An operating partner at a mid-market private equity firm running a buy-and-build should have the platform match each add-on's customer and supplier records against its own with AI within 60 days of closing. The report to the deal team comes before any system integration starts: the revenue from shared customers, the spend with shared suppliers.
FTI Consulting, which sells transaction and value creation services to private equity firms, published its 2026 Private Equity Value Creation Index on June 4, 2026, from a survey of 555 senior private equity leaders across 14 countries between January 19 and February 17, 2026, with no fund size or deal size stated for this sample. M&A is the slowest, with only 25% of firms achieving results within 12 months and just 35% of firms describing M&A implementation as efficient or very efficient, the lowest among all levers.
The matching is entity matching, the task of deciding whether two entity descriptions refer to the same real-world entity, as a paper published in the proceedings of the International Conference on Extending Database Technology in March 2025 defines it.
Its six benchmarks are e-commerce product listings and bibliographic records, never company, customer or supplier data, so what it shows here is evidence from a different domain, not a tested case for matching customer and supplier records. Its authors tested hosted and open-source large language models against two fine-tuned baselines built on pre-trained language models, a RoBERTa model and the Ditto matching system, and found that the best LLMs require no or only a few training examples to perform comparably to PLMs that were fine-tuned using thousands of examples and that LLM-based matchers further exhibit higher robustness to unseen entities, meaning entities absent from the model's fine-tuning data, a fact about training, not about which records a platform has looked at before.
The work runs on exports. The add-on's controller sends the customer master and the supplier master from its ERP and CRM, or the AR and AP subledgers where the add-on has neither, with trailing twelve-month revenue by customer and spend by supplier, and the platform's CFO sends the same four files from the platform. The matcher pairs records on name, address, tax ID and web domain, the language model decides the pairs an exact match cannot, and the add-on's controller confirms or rejects the pairs the matcher cannot decide alone, not every record, knowing which two names are one customer, the check that bounds a transfer the paper does not itself test.
The same matcher runs on the next add-on with the platform's records already in it, so each add-on adds to one customer and supplier list across the portfolio. The platform's CFO owns the day 60 report and pays for the matching from the integration budget, which covers the model's token charge and the controller's review hours; the add-on's controller supplies the exports in the first ten days and finishes reviewing pairs by day 40, on top of the first post-close close and its opening balance sheet already on that desk; and the deal team receives the report on day 60 and picks the first integration from it.
At a platform of four companies buying its fifth, the report to the deal team on day 60 carries four lines: shared customers, with the add-on's and the platform's revenue from them; shared suppliers, with each side's spend; the add-on's customers that no platform company sells to, which is the cross-sell list; and the pairs the matcher could not decide that the controller rejected, so the deal team knows how much of the list a person checked.
