Private equity has plenty of work that looks ideal for AI: long documents, inconsistent spreadsheets, repetitive research and recurring memos. It also has confidential data, judgment-heavy decisions and numbers that must survive a partner review.
The sensible starting point is the workflow. Give the tool a bounded job, a controlled set of inputs, an output schema and a verification test. Keep the investment decision with the team.
This guide maps eight practical workflows across the deal lifecycle. Each one has a detailed playbook with prompts, file structures and review methods.
Clear the data boundary first
Before choosing a prompt, answer four questions:
- What data may enter the tool?
- Where will the files be stored and processed?
- Which connectors or local folders can the tool access?
- Who reviews the output before it reaches a model, memo or external party?
Consumer access, business plans and enterprise deployments can have different controls. Firm policy and the specific engagement govern the answer. An approved tool does not make every file approved.
For deal work, remove personal data and unrelated documents from the working folder. Use a copy of source files. Record what the agent changed. Keep original evidence available to the reviewer.
Choose the harness around the job
Claude Cowork and ChatGPT Work suit knowledge-work tasks that span documents, spreadsheets and presentations in a controlled workspace. Claude Code and Codex suit local-folder and terminal work: scripts, CSV processing, workbook inspection and repeatable checks.
The product names will change faster than the workflow. Focus on five durable capabilities:
- access to the permitted source files
- structured output
- citations or evidence tags
- repeatable scripts and logs
- clear review points before any write or external action
Check the current product documentation and the firm’s approved configuration before using a connector or local folder.
1. Map the market
AI can expand search language, enrich a seed list and standardize evidence across hundreds of companies. The associate still defines the boundary and tests for missed companies.
The workflow starts with a written inclusion rule. Build a seed list from known companies and trade sources, then ask the harness to expand synonyms and routes to market. Every company row should carry a source URL, quote, date and confidence.
Run false-positive and false-negative tests before calling the map complete. Review companies just above the cutoff and hide a known-good set to test recall.
Detailed playbook: How to Use AI for Private Equity Market Mapping
2. Screen add-on targets
A market map becomes useful when the platform thesis turns into a scorecard. Separate evidence gathering from scoring. Define the anchors for service fit, customer fit, geography, size and acquisition practicality.
The harness can enrich 100 names and create a review queue. It should mark unknown fields, capture sources and explain each score from observed evidence. A person reviews entity matches, hard disqualifiers and rows near the cutoff.
Detailed playbook: AI for Add-On Acquisition Screening
For the strategic context, see the site’s bolt-on acquisitions guide.
3. Review the CIM
“Review this CIM” is too broad. Use a staged process:
- extract defined facts into a table
- reconcile internal inconsistencies
- reverse-engineer the forecast
- build the missing-evidence list
- draft management questions
Require page-level citations. Then audit a fixed sample against the source. The detailed workflow also covers the security gate and what should stay out of an unapproved environment.
Detailed playbook: AI for CIM Review: A Private Equity Due Diligence Workflow
4. Prepare for the management meeting
Combine the CIM findings, expert calls, model assumptions and open issues into a question tree. Each primary question should carry the evidence gap, follow-up branches and the document that would close the issue.
Stress-test the tree with simulated answers. Keep those simulations separate from facts. Limit the core agenda so the team has time to listen and follow up.
Detailed playbook: Using AI to Prepare for a Management Meeting
5. Build and check the LBO
The defensible uses are model setup, formula audit, scenario generation and commentary. One-prompt model creation hides too many decisions.
Give the agent a model-convention file, bounded ranges and approved source schedules. Build modules separately and save new versions. Check every hardcode, recalculate returns independently and open the finished workbook in Excel.
Detailed playbook: AI for LBO Modeling
6. Draft the IC memo
AI can assemble a first draft after the evidence has been structured. Build an evidence matrix first. Tag each fact by source, distinguish management statements from documented facts and draft one section at a time.
Run a sentence-level citation audit and a separate numeric reconciliation. Use the harness to challenge the case, then let the deal team decide which risks matter.
Detailed playbook: How to Draft an Investment Committee Memo With AI
7. Monitor the portfolio
Monthly packs need a common schema before automation. Extract reported values with source tags, compare them with budget and prior periods, and apply explicit alert rules.
Keep observed movement, management explanation and investor follow-up in separate fields. Recheck every severe alert and sample a company with no alerts to test for misses.
Detailed playbook: AI for Private Equity Portfolio Monitoring
For where monitoring sits in ownership, see The Life Cycle of a Private Equity Fund.
8. Maintain the evidence and workflow
The eighth workflow sits across the other seven: keep the process reusable. Save instruction files, schemas, source indexes, check scripts and decision logs. Version them when the deal thesis or reporting definition changes.
A good harness should make the next run easier without hiding the logic. If a new associate cannot see the inputs, rules and checks, the workflow has become another black box.
A prompt pattern that travels
Most PE tasks can start from the same structure:
Objective: [one bounded task]
Inputs: [named files and approved sources]
Definitions: [metric, thesis or model conventions]
Output: [table, memo section, review queue or workbook range]
Evidence rule: [URL, page, cell or evidence tag for every material claim]
Unknowns: use UNKNOWN; do not infer missing facts
Checks: [reconciliation, source sample, false-positive/negative test]
Write rule: do not overwrite originals; save a new version
Stop rule: list unresolved choices and stop rather than assuming
This is more reliable than a long role-playing preamble. It tells the agent where the job ends.
Build or buy
A general-purpose harness is often enough for early workflows when the team already has approved access and somebody can maintain the templates. A specialist product may make sense when the same process runs across many deals, needs licensed data, requires firm-wide permissions or must integrate with existing systems.
Test the workflow before buying the system. Take 20 real examples, define the expected answer and measure accuracy, review time and failure severity. A polished demo on one clean CIM says little about performance on a scanned appendix or a changing KPI definition.
Roll out in stages
Start with read-only work on low-risk material. A sensible ladder is:
- summarize and extract with citations
- create review queues and draft questions
- run checks on copies of files
- write bounded outputs into templates
- connect to broader systems only after access and logs are settled
At each stage, measure time saved after review. Gross generation speed does not matter if the team spends longer fixing the output.
Common failure modes
The same problems recur across the lifecycle:
- an ambiguous task produces confident filler
- missing evidence is treated as a negative fact
- management commentary is upgraded into proof
- a source link points to a homepage rather than the claim
- an agent writes into the only copy of a model
- simulated answers leak into the memo
- a clean summary hides a broken reconciliation
Design a test for the expensive error. In market mapping, test misses. In a CIM, audit citations. In an LBO, recalculate. In monitoring, inspect quiet companies.
What the investor still owns
The team owns the acquisition thesis, the quality of earnings, the operating case, leverage, valuation, key risks and the decision to invest. It also owns the conversation with management and the judgment about whether an explanation is credible.
AI can reduce the cost of getting evidence into a reviewable form. That gives the investor more time for the parts that do not compress well: asking the next question, seeing the commercial pattern and deciding what could break the deal.
Measure the workflow on real files
Before broad use, build a small evaluation set from completed work. Include clean and messy examples, scanned pages, conflicting definitions, missing data and a case that should make the agent stop. Remove or protect confidential material as required.
Score the results on factual accuracy, source accuracy, completeness, review time and severity of errors. A missed decimal in covenant headroom matters more than awkward prose. Weight the evaluation accordingly. Keep the set stable so a new prompt or product version can be compared with the prior run.
Track reviewer time as well as generation time. A workflow that saves 40 minutes of drafting and adds an hour of citation repair is moving work around.
Keep the operating instructions visible
Store the thesis, metric definitions, model conventions and output schema beside the task. Avoid relying on a long chat history that nobody else can inspect. Date-stamp tool-specific instructions and keep the underlying control independent of the product.
Name an owner for each workflow. That person reviews failures, updates the evaluation set and decides when a change is ready for the team. Without ownership, prompts fork across associates and the process drifts.
Use review effort where the loss sits
Different tasks need different checks. A market map needs recall testing. A CIM extract needs source sampling. A model needs recalculation. An IC memo needs sentence-level evidence and numerical reconciliation. Portfolio monitoring needs rule tests and a review of quiet companies.
Do not apply the same generic “human in the loop” instruction everywhere. Name the person, the evidence and the test. The closer the output sits to a decision or an external action, the tighter that gate should be.
Keep external actions separate
Research, drafting and checking can happen inside the workflow. Outreach, data-room uploads, model overwrites and distribution to an IC are separate actions. Give them their own review and permission step. A tool that can reach email or shared drives should not do so merely because it prepared the file.
This separation also improves error recovery. The team can discard a bad draft or rerun an extract without recalling a message or unwinding a changed workbook. Treat read, draft, write and send as different levels of authority.
In practice
A useful AI workflow in private equity has a narrow task, approved inputs, a visible evidence trail and a human check matched to the failure risk. Start there. The choice of tool comes after.
The measure is straightforward: the work should be faster to review and easier to defend.
Tool references
Claude Cowork: https://www.anthropic.com/product/claude-cowork
Claude Code: https://www.anthropic.com/claude-code
ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex
OpenAI Codex: https://openai.com/codex/