI as the execution mechanism.
When AI sits under the CTO, the conversation is about models, infrastructure, and technical debt. When AI sits under the transformation programme, the conversation is about margin, headcount, and revenue. PE firms want the second conversation. Most are having the first.
Having seen what doesn't work, here is what does. Four principles that we apply at every engagement.
1. Start with the P&L, not the technology. The first step is always a full operational audit. Map every cost line. Map every revenue driver. Identify where the margin is leaking. Only then ask: "Which of these problems can AI solve better, faster, or cheaper than the current approach?" This produces an AI roadmap that is directly tied to financial outcomes, not a technology roadmap that hopes to find financial outcomes later.
2. Deploy AI agents, not AI models. A model sits in a notebook and waits for someone to use it. An agent runs autonomously, makes decisions, and produces outputs. The pricing agent doesn't predict optimal prices and wait for a human to implement them — it generates markdown recommendations, monitors results in real-time, and adjusts. The CX agent doesn't score sentiment and produce a dashboard — it triages tickets, responds to first-line queries, and flags escalations. The difference between a model and an agent is the difference between advice and action.
3. Replace vendors before building net-new capabilities. The fastest ROI from AI is vendor replacement. Every SaaS tool that uses AI to process your data is a candidate for replacement with an in-house build. The pricing vendor (£75K saved). The CDP vendor (£45K saved). The visual search tool (£30K saved). The SEO platform (£25K saved). Each replacement not only saves the vendor cost but gives you ownership of a capability that can be customized and compounded. Vendor replacement delivers immediate, measurable EBITDA impact — and it builds the in-house AI muscle for more ambitious projects later.
4. Track everything against EBITDA. No AI initiative should exist without a specific, quantified EBITDA value. "Explore the potential of generative AI" is not an initiative. "Deploy an AI agent to automate product description generation, reducing merchandising headcount by 1 FTE (£42K annual saving)" is an initiative. The discipline of EBITDA-tagging forces the team to justify every project in financial terms before it starts. It also makes portfolio-level reporting straightforward: sum the EBITDA values across all AI initiatives, across all portfolio companies, and you have the total AI value creation number for the fund.
PE firms have a structural advantage that most companies do not: multiple portfolio companies that face similar operational challenges. The pricing engine built for one retailer can be adapted for another. The headcount optimization methodology applied to one business can be templated for the next. The automation-first approach to headcount reduction works across any business with repetitive operational processes.
This is where the real value creation lies — not in deploying AI at one portfolio company, but in building a repeatable AI value creation playbook that can be run across the entire portfolio. The first deployment is the most expensive. The second is 60% of the cost. By the third, you have a methodology, a set of proven agents, and an operational team that can execute in weeks, not months.
But this only works if the first deployment is done correctly: operationally led, EBITDA-tracked, and integrated into the business rather than bolted onto the technology stack.
Most PE-backed AI initiatives fail not because AI is immature, but because the approach is wrong. Hiring technologists to solve business problems. Buying rented capabilities instead of building owned ones. Treating transformation as a technology project instead of an operational programme.
The firms that are extracting genuine EBITDA improvement from AI are doing something different. They are starting with the P&L. They are deploying agents, not models. They are replacing vendor costs with owned capabilities. And they are tracking every initiative against its financial impact with the same rigor they apply to any other value creation lever.
AI is not a magic lever. It is a tool. Like any tool, its value depends entirely on who wields it and what problem they are solving. Give it to technologists and you get technical projects. Give it to operators and you get EBITDA improvement.
Book a free margin audit. We work with PE firms and their portfolio companies to build AI value creation programmes that deliver measurable EBITDA improvement — not slide decks.
We go into businesses and make them permanently more profitable. Every initiative is EBITDA-tracked.