AI in M&A Due Diligence: How to Build a Repeatable Workflow

Why One-off AI Prompting Does Not Create a Repeatable Diligence Process

One-off AI prompting helps with individual tasks. It does not create a repeatable diligence process. The AI starts each deal without knowing the team's chart of accounts, EBITDA definitions, or escalation rules. You upload a file, ask a question, and iterate until the output is usable. Most practitioners using AI in M&A follow this sequence. It’s fine for a single task, but it doesn’t work as a process.

That’s because every new deal starts from scratch, and the AI lacks critical knowledge. It doesn’t know how the team defines revenue, what payer concentration triggers a flag, what the IC deck needs to look like, or how compensation is reported in that industry. Re-explaining those rules for every deal isn't a workflow; it’s a waste of time.

Prompting AI for M&A use cases is a useful starting point. But diligence teams eventually need something more durable than a good prompt written on a good day.

The real gap is process memory. A repeatable workflow gives the AI that memory so the team stops re-explaining the same rules on every deal.

AI Does Not Fix a Poorly Scoped Diligence Process

AI can accelerate bad diligence as easily as good diligence. Speed on a generic review produces a faster generic review. Before any team builds an AI diligence workflow, it has to define what the workflow is supposed to catch.

The problem with most diligence processes is scope. Most deal teams apply the same template regardless of what they are buying or why: financial model, legal review, commercial diligence, IT assessment. The boxes get checked. But the diligence gaps that surface after close are almost always in the areas the template failed to address.

Jann Lau, Head of M&A at Thoughtworks, put it plainly: diligence needs to be anchored to the deal's core value drivers, not run as a parallel exercise alongside the thesis. The teams that get this right do not do more diligence. They scope it differently.

An AI workflow built on a poorly scoped process will surface faster findings but not better ones. The workflow needs to test the assumptions underlying the deal thesis: what must be true for this acquisition to create value, and what would break it? 

That’s what the rules in each skill should enforce.

What a Repeatable AI Diligence Workflow Actually Needs

A repeatable AI diligence workflow needs:

• defined input
• team-specific rules 
• source references
• assumption logs
• escalation logic
• a consistent output format
• hard gates
• a way to carry findings across workstreams

Most experienced deal teams carry that documentation in their heads. The workflow needs to make it explicit and executable.

One way to encode those rules is through Claude skills. A skill is a large structured prompt that tells the model exactly how a specific diligence job should run. It includes the team's chart of accounts and how to normalize against it, the EBITDA definition and how adjustments get made, the risk thresholds that trigger a flag versus those that pass, and the output format the IC team expects to receive.

"If you actually build a skill, it's like a repeatable process. You can think of it as a massive prompt that is doing exactly what you need, the same way every time. It knows exactly what your chart of accounts looks like and how to normalize them. It knows how you like to make EBITDA adjustments, what to look for. It knows what kind of questions you're gonna ask, and it spits out something that's ready to do modeling with the first time." — John Palusci

But a Claude skill isn’t magic. It’s process documentation written in a format the model can execute.

Every skill has two layers. Deterministic rules are the hard definitions: net income must tie, bank inflows reconcile this way, CPT codes map to these service line categories. These remove guesswork. The AI is not inventing its own definition of EBITDA. It’s simply following the team's rules.

Probabilistic reasoning is where judgment applies: flag this billing pattern for compliance review, surface payer concentration as a risk, and identify which findings require a decision versus which pass. The skill does not make those decisions. It gets them in front of the right person faster.

That division is what makes the workflow trustworthy. The rules layer removes ambiguity. The reasoning layer surfaces judgment calls for the team to resolve.

The Six Diligence Jobs to Turn Into an AI Workflow

A repeatable AI diligence workflow should not start with a single giant prompt. It should break diligence into the jobs the team already repeats on every deal.

In Palusci's example, that meant six jobs: financials, operational data, payroll, bank statements, pro forma modeling, and investment committee output. The exact categories will change by industry, but the principle holds: build around recurring work, not around the AI tool.

• Financials — map, remap, flag what does not tie
• Operational data — EMR, customer cohorts, service lines, usage data
• People and payroll — roster vs. P&L, key dependencies, cost structure
• Bank statements — cash proof, deposit reconciliation, outflow flags
• Pro forma assumptions — findings into model inputs, challengeable
• IC output — first-pass summary with sources, open questions, risks

Financials. The financial skill maps seller financials to the buyer's chart of accounts, rebuilds LTM views, documents every mapping assumption, and flags anything that does not tie. The output is a reconciled starting point. A mapping log shows exactly how each line was translated. That matters for financial diligence and for the negotiation that follows: findings surface earlier, when there is still time to act on them.

Operational data. In healthcare, this is EMR data: procedure codes, payer mix, provider volumes, collection trends, billing patterns, and compliance flags. In another industry, the same workflow logic applies to customer cohorts, SKU-level margin, usage data, contracts, service line performance, or project-level revenue. The skill encodes the categories and thresholds relevant to the deal. It also flags billing structures that pose compliance risk, unusual payer concentration, and revenue trends that do not align with the narrative.

People and payroll. The payroll skill checks the roster against the P&L, analyzes compensation structures, and surfaces mismatches: a provider generating revenue who does not appear in payroll; compensation that does not align with reported labor costs; and a staffing structure that would materially change post-acquisition.

Bank statements. The bank statement skill runs a light cash proof. It checks whether reported revenue shows up as deposits, flags distributions and unusual outflows, and does not smooth gaps. It documents what it found and where.

Pro forma modeling. The pro forma skill turns diligence outputs into assumptions and builds a working Excel model with linked cells and documented assumptions. Not a paste of values. A model the team can stress-test, with every assumption visible and open to challenge.

IC output. The IC skill drafts a first-pass investment committee summary with findings, source references, open questions, and risks. A faster starting point, not a final answer.

The orchestrator. The orchestrator is what turns six separate AI tasks into one diligence workflow. It sequences the six skills, enforces hard gates, and carries a running state file across the whole run. If net income does not tie, it flags the issue rather than passing the output forward. If the financials skill surfaces an anomalous charge, the bank statement skill knows to look for it. Findings accumulate instead of getting siloed by workstream.

Where the Workflow Was Tested

Schweiger Dermatology runs 170+ outpatient locations across nine states with more than 600 providers. The corp dev team closes 5-8 deals a year with 10+ active in pipeline at any given time. Most of those deals are smaller acquisitions where the finance team does the diligence in-house.

John Palusci, VP Strategic Finance, isn’t a developer. He built the workflow over a month of nights and weekends using Claude's skill-builder, which walks through the build interactively: clarifying questions about definitions, edge cases, and output requirements, then constructs the skill from the answers.

Healthcare adds EMR complexity that most general diligence processes aren’ built for. A general AI chat session can’t review electronic medical record data in a repeatable diligence process, so the value comes from encoding the team's categories, thresholds, and escalation rules before the file arrives.

What a Structured Workflow Can Surface Before the Analyst Starts

At M&A Science's AI in M&A Summit, Palusci ran the workflow on an allergy and dermatology practice, using real data with identifying information removed. He uploaded the EMR file, tagged the orchestration skill, and let it run.

Eight minutes later, the output included a twelve-month trailing collections trend with a year-over-year decline flagged; two primary providers identified, one flagged for a billing structure that carries compliance risk in outpatient medicine; and payer concentration flagged because one payer represented a disproportionate share of revenue. The contract transfer needed to be confirmed at close, and six open issues were logged with data references.

A skilled analyst would have taken a full day to build the same picture. And the output wasn’t just a confident summary; it was a documented analysis with assumptions logged, mapping tables showing how data was categorized, and Excel workbooks formatted the way the IC team expects.

Every flag had a source. Every assumption was logged. The skill forced every output to show its work. The objection to AI in M&A due diligence is rarely about speed. It is about whether the team can trust what it surfaces. Everything ties or it flags.

The workflow did not remove the need for diligence judgment. It moved the team faster to the point where judgment could begin.

What Teams Should Expect When Building an AI Diligence Workflow

This is not a no-work shortcut. It’s about making process work reusable.

Start by building one skill at a time. Pick a recurring diligence task, test the skill on anonymized data, add rules for edge cases, and confirm the output is traceable. Do that before building the next one. The orchestrator comes after the individual skills work independently.

In Palusci's case, the full build took about a month of nights and weekends. Token costs during development came to a few hundred dollars. Once built, the system runs on Sonnet for production deals; Opus was used during development where deeper reasoning was needed. No coding was required, but IT and legal approval still came first.

"Will it get you 100% of the way there? No. But AI gets you 80% of the way there really quickly — so that last 20% is the real value add, and you're able to get back to your seller that much quicker." — John Palusci

Security Checklist Before Loading Deal Data Into Claude

Before any team loads deal data into an AI platform, the setup needs to run through IT and legal review, not just a quick signup. In Palusci's case, the workflow ran on an enterprise Claude deployment cleared before any deal data was uploaded.

"Always work with your IT folks on this," Palusci said. "They're probably already worrying about it and have an opinion."

Security is not a reason to avoid the workflow. It is a reason to route the setup correctly before the first file is uploaded.

How to Build Your First AI Diligence Skill

Start with one recurring diligence task.

Pick something with repeatable inputs and repeatable outputs like financial mapping, bank statement reconciliation, or payroll comparison against P&L. Write the rules: what must tie, what gets flagged, what format the output takes, what sources get logged. Test it on anonymized data, and fix what breaks. Build the next skill only after the first one works consistently.

If the team is using Claude, that first task becomes the first skill. But the logic applies regardless of which AI platform is in use: any AI becomes more reliable in diligence when it is given explicit rules rather than asked to figure them out from context on each deal.

The first skill does not need to run the whole deal. It just needs to remove one recurring translation step.

The Translation Tax Is Optional

A deal team should not spend its first week translating files before it can start asking deal questions. AI will not decide whether the deal is worth doing. But it can handle the repeatable first pass, preserve the assumptions, flag what does not tie, and get the team to the judgment work faster. That is where AI in M&A due diligence starts to matter.

Frequently Asked Questions

How do M&A Teams Use AI for Due Diligence?

The most effective use is building a repeatable workflow, not asking one-off questions. Teams encode their chart of accounts, EBITDA definitions, risk thresholds, and output format into reusable skill prompts. The AI then runs the same process on every deal without re-explanation.

What is an AI Diligence Workflow for M&A?

An AI diligence workflow is a set of structured, reusable prompts that runs specific diligence jobs in sequence: financial mapping, operational data review, payroll verification, bank statement reconciliation, pro forma modeling, IC summary. Each skill encodes the team's rules and output format. An orchestrator connects the skills, enforces hard gates, and carries findings across all workstreams.

Can You Use Claude for M&A Due Diligence?

Yes, with an enterprise deployment cleared by IT and legal. Teams working with deal data need paid enterprise plans, web search turned off, data scoped to restricted projects, and a business associate agreement in place for healthcare deals. John Palusci, VP Strategic Finance at Schweiger Dermatology, demonstrated a six-skill Claude workflow at M&A Science's AI in M&A Summit in June 2026, running it on anonymized deal data in eight minutes.

Why Does Slow Financial Diligence Affect Deal Negotiations?

When findings surface late because data processing took too long, there is no time left to act on them. You cannot renegotiate price on something you found in week three of a four-week process. Getting financial analysis done early is what creates room to negotiate on price, structure, and representations and warranties.

What is the Difference Between One-off AI Use and a Repeatable AI Diligence Workflow?

One-off AI use helps with individual tasks but starts fresh on every deal. The AI has no memory of the team's chart of accounts, risk thresholds, or output format. A repeatable workflow encodes those rules into reusable skills so the AI runs the same process consistently across deals. The difference is process memory.

‍