Reading a portfolio company on day one: AI, the 100-day plan, and honest attribution

Start with the commercial problem. Why does the period right after a close matter so much, and where was the old approach falling short?

The first year of ownership is where a disproportionate share of value gets made or lost. Industry work has long put the first twelve months at roughly a third of total value creation, so the early planning quality matters out of proportion to its length. Yet the traditional 100-day plan spends much of that window on discovery: a small team, often with consultants, manually assembling pricing files, customer data, and operating metrics from systems that rarely talk to each other. Lower-middle-market companies are usually behind on data infrastructure, with fragmented systems, spreadsheets, and inconsistent definitions of what counts as a customer or a product. By the time you have a clean picture, weeks are gone and you are acting on a static snapshot. The commercial cost is concrete. Every week spent assembling data is a week not spent testing a price change or addressing churn, and the levers you find late are the ones that compound slowest. We wanted to compress discovery so that execution could start earlier, on better evidence.

Firms have tried analytics on portfolio companies for years. What specifically makes this viable now rather than five years ago?

Two things changed. The first is that models can now read messy, inconsistent, unstructured data and make sense of it. A great deal of a lower-middle-market company's reality lives in formats that defeated older tools: PDF price lists, free-text invoice lines, contract terms buried in documents, a CRM where the same account appears three different ways. Extraction and retrieval techniques let us pull that into a consistent structure without a six-month data-warehouse project first. The second is cost and speed. What used to require a dedicated data-science hire on each company, we can now do with a small central team and a repeatable process. That matters at our end of the market, where a company cannot support a large analytics function and the deal economics do not justify one. I would add a caveat. The technology did not remove the need for judgment. It removed a lot of the manual assembly that used to sit in front of the judgment.

Walk through how it actually works, concretely, from the day you take control.

On close we get read access to the core systems: the general ledger, the billing or point-of-sale data, the CRM, and whatever pricing files exist. The first pass is structural. We use extraction to normalize transactions, customers, and products into one consistent model, so that a customer means the same thing in the ledger and in the CRM. Then we run a standard set of analyses we have refined across deals. On pricing, we look at realized price by customer and product against list, discount leakage, and cases where similar customers pay very different amounts. On churn, we build a retention view and flag cohorts leaving faster than expected, then look for correlates: a price increase, a service problem, a departed sales rep. On operations, we surface where manual effort and exceptions cluster. The output is deliberately not a recommendation. It is a ranked set of hypotheses with the supporting evidence attached, which the operating team and management then pressure-test before anything reaches the plan.

This is sensitive material: customer lists, pricing, financials. How do you handle governance and confidentiality?

We treat portfolio company data as some of the most sensitive we hold, and it carries obligations that are both legal and fiduciary. A few rules are not negotiable. The data stays inside an environment we control, with access scoped to the specific deal team, so that someone working on one company cannot see another's information. We do not train shared or public models on a company's data, and we are explicit with management about that, because the fear that their numbers leak into someone else's tool is legitimate. We keep an audit trail of what was accessed and by whom. Where a company has its own confidentiality commitments to its customers, we honor them and keep personal information out of anything that does not need it. None of this is unique to AI; it is ordinary care for confidential deal information. What AI changes is the volume of data moving through the process, which raises the bar on getting access and retention controls right before the analysis starts, not after.

How do you keep the analysis from being confidently wrong, and who is accountable for a decision that follows from it?

The discipline is that the model produces hypotheses, not decisions. Every finding carries its underlying data, and the operating team traces it back before it goes anywhere. If the system says a customer segment is underpriced, we check the contracts and ask the commercial team why, because there is often a reason the model cannot see, such as a volume commitment or a strategic account. We have caught the tool over-reading a data artifact more than once, which is exactly why the check exists. Accountability sits with people, not software. The operating partner on the deal owns the value-creation plan and answers to the investment committee and the board for it; management owns execution. The AI is a faster analyst, and like any junior analyst its work is reviewed by someone senior who is accountable for the call. We are also candid that early findings are a starting read. We revise them as we learn the business, and we say so, rather than presenting a first pass as settled fact.

What about the human side: a management team that just got acquired, staff skepticism? What went wrong first?

The first mistake we made was arriving with answers instead of questions. A management team that has just been bought is watchful, and if the new owner shows up in week two with an AI-generated list of everything wrong with their pricing, you get defensiveness rather than partnership. The findings can be right and still go nowhere. So we changed the sequence. Now we present what the analysis surfaced as questions and ask the people who run the business to explain it. Often they already know the problem and simply lacked the time or the mandate to fix it, and the analysis gives them both cover and a prioritization. Who delivers it matters too. This works when the operating partner sits with the CFO and walks through the evidence together, not when a black-box report lands in an inbox. On the staff side, the worry is usually headcount. We are direct that this reallocates analytical effort toward advisory and execution rather than replacing the finance team, and we try to make that true in practice rather than merely say it.

What can you actually measure so far, and what should a reader be careful not to over-read?

The clearest effect is speed. Discovery that used to consume much of the 100 days now produces a first, evidence-backed view of pricing, churn, and operational drag within days, so the plan starts earlier and rests on data rather than anecdote. Because we run the same process on each company, it compounds: the second and third deals are faster than the first, and we are building a genuinely portfolio-wide playbook instead of reinventing the analysis each time. On outcomes I would be careful. Where a pricing or retention lever has actually been executed, we have seen margin improvements in the low hundreds of basis points on individual companies, roughly in line with what others in the market report, but I would not present that as a clean AI attribution. A pricing gain comes from the analysis, plus a management team that executed it, plus a market that allowed it. We track results in a value bridge and try to attribute them accurately. The dependable claim is that we find levers earlier and act on them sooner, not that software creates the value by itself.

What would you say to a skeptical peer, or to a management team nervous about this?

To a peer, I would say the failure modes are well documented and mostly are not about the technology. Most of these efforts stall on messy data, on deployments never tied to a specific value milestone, and on treating the work as an IT project rather than a business change. If you cannot connect the analysis to a line on the income statement and a person who owns it, you will produce interesting charts and no return. I would also warn against believing the output too easily; a confident, polished answer built on a data error is more dangerous than an obvious mistake. To a nervous management team, I would be plain about the limits. The tool does not understand their business better than they do. It reads patterns quickly and misses context routinely, which is precisely why their judgment stays in the loop. It is not a verdict on their competence. It is a way to spend the first months on the two or three things that matter most, instead of on assembling spreadsheets.

Give me a concrete moment where this mattered on a specific deal.

On one distribution business, the story going in was volume. The prior owner believed growth would come from winning larger accounts. In the first week, the normalized pricing analysis showed something different. A large group of mid-size customers was quietly receiving discounts that had drifted well below policy over years of renewals, with no volume to justify them. The pattern was hard to see in the raw system because the same customer was recorded inconsistently, which is exactly the mess the extraction step resolves. We took it to the CFO as a question rather than a finding, and it turned out the sales team had suspected it for a while but never had it quantified. That reframed the 100-day plan. Instead of chasing new logos first, the early work became disciplined price realization on the existing base, which is faster and lower-risk. To be clear, the people executed the change, not the model. But we would not have prioritized it in week one without the analysis, and prioritizing correctly in week one is most of the value.

Looking ahead, what does this mean for how your firm competes and how the profession works?

I think it changes what an operating partner is expected to do. The scarce skill used to be pattern recognition built over many deals, knowing where to look. That intuition still matters, but the tooling makes it faster to test and less dependent on any one person carrying it in their head. For a firm our size, that is leveling. We can bring a discipline to the lower middle market that used to require the analytical bench of a much larger firm. Over the next few years I expect diligence and the post-close plan to connect more directly, so that the value-creation thesis you underwrite is the same model you execute against, rather than two disconnected exercises. I am cautious about the enthusiasm. More than half of mid-market portfolio companies now report AI initiatives, so having one is no longer a differentiator; proving it moved a number is. The firms that do well will be the ones that stay disciplined about attribution and keep judgment where it belongs. The technology is an accelerant, not a substitute for knowing what to do with it.