AI Across PE Portfolios: Use-Case Underwriting Before Spend

The operating partner does not need another AI trend briefing.

What the operating partner needs is a repeatable way to decide which portfolio companies, which functions, and which use cases deserve AI funding — and which ones should wait, fix their foundation, or stop entirely.

Many PE-backed portfolios already have AI pressure from multiple directions: portfolio company leadership asking for AI budget, vendors offering demos, board members asking about AI strategy, and competitors making announcements. The volume of requests is growing. What is not growing is the rigor of the funding decisions.

The result: AI spend that is driven by enthusiasm, vendor relationships, or competitive anxiety rather than by evidence of readiness. Some initiatives get funded before the data infrastructure can support them. Others duplicate effort across portfolio companies that share the same functions but were never coordinated. And some should have been stopped months ago but persist because no one has the framework to recommend stopping.

The Funding Triage Problem

The most expensive PE AI mistake is not funding the wrong initiative. It is funding initiatives before assessing whether they are ready.

A portfolio company with fragmented data infrastructure, no labeled validation data, and no governance framework is not ready for an AI production initiative — regardless of how compelling the use case sounds in a board presentation.

Funding that initiative before fixing the foundation creates a predictable outcome: the team spends months building a system that works in demo but cannot survive production data quality, integration complexity, or organizational readiness.

The triage framework that avoids this outcome has four possible outputs:

• Fund: the initiative has a clear problem, adequate data, appropriate architecture complexity, governance infrastructure, and a credible production path
• Defer: the initiative has merit but is missing one or two readiness dimensions that can be resolved in a defined timeframe
• Fix the foundation: the initiative needs infrastructure work (data, governance, team capability) before any AI investment will produce value
• Stop: the initiative cannot articulate the problem, does not have the data, or has no credible path to production

The operating teams that do this well share one trait: they treat the stop recommendation as a capital return, not a failure. Stopping an initiative that was never ready is not a missed opportunity — it is the point of the framework. The operating teams that resist stopping are the ones that funded enthusiasm, and now have to explain why AI spend produced nothing in production. That explanation is more expensive than the stop would have been.

Readiness Varies More Than Operating Teams Expect

Portfolio companies that share the same functions — sales ops, customer support, finance, procurement — rarely share the same AI readiness.

Company A may have centralized CRM data, documented workflows, and an internal data team. Company B may run the same function on spreadsheets, email, and tribal knowledge. Both may request AI budget for “sales automation,” but their readiness profiles are fundamentally different.

The readiness scan should assess each company independently across six dimensions:

1. Problem clarity: can the company state what decision AI would improve, in one sentence?
2. Data readiness: does labeled, versioned, accessible data exist with a known owner?
3. Complexity fit: does the use case actually require AI, or would deterministic automation suffice?
4. Governance maturity: do approval gates, logging, and escalation paths exist?
5. Integration surface: how many systems must the AI initiative connect to, and are they documented?
6. Production trajectory: is there a credible path to production with stage gates and exit criteria?

These dimensions are scored per initiative, not per company. A single portfolio company may have one initiative that scores green and another that should stop.

Common Functions Worth Coordinating

Some functions recur across PE portfolios often enough that coordinated evaluation is more efficient than company-by-company assessment:

Sales Operations and CRM

Sales ops AI — lead scoring, pipeline forecasting, activity recommendation — is one of the most frequently requested and most frequently disappointing AI use cases in PE portfolios. The reason: CRM data quality. If the CRM is not maintained, the model is not useful.

The readiness gate: is CRM data complete, current, and reliably maintained? If not, the investment should be data quality, not AI.

Customer Support

Customer support — classification, routing, response suggestion — is a strong automation candidate when ticket data is structured and the escalation workflow is documented. It is a weak candidate when support operates through unstructured email, chat, and phone with no centralized system.

Finance and Reporting

Financial reporting automation — extraction, reconciliation, anomaly detection — has clear value and relatively low complexity. The data usually exists in structured accounting systems, and the accuracy is measurable.

Procurement

Procurement AI — spend analysis, supplier risk monitoring, contract extraction — is valuable at portfolio companies with significant vendor spend. The readiness gate is similar to sales: does the procurement data exist in a structured, accessible system?

Governance Standards Without Architectural Uniformity

The operating team should enforce standards, not platforms.

A standard vendor evaluation framework, minimum governance requirements, and readiness assessment criteria create comparability across the portfolio without forcing architectural uniformity that ignores company-specific constraints.

The minimum governance standard should include:

• Decision owner: every AI initiative has a named person accountable for outcomes
• Approval gates: defined by blast radius — internal analytics require less oversight than customer-facing automation
• Logging: all AI system actions logged with enough context to investigate unexpected behavior
• Rollback: a defined procedure for reverting to the pre-AI process
• Evaluation: measurable criteria for whether the AI system is improving the target workflow

In practice, the operating partner who arrives at a portfolio company with this scorecard in hand changes the conversation immediately. Instead of arguing about which vendor to pick, the conversation becomes: does this initiative have a named decision owner? Is there a rollback plan? What is the measurable accuracy target? Those questions either get answered or they reveal that the initiative was not ready. That shift — from vendor selection to readiness verification — is where the capital protection actually happens.

from pydantic import BaseModel, Field
from typing import Literal, Optional
from enum import Enum


class ReadinessVerdict(str, Enum):
    FUND = "fund"
    DEFER = "defer"
    FIX_FOUNDATION = "fix_foundation"
    STOP = "stop"


class GovernanceChecklist(BaseModel):
    has_named_decision_owner: bool
    approval_gates_defined: bool
    logging_requirements_documented: bool
    rollback_procedure_exists: bool
    evaluation_criteria_measurable: bool

    def governance_score(self) -> int:
        return sum([
            self.has_named_decision_owner,
            self.approval_gates_defined,
            self.logging_requirements_documented,
            self.rollback_procedure_exists,
            self.evaluation_criteria_measurable,
        ])


class UseCaseReadinessScorecard(BaseModel):
    initiative_name: str
    portfolio_company: str
    function: str  # e.g. "customer support", "sales ops", "finance"

    # Six readiness dimensions — example scale: 0 = blocker, 1 = partial, 2 = clear
    problem_clarity: int = Field(ge=0, le=2, description="Can the team state the target decision in one sentence?")
    data_readiness: int = Field(ge=0, le=2, description="Labeled, versioned, accessible data with a known owner")
    complexity_fit: int = Field(ge=0, le=2, description="Does the use case actually require AI vs deterministic automation?")
    governance_maturity: int = Field(ge=0, le=2, description="Approval gates, logging, escalation paths exist")
    integration_surface: int = Field(ge=0, le=2, description="Connected systems are documented and accessible")
    production_trajectory: int = Field(ge=0, le=2, description="Credible path to production with exit criteria")

    governance_checklist: GovernanceChecklist
    fund_score_threshold: int = Field(..., ge=0, le=12)
    defer_score_threshold: int = Field(..., ge=0, le=12)
    governance_score_threshold: int = Field(..., ge=0, le=5)
    blocking_gap: Optional[str] = Field(None, description="Primary gap preventing fund verdict, if any")
    recommended_action: Optional[str] = None

    def total_score(self) -> int:
        return (
            self.problem_clarity
            + self.data_readiness
            + self.complexity_fit
            + self.governance_maturity
            + self.integration_surface
            + self.production_trajectory
        )

    def verdict(self) -> ReadinessVerdict:
        score = self.total_score()
        has_blocker = self.problem_clarity == 0 or self.data_readiness == 0
        if has_blocker:
            return ReadinessVerdict.STOP
        if score >= self.fund_score_threshold and self.governance_checklist.governance_score() >= self.governance_score_threshold:
            return ReadinessVerdict.FUND
        if score >= self.defer_score_threshold:
            return ReadinessVerdict.DEFER
        return ReadinessVerdict.FIX_FOUNDATION

This model enforces the six-dimension readiness framework programmatically: each initiative becomes a scorecard instance, and the verdict distribution tells the operating team where capital is actually ready to deploy. If every instance returns fund, the scoring thresholds need recalibration.

pe_portfolio_readiness_scan:
  fund: "Zenith Corp"
  portfolio_company: "Zenith Corp — B2B SaaS"
  initiatives_assessed: initial_tranche
  results:
    - initiative: "customer support ticket classification"
      readiness: fund
      rationale: "Structured ticket data, documented escalation, measurable accuracy target"
      recommended_architecture: deterministic_classification_with_human_review
      timeline: "bounded path to production pilot"
    - initiative: "sales pipeline forecasting"
      readiness: defer
      rationale: "CRM data incomplete — material share of opportunities missing stage updates"
      blocker: "CRM data quality — requires cleanup before model training"
      recommended_action: "Fund CRM data quality initiative first"
    - initiative: "AI-powered product recommendations"
      readiness: stop
      rationale: "Cannot articulate target decision, no labeled data, no evaluation criteria"
      recommended_action: "Return to problem definition — revisit after customer support initiative proves deployment capability"
  portfolio_level_recommendations:
    - "Establish standard vendor evaluation criteria across all companies before next procurement cycle"
    - "Require decision owner and rollback procedure for every funded initiative"
    - "Share customer support classification playbook with other portfolio companies running similar functions"

The Board Memo

The operating team’s job is not to present AI enthusiasm to the board. It is to present AI judgment: which initiatives are funded, which are deferred, which are stopped, and what evidence supports each decision.

The board-ready artifact should include:

• A portfolio map showing each company’s AI initiatives and their readiness scores
• Fund/defer/stop recommendations with specific rationale for each
• Common functions where coordinated evaluation would reduce duplicated effort
• Governance gaps that create portfolio-level risk
• Capital allocation recommendation tied to readiness evidence, not vendor promises

For the six-dimension readiness scorecard applied to individual initiatives, see The 6 Dimensions To Score Before Recommending an AI Engagement. For the discovery questions that reveal whether any initiative is ready for advisory, see Enterprise AI Use-Case Intake System. For the governance review that produces operating decisions rather than awareness decks, see What an Enterprise AI Governance Review Should Produce in 30 Days. For what the portfolio review artifact should contain when delivered to the board, see What an Enterprise Agentic Portfolio Review Should Produce in 30 Days. For the procurement mistakes that let vendors set the architecture narrative before readiness is assessed, see The Procurement Problem: What Enterprise AI Buyers Get Wrong Before the First Vendor Call.

FAQ

The Decision Rule

Do not fund AI across a PE portfolio from vendor pressure, board anxiety, or platform standardization. Underwrite each use case like capital allocation: problem clarity, data readiness, governance, integration risk, and production path first; spend second.