Most enterprise AI is funded the way you fund a tool, and that is why most of it dies at the first cost review. A seat count, a token bill, a prompt volume — these measure activity, not return, and no CFO funds activity with conviction. The pilot impresses, the demo lands, and then someone asks what it moved on the P&L, and the answer is a usage chart. Usage is not an outcome. An agent measured only by usage is an expense looking for a justification, and it loses every budget fight to a project that can name its number.
The fix is not better dashboards. It is a different funding model: stop funding agents like IT projects and start funding them like value streams. A value stream has an owner, a metric, and a line of sight to cash. So should every agent the enterprise puts into production.
No value card, no scale funding
Every agent in a harness carries a value card — five fields, no exceptions:
- Baseline — what the metric reads today, before the agent.
- Target — what it should read after, and by when.
- Value lever — which of five it moves: revenue, margin, working capital, productivity, or risk.
- Owner — a named executive accountable for the number, not the agent.
- Attribution method — how you will prove the agent caused the change.
The rule is blunt: no value card, no scale funding. An agent can earn a pilot budget on a hypothesis. It cannot earn a scale budget — production access, headcount reallocation, a line in next year's plan — without a card. That single gate is what separates a portfolio you can defend to a board from a pile of experiments you cannot.
The value spine
A value card is credible only if it connects what the agent does to something a shareholder cares about. That chain is the value spine, and every link has to hold:
Agent activity → process metric → financial/risk metric → executive value lever → shareholder value.
A reconciliation agent clears items (activity). Recon aging falls and unresolved items drop (process metric). The close completes faster with fewer manual journals (financial/risk metric). That is productivity and risk reduction (lever), which shows up as faster, more reliable reporting and lower control cost (shareholder value). Break any link and the card is a story, not a thesis. The discipline is refusing to fund a card whose spine has a missing vertebra — an agent that moves a process metric no one can tie to a financial one is busy, not valuable.
What a funded portfolio looks like
Here is what the office of the CFO actually approves — agents as value-card line items, grouped by the process domain they run in:
| Agent | Process domain | Process metric | Value lever |
|---|---|---|---|
| Month-End Closer | R2R | Close cycle time, manual journals | Productivity, risk |
| GL Reconciler | R2R | Recon aging, unresolved items | Productivity, risk |
| Billing Dispute Triage | L2C | Dispute aging, first-pass resolution | Working capital, revenue |
| Cloud Spend | S2P | Commitment utilization, idle resources | Margin |
| Usage & Rating Integrity | L2C | Usage-to-invoice mismatch, leakage rate | Revenue, margin |
Read the table as a budget. Month-End Closer and GL Reconciler are funded out of the productivity and risk levers — they buy back finance hours and shrink audit exposure. Billing Dispute Triage is funded out of working capital and revenue — every day it takes off dispute aging is cash pulled forward and revenue defended. Cloud Spend is a margin play. Usage & Rating Integrity closes the gap between what was consumed and what was billed — pure leakage, pure margin and revenue. None of these is "an AI project." Each is a claim on a specific lever, owned by the executive who already owns that lever.
Fund by lever, kill by metric
Run the portfolio the way a CFO runs any portfolio of bets — by lever, with a kill rule. Allocate to the revenue, margin, working-capital, productivity, and risk levers, fund the agents that map to each, and review them against their targets. An agent that moves its metric earns more scope and more budget. An agent that does not move its metric after a fair window gets killed — not iterated indefinitely, not quietly left running, killed. The willingness to kill is what makes the funding credible; a portfolio with no kill rule is a slush fund, and boards know it.
This reframes what the CFO is actually buying. Funded as a project, an agent is an expense that expires — the budget is spent, the quarter closes, the capability is rented for as long as the invoice runs. Funded as a value stream, the same agent is an owned, governed capability that compounds: it keeps moving its lever, the attribution keeps accruing, and next year's plan starts from a higher baseline. The CFO owns that conversion — turning AI spend into a compounding asset on the books rather than a recurring cost off them. That conversion is the whole reason finance, not IT, should hold the agent budget.
Be honest about attribution
The value card is only as good as its attribution method, and attribution is where most AI business cases quietly cheat. The honest method is baseline-versus-counterfactual: measure the metric before, hold the comparison against what it would have read without the agent, and attribute only the delta the agent can actually claim. Seasonality, a pricing change, a headcount shift — strip them out. Over-claim and the first audit of the value case torches the credibility of the entire portfolio.
The corollary is uncomfortable and load-bearing: an agent whose impact you cannot attribute gets no scale budget. Not a smaller budget — no scale budget. It can stay a pilot, it can keep collecting evidence, but it does not graduate until its card has a defensible number. This is not finance being difficult. It is the only stance that lets the CFO say yes to the next ten agents with a straight face, because every one already in production has a number that survived scrutiny.
Funding agents like value streams is what turns a noisy AI line item into a managed book of bets with owners, targets, and a kill rule. It is also what lets the enterprise scale past the pilot, because the budget conversation finally has the one thing it always lacked: a number the CFO can defend.
This is the third essay in The Harness Era. The full-length analysis — the value spine in detail, the worked value cards, the portfolio funding model, and the attribution discipline — is linked below. Next, #4 turns from the money to the machinery: The CIO's Agent Control Plane — the architecture that lets a portfolio of agents run safely at scale.