A portfolio operator rarely sees one AI proposal at a time.
The requests arrive from several directions at once: a vendor demo, an internal prototype, a board question, a portfolio company leader asking for budget, a platform team proposing standardization. Most proposals sound plausible. Fewer arrive with the evidence needed to make a capital decision.
The problem is not a shortage of AI ideas. It is the absence of a triage process that separates fundable workflows from expensive experiments.
This post describes a four-decision model for reviewing AI initiatives across a portfolio. Each candidate gets classified as Fund, Defer, Redesign, or Kill based on workflow evidence, ownership, data readiness, and maintenance burden.
The Portfolio AI Triage Problem
Portfolio operators face a different AI intake problem than a single-company technology leader. A single company can sometimes absorb a few experiments that do not work. A portfolio review has to compare initiatives across different companies, systems, teams, and maturity levels.
That makes the real constraint operating attention. Every funded pilot that has no accountable owner becomes a maintenance liability. Every vendor proposal that skips data ownership becomes an integration surprise. Every board-driven AI request that lacks a measurable workflow competes with work that may already have the evidence to proceed.
The triage model below forces each candidate initiative through the same evidence gates before it receives a classification.
The Four Decisions: Fund, Defer, Redesign, Kill
Each candidate AI initiative receives one of four classifications. There is no “maybe” category.
Fund. The initiative has a named owner, documented workflow evidence, accessible source data, a defined permission boundary, and a review path that does not depend on the project team grading its own work.
Defer. The workflow is plausible but one or more evidence gates are not yet met. Defer is not a soft rejection. It is a conditional hold with specific unblock criteria.
Redesign. The underlying problem is real but the proposed approach carries unnecessary risk. Common triggers: single vendor with no exit path, undefined permission boundary, or assumed data quality. Redesign sends the initiative back for a revised proposal with specific gaps named.
Kill. The initiative has no accountable owner, no measurable workflow, or depends on data the company cannot govern. Killing protects the portfolio from orphan systems.
| If the candidate initiative has... | Then... |
|---|---|
| Named owner, workflow evidence, accessible data, defined permission boundary, independent review path | Fund. Assign capital and engineering time. Define the first review gate before work starts. |
| Plausible workflow but missing data access, owner, or review path | Defer. Name the specific gap. Do not fund until the unblock condition is visible. |
| Real problem but vendor lock-in, undefined permissions, or assumed data quality | Redesign. Return for revised proposal. Name the structural gaps that must be addressed. |
| No accountable owner, no measurable workflow, or ungovernable data dependency | Kill. Document the reason. Do not revisit without a materially different proposal. |
The Evidence Each Candidate Workflow Needs
Before any initiative reaches the triage table, it should arrive with evidence across eight dimensions. If the proposal lacks evidence on any dimension, that gap becomes part of the classification.
Accountable owner. A named person inside the operating company who will own the system after deployment. Not the vendor. Not only the operating partner who approved the budget. Someone whose real operating responsibilities include the workflow.
Workflow definition. A specific description of what the AI system will do, with measurable inputs and outputs. “Automate document review” is not enough. A better version names the document type, the review action, the escalation route, and the acceptance criteria.
Source systems. The systems where input data lives. Named platforms, named APIs, named data stores. If the proposal says “we will integrate with existing systems” without naming them, integration risk is undefined.
Data owner. The person or team that controls access to the source data and can confirm its quality, freshness, and completeness. Data ownership and system ownership are often different people.
Permission envelope. What the AI system is allowed to do without human approval, and what requires human review before action. A system that classifies documents needs different permissions than a system that sends responses to customers. The boundary must be explicit before deployment. See Agent Governance Advisory for how permission boundaries apply to agentic systems.
Quality signal. How the team will know the system is producing correct output. This requires a measurement method that does not depend on the system evaluating its own work. Manual spot checks, independent validation sets, or downstream outcome tracking qualify. Self-reported accuracy metrics from the vendor do not.
Review path. The process for detecting degradation, escalating failures, and deciding when the system needs retraining, redesign, or retirement. A system with no review path will run until it produces a visible failure.
Rollback and stewardship. What happens when the system needs to be turned off. Who reverts to the prior workflow. What data is retained. Initiatives that cannot describe their own rollback path are not ready for production.
- Named accountable owner inside the operating company
- Specific workflow with measurable inputs and outputs
- Named source systems and confirmed data access
- Identified data owner who can confirm quality and freshness
- Explicit permission envelope with human-review boundaries
- Independent quality signal that does not rely on self-reporting
- Defined review path for degradation detection and escalation
- Documented rollback path with named steward
The Enterprise AI Portfolio Triage Worksheet provides a structured way to capture this evidence before a funding conversation.
What Not To Fund Even When The Vendor Deck Is Strong
Some initiatives sound compelling and still should not receive capital. The triage model catches missing evidence. This section catches initiatives that are strategically wrong for the portfolio at the current stage.
The cross-company platform play. A vendor proposes a unified AI platform across all portfolio companies. The deck is polished. The pricing may look attractive. The problem: no single company has validated the workflow independently. A platform decision before a workflow decision locks the portfolio into architecture that may not fit any individual company. Fund the workflow first. Evaluate the platform after evidence exists.
The prototype without an operating owner. An internal team has built a working demo. The demo is impressive, but there is no workflow owner outside the builder group, no production integration path, and no maintenance plan that survives the team’s next priority shift. A prototype without an operating owner is a research artifact, not a funded initiative.
The board-driven urgency. A board member saw a competitor announcement. The portfolio company receives pressure to “move on AI.” The resulting proposal is broad, the use case is vague, and the timeline is compressed by external anxiety. Run the evidence gates. If the initiative passes, fund it. If it does not, the board pressure is the reason to defer, not the reason to accelerate.
Where Vendor Proposals Usually Hide Risk
Vendor-led AI proposals follow a predictable pattern. They describe the use case, the technology stack, the implementation plan, and the expected outcome. What they often omit is where the portfolio operator carries risk.
Integration scope. Proposals describe integration as a line item. They may not name the specific systems, data format mismatches, or API constraints. When integration scope is vague, the operating company absorbs the cost of discovery.
Permission boundaries. Most vendor proposals describe what the system will do. Fewer describe what it is not allowed to do. A system that processes customer communications, financial records, or regulatory filings needs explicit permission limits. Without them, the first incident defines the boundary instead of the design.
Data quality assumptions. Vendor proposals assume clean, accessible data. Portfolio operators know this assumption often fails inside operating companies. The triage question is whether the system works with the data the company actually has.
Maintenance and renewal. Proposals focus on deployment. They may not address what happens after launch, when the data changes, the workflow shifts, or the vendor implementation team moves on. The portfolio operator needs to know who maintains the system, at what cost, and under what conditions it gets retired.
An Agentic Portfolio Review can surface these gaps across a portfolio, identifying which vendor proposals carry hidden risk and which internal builds lack evidence to proceed.
How To Run The First Triage Pass
The first triage pass is a classification pass, not a deep implementation audit. The goal is to make the initiative set legible enough for capital and operating decisions.
Step 1: collect all active and proposed AI initiatives. Include vendor proposals, internal builds, pilots already running, and initiatives that were approved but never started.
Step 2: score each initiative against the eight evidence dimensions. Use a simple signal: evidence present, evidence partial, evidence absent. The pattern across dimensions matters more than any single score.
Step 3: classify. Initiatives with evidence across all dimensions are Fund candidates. Initiatives with specific missing evidence are Defer candidates. Initiatives with real problems but structural flaws are Redesign candidates. Initiatives with no owner, no measurable workflow, or ungovernable data are Kill candidates.
Step 4: present the classification to the operating group. Show the evidence, the classification, and the specific gaps. Disagreement is acceptable when it is about evidence, not enthusiasm.
Step 5: define review gates. Fund candidates need a review gate before work starts. Defer candidates need unblock criteria. Redesign candidates need a revised proposal boundary. Kill candidates need a documented reason and closure.
This process does not require every reviewer to be an AI specialist. It requires discipline in asking the same questions for every initiative and refusing to fund what cannot answer them.
For portfolios that need deeper technical evaluation, Enterprise Agentic Advisory provides principal-level assessment of AI system architecture, vendor proposals, and production readiness.
FAQ
How many AI initiatives should a portfolio operator fund at once?
Fund only the initiatives that have a named owner, documented workflow evidence, accessible data, and a defined review path. The constraint is not the number of ideas. It is the operating attention available to steward each system after launch.
What evidence does an AI initiative need before it deserves a Fund decision?
At minimum: a named internal owner, a specific workflow with measurable inputs and outputs, confirmed source data access, an explicit permission boundary, and a review path that does not require the project team to self-certify quality.
When should a portfolio operator kill an AI pilot instead of deferring it?
Kill when the initiative has no accountable owner, depends on data the company cannot access or govern, or has no measurable acceptance criteria. A deferred project retains a path to Fund. A killed project should not return without a materially different proposal.
The Decision Rule
Do not fund AI across a portfolio from vendor pressure, board anxiety, or platform standardization. Underwrite each initiative like capital allocation: workflow evidence, accountable ownership, data readiness, and maintenance burden first; spend second.
For the broader operating framework, see Portfolio AI Capability. For the self-assessment artifact, start with the Enterprise AI Portfolio Triage Worksheet.
Portfolio AI Triage
If your portfolio needs a structured review of AI initiatives, Portfolio AI Capability provides the evidence framework to classify what deserves funding, what should wait, and what should stop.