# Why most PE portfolio AI projects fail — and what to do differently PE portfolio AI projects fail for predictable reasons. Learn the sequencing mistakes operating partners make — and how to protect capital and create real value. Published: 2025-03-13 Author: Rodan Analytics There is a pattern that repeats itself across PE portfolios with uncomfortable regularity. A portco management team pitches AI as a value creation lever during the hold. The operating partner nods. A vendor gets selected, often quickly. Six months later, the project is either stalled, descoped or quietly abandoned — and nobody is quite sure what went wrong. What went wrong is usually not the technology. It is the sequencing. Most portfolio AI projects fail because they are initiated as technology deployments rather than business transformations, and because the infrastructure required to make AI work — clean data, clear ownership, integrated systems — was never in place to begin with. This article will tell you why the conventional approach to AI in portfolio companies is structurally broken, what the actual failure modes look like in practice, and what a better sequencing looks like — one that protects capital and creates genuine equity value. --- ## The failure mode nobody talks about in the IC meeting Portfolio AI projects rarely die with a bang. They die slowly, through scope reduction and distraction. The typical trajectory: a promising use case gets identified — usually something in commercial ops, pricing or customer service. A vendor demo impresses the CFO. A budget is approved. Implementation begins. Then reality arrives: the data is fragmented across three systems, the ops team is too stretched to support the change, and the vendor's definition of "done" turns out to mean the model is trained, not that anyone is using it. The project limps to a soft launch. Adoption is low. The management team moves on to the next priority. The vendor renews on a reduced contract or exits. The board never gets a clean post-mortem. This is not a technology story. It is a change management and data readiness story. A manufacturer in the industrials sector, for example, might spend £300k on a demand forecasting tool, only to find that their ERP data has three years of gaps caused by a system migration. The model cannot learn from data that does not exist. The vendor's commercials assumed clean inputs. Nobody checked. The lesson is not "AI does not work in manufacturing." The lesson is that AI amplifies the quality of your data infrastructure — it does not replace it. --- ## Why the standard operating partner playbook does not transfer Operating partners are exceptionally good at applying proven playbooks: lean operations, procurement consolidation, commercial excellence frameworks. These work because the underlying variables are well understood and the interventions are proven at scale. AI does not behave like a procurement initiative. The value is highly context-dependent. A pricing optimisation model that delivers 4% margin improvement at one portco may be useless at another with different SKU complexity and customer concentration. You cannot lift and shift. The second issue is talent. Most portcos at the £500m–£1bn revenue scale do not have a chief data officer. They may have a capable head of IT and a small analytics team running Excel and Power BI. Deploying an enterprise AI system into that environment without interim data leadership is like fitting a performance engine into a car with no brakes. The operating partner's instinct — move fast, drive adoption, demonstrate value in the first hundred days — is exactly the wrong cadence for AI. The first hundred days should be diagnostic, not deployment. Rushing to deployment before assessing data maturity is the single most common and most expensive mistake in the market. A practical reframe: instead of asking "which AI use case should we prioritise?", the first question should be "do we have the data and the organisational capability to execute any AI use case at all?" --- ## What good sequencing actually looks like The projects that work follow a consistent pattern. It looks like this: - **Conduct a proper data and AI readiness assessment** before any vendor conversation. Map data sources, assess quality and completeness, identify ownership gaps and integration blockers. - **Define a narrow, high-value use case** that maps to a specific P&L line. Not "improve customer experience" — "reduce churn in the top-200 accounts by identifying at-risk signals 60 days earlier." - **Appoint or embed data leadership** before deployment begins. This might be a fractional CDO engagement, an operating partner with data specialism or a dedicated resource from an external adviser. - **Build the data foundation in parallel** with vendor selection, not after. Data pipeline work and system integration should start on day one. - **Pilot with a defined success metric** and a defined exit condition — both for continuing and for stopping. A B2B software portco following this approach recently ran a six-week diagnostic before committing to a £400k AI-led churn prediction programme. The diagnostic revealed that their CRM data was missing usage signals for 40% of the customer base. They spent eight weeks fixing the data infrastructure first. The model they subsequently deployed achieved meaningful predictive accuracy. The programme delivered. The diagnostic cost £8k and saved them from a failed deployment that would have cost multiples of that. Sequencing is not bureaucracy. It is how you protect the investment. --- ## The due diligence blind spot Most technology due diligence in M&A focuses on system architecture, technical debt and cybersecurity exposure. Very few processes include a serious assessment of data maturity. This is a significant oversight at a time when AI is increasingly cited as a value creation thesis at acquisition. If the investment memo says "we will deploy AI-driven pricing optimisation in Year 2," the technical due diligence should validate whether the data infrastructure makes that feasible. If it cannot validate that, the assumption should be treated as speculative — and the cost of building the required infrastructure should be modelled into the deal. A consumer brand acquired on the basis of D2C growth potential, for example, may have years of transaction data sitting across three separate e-commerce platforms with no unified customer identity. The "AI personalisation" thesis depends entirely on resolving that fragmentation first. That is an 18-month programme, not a 90-day quick win. Operating partners and deal teams who build data maturity assessments into their diligence process will consistently make better acquisition decisions and set more realistic value creation timelines. Those who do not will keep inheriting the same problem after close. --- ## From project to programme: building AI capability that compounds The portcos that extract sustained value from AI are not the ones that ran the most projects. They are the ones that built a capability. There is a meaningful difference between deploying a point solution and building an organisational capability that can run, iterate and expand AI applications over the hold period. The former requires a vendor. The latter requires leadership, governance and a data foundation that persists beyond any single initiative. This matters for exit. Buyers — both strategic and financial — are increasingly assessing AI capability as a valuation input. A portco that can demonstrate repeatable, governed AI deployment across commercial and operational functions is a more attractive asset than one that has a single AI tool bolted onto an otherwise unchanged business. Building that capability is not a five-year programme. With the right sequencing, a portco can move from low data maturity to a functioning AI capability within 12 to 18 months. But it requires treating AI as a strategic initiative with dedicated resource and executive sponsorship — not as a vendor contract managed by the IT team. --- The operating partners who are getting this right have changed one thing: they ask harder questions earlier. They do not let management teams lead with the use case. They ask about the data first. They insist on a diagnostic before a deployment decision. They treat AI readiness as a hold period asset, not a Year 3 problem. The cost of not doing this is not just a failed project. It is a compressed multiple, a missed value creation lever and an exit story that does not hold up to scrutiny. If you are sitting on a portfolio AI initiative that has stalled, or preparing to launch one, the right first step is a structured diagnostic — not another vendor conversation. Rodan's diagnostic engagements are designed specifically for PE-backed businesses: a focused, time-limited assessment that tells you what is actually true about your data and AI readiness, and what to do about it. Book a diagnostic at rodan.io. --- **Meta description:** PE portfolio AI projects fail for predictable reasons. Learn the sequencing mistakes operating partners make — and how to protect capital and create real value. HTML: https://rodan.io/insights/why-most-pe-portfolio-ai-projects-fail-and-what-to-do-differently