On this page
- Key Takeaways
- The AI Line in the Thesis Needs Its Own Workstream
- IT Diligence Stops at Whether the Systems Run
- Two Kinds of Targets Need Two Different Checks
- Data Debt Is a Purchase Price Problem
- The Four Layers Under Every AI Assumption
- Run the Readiness Read Before the IC Vote
- Diligence Is Where AI Upside Becomes Real or Doesn't
- FAQ
Key Takeaways
- AI is now in the investment thesis. In one recent survey, 80% of sponsors said they build AI into the thesis as a driver of value. The diligence that tests that assumption has not caught up.
- Standard IT diligence answers a different question. It checks whether the systems run. It rarely checks whether the data inside them can support the AI use cases in the value creation plan.
- Two kinds of targets need two kinds of checks. If the target sells AI, test whether the AI is real. If AI is upside in your plan, test whether the data can carry it.
- Data debt is a price problem. Every gap you find before signing is a cost you can model, negotiate or plan for. Every gap you find after close comes out of the hold period.
- Run the readiness read before the IC vote. Scope it to the use cases in the thesis, keep it short, and turn every finding into a line in the model.
Somewhere in almost every value creation plan right now, there is a line about AI. Automate the back office. Cut cost-to-serve. Speed up quoting, claims, collections, whatever the business runs on. The line is usually one sentence, and it is usually carrying more of the return than anyone has checked.
Alvarez & Marsal's 2026 European Due Diligence Report put a number on it. In a survey of more than 80 senior private equity professionals across Europe, 80% of sponsors said they actively build AI into their investment thesis as a driver of operational and commercial value. The same survey found that only 18% say their diligence providers offer a fully integrated diligence offering. (It is a European sample, and a small one. I don't expect the picture to look very different on this side of the Atlantic.)
So most sponsors underwrite AI upside. A much smaller share get diligence that connects that story to the operational reality underneath it.
That gap is the whole argument of this post. If AI shows up in your thesis, then whether the target's data, systems and people can support that AI is a diligence question. It deserves the same rigor as a quality of earnings report. Without it, the diligence is incomplete in exactly the place where you are counting on upside.
The AI Line in the Thesis Needs Its Own Workstream
Think about how the other assumptions in a deal get tested. Revenue gets a quality of earnings review. Market size gets a commercial diligence team. Legal exposure gets counsel. Each big assumption has an owner and a workstream.
The AI assumption usually has neither. It gets written by the deal team, maybe sanity-checked by an operating partner, and then it rides into the model as a margin improvement in year two.
I understand why. AI diligence is new, the questions are unfamiliar, and the deal timeline is already tight. But the assumption is doing real work in the model. If you are paying for a 300 basis point margin story and part of it depends on automating a workflow, someone should confirm the workflow can be automated with the data the company actually has.
Ask who signed off on the AI assumption
Before the IC meeting, find the AI line in the value creation plan and ask one question: who tested this, and against what? If the answer is "the management presentation," you have a gap. Management teams are optimistic about their own data. Almost all of them are.
IT Diligence Stops at Whether the Systems Run
Most deals already include some form of IT or technology diligence. That work is useful. It tells you what systems the company runs, whether they are supported, what the security posture looks like, and what it will cost to keep the lights on.
What it usually does not tell you is whether those systems hold data an AI system can use.
Those are different questions. A company can run a stable, well-maintained ERP and still have customer records split across three systems with no common ID. It can have a clean CRM and still keep the real pricing logic in a sales manager's spreadsheet. It can pass a security review and still have no policy for what data is allowed to go into a model.
Gartner's research on AI-ready data found that 63% of organizations either do not have or are unsure if they have the right data management practices for AI. Gartner also predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data. That survey covers data management leaders broadly, not PE targets specifically. But I don't see a reason to think the average mid-market portfolio company is ahead of the average enterprise here. If anything, it is behind.
Read the IT diligence report with an AI question in mind
You don't need to throw out your existing tech diligence. Add a lens to it. For every system it covers, ask: does this system hold data that one of our AI use cases needs, and can we get that data out in a usable form? A lot of the answers are already in the report. Nobody has asked them in that order.
Two Kinds of Targets Need Two Different Checks
This is where I'd slow down, because "AI diligence" means two different things depending on the deal.
The target sells AI. The company's product or pitch depends on AI. Here the question is whether the AI is real, whether it works, and whether it will scale. You are testing the asset itself.
AI is upside in your plan. The company is a distributor, a services firm, a manufacturer. It doesn't sell AI. You plan to use AI to improve margins after close. Here the question is whether the company's data, systems and people can support the use cases in your value creation plan. You are testing the plan.
PressW has done the first kind. When Ecliptic Capital was evaluating a healthcare AI company, the investment team knew the industry well but needed a technical read on the target's AI. PressW built a question set from our 5 Pillars of AI Competency (data quality, data governance, modeling sophistication, testing and evaluation, and scalability), interviewed the target's technical team directly, and delivered a report with the positive and negative technical flags that fed into Ecliptic's final investment decision.
The second kind is more common, and it is the one most sponsors skip. It is also the one an AI & Data Readiness Assessment is built for.
Match the check to the deal
Before you scope anything, decide which kind of target you have. Sometimes it is both: a software company that sells an AI feature and also has back-office upside. Then you need both checks, and they should be run by people who can tell a demo from a production system.
Data Debt Is a Purchase Price Problem
Here is the way I think about it. Every gap between the AI plan and the data reality is a cost. The only question is who pays it and when.
If you find the gap before signing, you have options. You can price it in. You can push the timeline for the AI upside out a year in the model. You can budget the remediation into the 100-day plan, so it isn't a surprise. You can decide the upside isn't there and walk away from that part of the thesis.
If you find the gap after close, you have one option. You pay for it, out of the hold period, while the board asks why the margin story is late.
Data remediation is slow and unglamorous. Consolidating customer records, documenting where the real business logic lives, building pipelines out of legacy systems, setting access rules. None of it is hard in the abstract. All of it takes longer than a deal model assumes, because the people who know where the data lives are also the people running the business.
Turn every finding into a cost line
A readiness finding that says "data quality: moderate" is useless to an IC. A finding that says "customer data sits in two systems with no shared key; consolidating it is a prerequisite for the collections use case and pushes that benefit out by two quarters" is something you can put in a model. Ask for findings in that form.
The Four Layers Under Every AI Assumption
When you strip the AI line in a thesis down to what has to be true, it rests on four layers. If any one of them fails, the use case stalls.
| Layer | Diligence question | Red flag | What it changes in the deal |
|---|---|---|---|
| Data | Does the data each use case needs exist, and is it complete, current and accessible? | Key data lives in spreadsheets, email or a legacy system nobody can export from | Remediation cost and a later start date for the AI benefit |
| Systems | Can AI connect to the core systems through APIs or exports today? | Critical systems with no API, end-of-life platforms, heavy customization | Integration budget, or a system replacement in the 100-day plan |
| People | Who would own and adopt the new workflow, and do they have the time and skills? | Lean teams with no one to own adoption, key knowledge held by one or two people | Hiring, training or change management cost, plus key-person risk |
| Controls | Do data sensitivity, contracts and regulation allow the data to be used this way? | Customer contracts that restrict data use, regulated data with no governance | Legal review, governance build, or use cases that come off the table |
Each layer maps to work that an assessment can do quickly: profiling the key data sources, reviewing integrations, interviewing the people who do the work, and checking access and compliance. That is the same ground our AI & Data Readiness Assessment covers, with each use case scored on value, feasibility and risk.
Score the use cases in the thesis, not every possible use case
In diligence, you don't need a full AI strategy. You need to know whether the specific use cases in your value creation plan hold up. Scope the assessment to those, and leave the broader roadmap for after close.
Run the Readiness Read Before the IC Vote
Timing is where I'd push hardest. Too often the readiness assessment gets scheduled as a first-100-days activity, after close, when its findings can no longer change the price.
I'd move it. A focused readiness read on three to five use cases fits inside a normal diligence window if you scope it tightly and get access to the right people early. The output doesn't need to be a full roadmap. It needs to answer one question per use case: can this company's data, systems, people and controls support this, and what will it cost to get there?
There's a real limit here, and I want to be honest about it. Diligence access is constrained. You often can't profile production data or interview every team before signing. Some of the read will be based on interviews and documentation rather than hands-on testing. That is still far better than taking the management presentation at face value, and it gives you a short list of what to verify first after close.
Plan the post-close assessment before you sign
Whatever the pre-signing read can't cover, write it into the 100-day plan as a named workstream with an owner and a date. The pre-signing read tells you whether the thesis holds. The post-close assessment turns it into a plan your team can execute.
Diligence Is Where AI Upside Becomes Real or Doesn't
I'm bullish on AI inside portfolio companies. The margin opportunities in back-office and operational work are real, and they are getting bigger as the tools improve.
But I don't think we have a settled playbook for AI diligence yet. How much access should a sponsor push for? How should an IC weigh a readiness score against a strong commercial story? How do you price a two-quarter delay on an AI benefit? Most firms, including the ones we work with, are still working those questions out deal by deal. What I'm confident about is that the firms who test the AI assumption before they sign will make better calls than the firms who discover it after.
If AI is in your next thesis, start with one question: which use cases in the plan depend on data you haven't seen yet? If you want help answering that before your next IC, that is what our AI & Data Readiness Assessment is for. Book a Discovery Call.
FAQ
An AI & Data Readiness Assessment in due diligence tests whether a target company's data, systems, people and controls can support the AI use cases in the investment thesis. It looks at data quality and access, integrations, adoption capacity and governance. The output is a view of which use cases hold up and what it will cost to deliver them.
IT due diligence checks whether a company's systems run, are supported and are secure. AI due diligence asks whether the data inside those systems can be used by AI for specific use cases. The two overlap, and a good AI readiness read builds on the IT findings instead of repeating them.
Before the investment committee vote, if AI is part of the value creation plan. A focused pre-signing read can inform price and timing. A fuller assessment after close turns those findings into a roadmap for the first 100 days.
Yes, though with limits. Much of a pre-signing read comes from interviews with the target's team, system documentation and sample data. Gaps that can't be tested before signing should become named workstreams in the 100-day plan.
Common red flags include key data held in spreadsheets or email, core systems with no API, no owner for adoption, knowledge concentrated in one or two people, and customer contracts that restrict how data can be used. Each one adds cost or delay to the AI upside in the plan.
Readiness findings turn the AI line in a thesis into costs and timelines. A data gap might add remediation cost or push an AI benefit out by a few quarters in the model. In some cases, a finding takes a use case off the table, and the plan changes before you sign instead of after.
