The Synergy Companies All articles
Partnership Strategy

The Signature on the Output: A Decision-Rights Charter for Portfolio AI

The Synergy Companies
The Signature on the Output: A Decision-Rights Charter for Portfolio AI

Photo: UW Digital Collections, No restrictions, via Wikimedia Commons

A portfolio company can now deploy a customer-facing AI assistant in a week, often without a purchase order, a security review, or a conversation with the parent company. That speed is useful. It is also how a holding company discovers, after a customer complaint, that three portfolio businesses are using the same model vendor under different terms, with prompts written by contractors and no clear owner for the output.

The shared AI stack is not a single asset. It is a chain of decisions: which use case to approve, which prompt becomes the standard, whether a fine-tuned model belongs to the portfolio company or the center, which vendor contract governs, what data may cross which boundary, and who answers when an output reaches a customer and is wrong. A charter is not a brake on adoption. It is the assignment of signatures before the machine writes something that requires one.

The stack has layers, and each layer needs an owner

Most governance failures begin with the word "AI" used as if it described one thing. In practice, a shared stack has at least four separable layers. The first is the prompt and the surrounding workflow: the instructions, retrieval rules, guardrails, and escalation paths that turn a general model into a specific business tool. The second is the model itself, including any fine-tuned version, evaluation set, or vendor-hosted configuration. The third is the commercial contract: seats, API usage, data-retention terms, indemnity, audit rights, and renewal. The fourth is the data boundary: what the tool may read, where outputs may be stored, and which customers or employees are in scope.

A workable default is to assign the prompt to the business unit that owns the customer experience, the fine-tuned model to a shared technical review group, the vendor contract to the center's procurement and legal functions, and the data boundary to a joint review that includes privacy and information security. None of these assignments should be absolute. The point is that each layer has a named accountable executive, a documented approver, and a place where exceptions are recorded.

Two keys for use-case approval

Portfolio companies should not need center approval for every internal experiment. A shared AI charter can distinguish between sandbox use and production use. In the sandbox, a business unit may test prompts and models with synthetic or properly de-identified data, subject to ordinary security rules. Production use, especially where an output influences a customer, a price, an employment decision, or a regulated record, should require two keys: the portfolio company's accountable executive and a center reviewer with authority over cross-portfolio risk.

The two-key rule prevents both paralysis and unilateral exposure. It also creates a useful byproduct: a use-case register. The register need not be elaborate. It should record the purpose, the model and vendor, the data categories, the human review step, the customer-facing status, and the date of the last evaluation. When a portfolio company later wants to reuse the same tool in another business, the register becomes the starting point for a fresh review rather than a fresh invention.

Paying for seats without subsidizing the wrong experiment

The funding question is where shared AI programs often lose credibility. If the center pays for every seat and every model call, portfolio companies have little reason to retire weak use cases. If each business unit pays for its own experimentation, the center may never build the common infrastructure that makes the portfolio safer and faster.

A practical split is to centralize the cost of the platform layer and decentralize the cost of consumption. The center can fund the vendor relationship, the security review, the standard evaluation harness, and a small shared engineering group. Portfolio companies then pay for their own seats, API consumption, and business-specific integration work through a transparent allocation. This keeps the center from becoming an uncritical sponsor of every pilot while preserving a common floor for contracts and controls.

The allocation should be visible in the same way as other shared-service charges: a unit cost, a usage estimate, and a monthly reconciliation. AI costs are volatile enough that a fixed annual charge will either overcharge early adopters or leave the center absorbing overruns. A consumption-based recharge with a cap for the first year is more honest than a flat fee dressed up as strategy.

When an output reaches a customer and is wrong

Incident liability cannot be resolved after the fact by asking who wrote the prompt. The charter should define a single incident owner for any production use case: normally the portfolio company whose customer or employee received the output. That company owns the customer communication, the remediation, and the internal review. The center owns the cross-portfolio response: vendor notification, contract remedies, model suspension if necessary, and a decision about whether other portfolio companies must pause the same tool.

This division matters because AI incidents rarely stay local. A prompt that produces a misleading answer in one business may be copied into another. A vendor outage may affect several portfolio companies at once. A data-retention failure may implicate more than one customer contract. The charter should require notification to the center within a defined window, not to seek permission but to trigger a portfolio-level assessment.

A one-page charter for 2026

The most useful charter fits on a page. It names the accountable executive for each layer. It separates sandbox from production. It requires two keys for customer-facing or regulated use. It sets a consumption-based funding model with a first-year cap. It defines the incident owner and the notification path. It requires an annual review of the use-case register and a quarterly review of vendor terms.

None of this guarantees that every AI output will be correct. It does guarantee that when an output is wrong, the portfolio can identify who owns the prompt, who approved the model, who signed the contract, who set the data boundary, and who is responsible for the customer. That is the minimum price of moving quickly with a shared stack. The alternative is a faster deployment followed by a slower explanation.

All Articles

Related Articles

The Untouched Quarter: Why a New Acquisition Deserves Ninety Days Before Synergy Arrives

The Untouched Quarter: Why a New Acquisition Deserves Ninety Days Before Synergy Arrives

When the Center Competes: A Fair Market Test for Shared-Service Mandates

When the Center Competes: A Fair Market Test for Shared-Service Mandates

The Reclaim Clause: A Credible Insourcing Path for Shared-Service Portfolios

The Reclaim Clause: A Credible Insourcing Path for Shared-Service Portfolios