Turn a Private Equity Investment Thesis Into an AI Opportunity Map

A portfolio company can list dozens of AI ideas and still have no reason to fund most of them. The investment thesis gives you a stricter test: which piece of work changes a measure in the deal model, what must be true for that change to appear, and who owns the result? An AI opportunity map answers those questions before the company buys tools or launches pilots.

Start with the deal model

Open the investment case, value-creation plan, and latest management accounts. Mark the assumptions that need operating change: price realisation, sales conversion, labour cost, inventory, debtor days, service levels, or losses. Give each assumption a baseline, target, date, and executive owner. If a number has already changed since acquisition, refresh the case before sizing an AI project against it.

Meet the people who run the work behind each assumption. A commercial lead can explain why quotes lose margin; a finance lead can show how invoice disputes delay cash; a service manager can identify the repeat issues that harm renewals. Ask for actual cases and records. A model has no economic value merely because it makes a task look faster in a demo.

McKinsey's guidance on PE value-creation plans calls for a quantified hypothesis, a named owner, and frequent review of each initiative. An opportunity map applies that discipline to AI: it connects a workflow change to a financial or operating measure that management already tracks.

Give each idea one primary value lever

Put every candidate in one primary column. Add secondary effects as notes, then decide later whether the evidence supports them. This keeps the first estimate clear and helps the CFO prevent double counting.

• Revenue: More won orders, higher realised prices, or better retention. Measure volume and price separately, and check what happens to gross profit after delivery cost.
• Operating margin: Less paid rework, overtime, contractor spend, waste, or avoidable support cost. Time returned to salaried staff is capacity until a budgeted cost or funded growth requirement changes.
• Working capital: Faster collections, lower inventory, or fewer billing delays. Cash released from shorter debtor days is a balance-sheet effect; it isn't automatically EBITDA.
• Service quality: Faster correct resolution, fewer missed appointments, or better on-time delivery. Start with a measurable operating outcome, then test whether it affects retention or price rather than assigning it an invented revenue value.
• Risk: Fewer compliance breaches, payment errors, or unsafe actions. Record exposure and control effectiveness; treat uncertain avoided losses as a scenario, not booked savings.

Some ideas genuinely touch more than one lever. An assistant that resolves billing disputes may reduce handling cost and release cash by shortening debtor days. Keep the underlying cases in one shared record so finance can separate labour savings from cash timing and check that neither benefit already sits in another initiative.

Trace the workflow and its limits

For each idea, write a one-page chain: input, decision, action, completed outcome, and the line in the value-creation plan it may move. Name the system that holds the source record, the person who may approve a change, and the exception that sends the work back to a person. Then compare the current process with the proposed process on real cases. Count checking, correction, integration, monitoring, and support, not only model usage.

Mark prerequisites beside the value estimate. A pricing assistant may need clean customer terms and cost data. A collections assistant may need dispute reasons linked to invoices and a clear authority to contact customers. If the data or approval path is missing, the opportunity belongs behind a prerequisite task with its own cost and owner.

BCG's 2026 survey of 100 senior PE investors reports that unclear ROI and competing priorities hinder digital work, while ERP and CRM records matter to later AI use. Those survey results aren't a return forecast for a portfolio company. They support pricing the data and integration work alongside the use case rather than hiding it after approval.

Work through a distributor example

Consider an illustrative industrial distributor with £40 million in annual sales. Its deal thesis depends on better quote discipline and faster cash collection. Sales staff prepare 12,000 quotes a year, and finance handles invoice disputes through email. The sponsor and management team map two candidates: a quote-review assistant and a dispute-triage assistant. The numbers below show the method, not a forecast for another company.

For quote review, the team identifies 2,400 comparable quotes a year where discount exceptions need a manager's decision. Suppose a controlled test shows £75 more contribution per eligible quote on average, counting lost quotes as zero, with the same overall win rate and product mix. That is £180,000 of annual gross-profit potential. Subtract £30,000 of recurring software, review, and support cost for a £150,000 annual run-rate contribution estimate. A £20,000 one-time integration bill brings the first-year estimate to £130,000. Before the board accepts it, the board should inspect the test's sample, the actual quote mix, and whether managers can repeat the result without extra sales effort.

For disputes, assume the company collects cash two days sooner at the same annual sales rate. Two days of sales are about £219,000 (£40 million divided by 365, then multiplied by two). That is a potential one-time release of receivables if the improvement holds, not £219,000 of recurring earnings. Any lower financing cost belongs in a separate calculation using the company's actual borrowing and cash position. Faster dispute handling may also reduce labour time, but the team must measure paid hours and avoid claiming the same staff capacity under another programme.

If a quote assistant also makes order entry cleaner, dispute volume might fall. That makes the two projects related, not additive by default. Test the overlap in the customer and order records before combining their values in the deal model.

Rank by value and proof, not excitement

Give each candidate a range for its annual recurring contribution, one-time cash effect, implementation cost, and time to first result. Add an evidence grade: observed in live operations, measured in a limited test, or still an assumption. Then assess access to data, owner capacity, staff adoption, and the consequence of a wrong output. A modest project with clean records and a willing owner may earn funding before a larger forecast built on missing data.

Use a stop rule before the pilot begins. For example, the quote assistant cannot expand if it lowers win rate, breaches price authority, or fails to lift contribution after review cost. Give the business owner the right to pause it. The map should show dependencies too: fixing customer terms once may enable both quote review and collections, while two separate data-cleanup budgets would count that work twice.

Put the map into a board decision

Bring the board a short list, not a catalogue of tools. For each candidate show the deal-model line, baseline, outcome measure, value range, evidence grade, full cost, dependency, owner, and next decision date. Keep recurring EBITDA, one-time cash, and service or risk measures in separate columns. Show how each estimate changes if volume, quality, or adoption disappoints.

Fund the next test only when it can settle an important uncertainty. A board can then compare AI work with pricing, procurement, hiring, and every other use of management time and capital. Keep the opportunity map current after each pilot: add the actual operating result and remove ideas whose value no longer survives contact with the business.