--- title: "AI value creation due diligence: what the IC needs" description: "AI value creation due diligence should hand the IC a costed plan, not a risk appendix. What the AI section of the paper holds and how it hands over." url: https://anywayventures.com/insights/ai-due-diligence-value-creation-plan updated: 2026-09-25 --- [← ALL WRITING](https://anywayventures.com/insights) DILIGENCE # AI due diligence is a value-creation plan, not a risk register Most funds now ask 'what's the AI risk?' in diligence. The better question is what AI makes this business worth, how fast, and what the IC paper should say about it. WRITTENSeptember 2026 READING · Ten minutes FOR · Deal teams, investment committees RELATES TO[AI due diligence](https://anywayventures.com/ai-due-diligence) Technical diligence has a settled shape: architecture, code quality, security, team, a red, amber and green summary. It's useful, and it's almost entirely defensive. AI has been bolted onto that list as another risk category: is the target's AI real, is the data clean, is there IP exposure. My view is that AI value creation due diligence should produce something different, a costed plan that sits in the investment case, and that the AI section of the IC paper should be written so someone can deliver it on day one. For most mid-market targets the AI question isn't "what could go wrong". It's how much cost comes out of this operation, how much faster it grows, and how quickly a new owner can make that happen. That's a value-creation question. It belongs in the investment case, not the risk appendix. > The plan you can underwrite is the one that knows what broke last month. I've covered what AI due diligence has to test in [AI due diligence for private equity](https://anywayventures.com/insights/ai-due-diligence-private-equity) , and how to build the plan properly after close in [how to build an AI value creation plan](https://anywayventures.com/insights/ai-value-creation-plan) . This piece is about the thing in between: the deliverable. What the IC actually reads, how the deal team uses it, and how it gets handed over so it's delivered rather than filed. ## Why the AI section of the IC paper usually disappoints Funds are already using AI on themselves. The [Bain and StepStone 2026 GP Outlook](https://www.bain.com/insights/private-equitys-reality-check-gp-outlook-2026/) found GPs report their highest returns from generative AI in deal sourcing and due diligence, while benefits inside portfolio companies skew to cost savings. The same survey found nearly 40% of GPs don't expect material financial impact from AI in their portfolio companies in 2026. Read those together. Funds believe AI works, because they've seen it work in their own diligence. They don't yet believe it will show up in portfolio EBITDA. That gap is mostly a paper problem. The AI section of a typical IC paper is two paragraphs of adjectives, a vendor's maturity score and a risk table. Nobody can underwrite adjectives, so the IC quite reasonably gives the AI case a value of zero. There's a second reason to take it seriously. The sell-off in listed software stocks in early 2026, driven by fear that AI agents undercut seat-based SaaS, didn't stay in public markets. [J.P. Morgan Asset Management](https://am.jpmorgan.com/us/en/asset-management/institutional/insights/market-insights/market-updates/on-the-minds-of-investors/what-does-the-software-sell-off-mean-for-private-markets/) pointed out the knock-on for private equity and private credit, where software is one of the largest sector exposures. ICs now ask about AI on every deal. The paper should give them an answer they can price. ## What goes in the paper Done properly, the AI section has seven parts. I'd hold any adviser, including me, to all of them. A one-page summary: the base case number, the three or four lines that make it up, the cost to get there and the single biggest precondition. The function-by-function walk: finance, operations, sales, customer service, product, ranked by hours and margin at stake, tested against what the target's data and systems can support today. The evidence behind each number: ticket exports, timesheets, headcount by role, system logs, sample reviews. Every figure traces to a source in the data room or a session with management. Conservative, base and stretch cases for the model, with the capture rate, loaded cost and run cost stated for each line. A sequenced first hundred days: what starts in week one, what waits for a data fix, who owns each line. Dependencies and preconditions: the systems, data, people and contracts each line needs before it can pay back. What would make us wrong: the two or three assumptions that, if false, take a line to zero. The last part is the one most papers leave out. It's also the one an experienced IC member reads first, because it tells them whether the author has thought about failure or only about the upside. The one-page summary deserves care. If a partner reads nothing else, they should come away knowing the base case in pounds, what it costs to deliver, when it starts to show in the numbers, and the one thing that must be true. Four sentences, a small table and nothing about "the art of the possible". ## How a deal team uses each part Each part of the paper feeds a different decision, and it helps to write it that way. The conservative, base and stretch cases are model inputs. The deal team picks which one goes into the base model, and in most lower-mid-market deals that should be the base case with a delayed start, or the conservative case in full. The stretch case goes in the upside scenario and nowhere else. Price follows from that choice. If the fund pays for the stretch case at entry, the vendor has taken the upside and the new owner has bought the work. The paper should say plainly which case it thinks the price can carry. The sequenced hundred days becomes a workstream in the 100-day plan, with the same owners and dates. If it can't be lifted straight across, it wasn't sequenced properly. The function walk and the owners feed management incentives. When a line has a named owner in the management team and a metric, it can sit in that person's objectives or the MIP hurdles. [Bain's 2025 Global Private Equity Report](https://www.bain.com/insights/field-notes-from-generative-ai-insurgency-global-private-equity-report-2025/) describes Vista Equity Partners asking portfolio companies to submit goals and quantified benefits for their GenAI initiatives, and running a GenAI CEO council. The same report found nearly 20% of portfolio companies had operationalised generative AI use cases and were seeing concrete results. Asking management for quantified benefits only works if the first set of numbers came from somewhere sensible. Diligence is where they come from. The preconditions and the "what would make us wrong" list feed the SPA and the first board pack. A precondition the vendor can fix before completion is a negotiation point. One that can't be fixed is a reason to push the start date back in the model. ## One line, from diligence finding to 100-day action Picture a £60m distribution business with 300 staff. Fourteen of them work in customer service. Here is how a single AI line travels through the paper. The numbers are illustrative. Stage What the paper says Arithmetic Diligence finding Ticket export shows 9,000 tickets a month at an average of 12 minutes each. A sample of 200 shows 55% are order status or delivery date queries answerable from the ERP and carrier data. 9,000 × 12 min = 1,800 hours a month. 1,800 × 55% = 990 hours in scope Capture rate Conservative 20%, base 35%, stretch 50%, because customers still phone and some carriers don't send tracking data 198, 347 and 495 hours a month returned Value Loaded cost of £30 an hour, converted through not backfilling leavers and absorbing volume growth without hiring £71k, £125k and £178k a year Cost £50k to build, £1.5k a month to run £18k a year run cost. Base case net run rate £125k less £18k = £107k Model input Base case, starting month 7, not month 1 Year one: 6 months × £107k ÷ 12 = about £54k, less £50k build = about £4k. Year two: £107k Precondition Only 2 of 5 carriers feed tracking data into the ERP today The other 3 carry 40% of volume. Connecting them by day 30 is assumed. If it slips, the line falls to the conservative case 100-day action Days 1 to 30: baseline and carrier feeds. Days 31 to 70: build with three service agents. Days 71 to 100: live across all connected carriers Owner: head of customer service. Metric: hours returned per month Notice what the table does. The £125k never goes into the model as a year-one figure, because nothing gets built in month one and the carriers aren't connected. Year one is roughly breakeven. That's an honest number, and it's the kind of number a CFO will sign up to. It also tells the incoming chair exactly what to ask about at the day-60 board meeting. The same line in a weak paper reads: "AI-enabled customer service could reduce contact centre costs by 30 to 40%." Nobody can check that, so nobody can hold anyone to it. ## Tests an investment committee member can apply Any AI value-creation claim in a paper should survive these. They take ten minutes and need no technical background. Can the paper name the process, the team that does it today and the hours it takes, from the target's own data rather than a benchmark? Is there a baseline, measured before anything is built? Does each capture rate come with a reason, and would you accept the conservative case if that's all you got? Are run costs in, per year, alongside the build cost? Does the paper say how hours become pounds: redeployment, not backfilling leavers, avoided hires, or faster billing? Has the data precondition been tested in the data room, or assumed? Is there a named owner in management who has seen the number and agrees with it? Does the paper say what would make it wrong? Does the start date in the model reflect the time it takes to fix the preconditions? Will the person who wrote this still be involved after completion? If a claim fails three or more of those, treat it as colour. Keep it in the equity story if you like. Keep it out of the price. ## Who should write it The uncomfortable part is that the plan can only be written well by someone who has recently built these systems in production. Vendor decks and analyst frameworks will tell you what is theoretically possible. The plan you can underwrite is the one that knows what broke last month. The difference shows up in the capture rates. Someone who hasn't shipped an AI system into a live operation recently tends to assume the model is the hard part. It rarely is. The hard parts are carrier feeds that drop out on Fridays, a CRM where half the fields are free text, the one person in finance who holds the process in their head, and a run cost that triples when usage goes up. Those are what move a line from base to conservative. You only price them in if you've paid for them yourself. It also shows up in sequencing. A generalist will put the most valuable line first. Someone who has built these things will put first the line whose data is ready, because early proof buys the management team's patience for the harder work. I spent thirteen years running Netsells, a York digital product consultancy, and was Chief Product Officer, COO and then CEO of YourParkingSpace. These days I build AI systems inside PE-backed and mid-market businesses every week. That's the reason I'm comfortable putting a capture rate in a paper and being held to it. ## Handing over at completion Most AI plans die in the gap between signing and the first board meeting. The diligence adviser writes the paper and leaves. The operating team inherits a PDF. Three months later a delivery partner is hired, reads the paper, disagrees with half of it, and writes their own. By then the IC has approved one thesis and management is working to a second. That's thesis drift, and it's expensive. Every handover between advisers loses the context behind the numbers: why a capture rate was set at 20% and not 35%, which carrier was the problem, which manager was sceptical and why. The paper records the conclusion. The reasoning lives in someone's head. Three things stop it. First, write the paper so it can be lifted directly into the 100-day plan, with the same lines, owners and metric. Second, hold a handover session in the first fortnight after completion where the author walks the management owners through each line and its preconditions, and the owners agree or amend the base case on the record. Third, keep the author involved for at least the first hundred days, whether as an operating partner, a delivery lead or simply the person the chair calls when a number looks off. The metric should carry across unchanged. I use hours returned to the business per month, and later, what is still running, unattended, six months after it went live. If the paper reports one metric and the board pack reports another, the thesis has already drifted. Continuity also keeps the author honest. It's easy to write a generous stretch case if you won't be around when it's measured. It's much harder when you'll be sitting in the board meeting where it isn't met. ## A paper someone can deliver A good AI section is short, specific and slightly boring. It names processes, cites evidence, prices the base case with a sensible start date, and says what would make it wrong. It leaves the IC with a number it can underwrite and leaves management with a plan it has already agreed to. If you have a deal in exclusivity and want the AI section written that way, that's what our [AI due diligence](https://anywayventures.com/ai-due-diligence) work produces, with a fixed quote in 48 hours. The [first call](https://anywayventures.com/contact) is free. ## Questions ### What should the AI section of an investment committee paper include? A one-page summary with the base case number, a function-by-function walk of where hours and margin sit, the evidence behind each figure, conservative, base and stretch cases for the model, a sequenced first hundred days, the preconditions each line depends on, and a short list of what would make the case wrong. It should read as a plan someone can deliver, not a list of risks. ### How is AI value creation due diligence different from AI risk diligence? Risk diligence asks whether the target's AI is real, whether the data is clean and whether there is IP or regulatory exposure. Value creation diligence asks how much cost comes out, how much faster the business can grow and how quickly a new owner can make that happen. Both matter. Only the second produces numbers the deal team can put in the model and hold management to. ### Should AI upside be priced into a private equity deal? Usually only the base case, and sometimes only the conservative one. The stretch case belongs in the equity story and the management incentive plan, not the price. If the fund pays for the stretch case at entry, the upside has gone to the vendor. A costed, evidenced base case gives the IC something to underwrite without paying twice for the same work. ### Who should write the AI value-creation plan in diligence? Someone who has recently built and run AI systems in production, ideally the same person or team who will help deliver it after completion. Recent build experience keeps the capture rates and costs honest. Continuity stops the thesis drifting between the diligence adviser, the operating team and whoever is hired to deliver it. Brannan Coady Founder of Anyway Ventures. Exited founder, former CEO of YourParkingSpace, investor and board member; now building AI inside PE-backed and mid-market businesses. [More →](https://anywayventures.com/brannan-coady) ## Keep reading [ALL WRITING →](https://anywayventures.com/insights) [DILIGENCE · 10 MIN AI due diligence for private equity: what to test before you sign Most AI diligence reports check whether the target's AI is real and stop there. That is one job of three, and it is rarely the one that moves the price.](https://anywayventures.com/insights/ai-due-diligence-private-equity) [OPERATING PARTNER · 10 MIN How to build an AI value creation plan for a portfolio company A working method, step by step: where the hours go, what to rank first, how to size it in pounds, and how to report it so the exit story is evidenced rather than asserted.](https://anywayventures.com/insights/ai-value-creation-plan) [DILIGENCE · 10 MIN AI disruption risk assessment: will AI hollow out the target? AI washing covers targets that overclaim AI. This covers the other problem: targets that AI threatens. An eight-factor grid, the evidence to request, and what the score should do to the deal.](https://anywayventures.com/insights/ai-disruption-risk-assessment)