# Fund AI Work That Can Earn Acceptance

This is the executive route through The Applied AI Field Guide. It is for the person deciding whether one AI initiative should receive another tranche of money, time, or organizational attention.

The decision is not “Is the demo impressive?” It is:

> Does this workflow have an owned outcome, a credible way to verify it, a tolerable authority and loss boundary, and a plausible positive value case after full cost?

If one of those four gates is missing, fund discovery—not deployment.

## Ask for one decision contract

Before a pilot, require a short, inspectable answer to these questions:

| Question | Evidence worth funding | Warning sign |
| --- | --- | --- |
| What work is changing? | One observed workflow, named user, recent cases, exceptions, and current fallback | A broad function, persona, or model capability |
| What outcome matters? | A source event or accountable reviewer can independently accept it | “The agent completed the task” |
| Who owns the decision? | Business, operational, risk, technical, and receiving-service owners are named | The delivery team owns every decision indefinitely |
| What may the system do? | The maximum effect, approval path, stop conditions, and recovery are explicit | Autonomy is treated as the objective |
| What must be true economically? | Baseline, eligible population, attribution, full cost, residual loss, and stop threshold are falsifiable | Time saved is annualized before adoption or acceptance is measured |
| What happens when the brief is wrong? | Competing claims stay cited; an authorized person accepts, rejects, or defers a scoped reframe | The project quietly rewrites its goal after failure |
| Can the operating team own it? | Users can inspect and correct the work; the receiving team exercises release, support, incident, rollback, and retirement | Handoff means a document folder and a training call |

## Release money in evidence stages

Use a sequence with a decision at each boundary:

```text
observe the work
  -> bound the outcome and economics
  -> make data fit for the decision
  -> prove one controlled slice
  -> shadow with real users
  -> canary one named segment
  -> continue, reshape, expand, pause, or stop
```

Keep technical performance, operator acceptance, adoption, business value, full economics, production readiness, and receiving-team capability as separate gates. A strong score in one cannot average away a failed authority or safety boundary.

## Read one worked decision

The [invoice-exception engagement](https://davidahmann.github.io/applied-ai-field-guide/worked-engagement/invoice-exception/) begins with a fictional sold promise: automatically resolve and post eligible exceptions. Field evidence shows that a reviewer must approve a staged correction before posting.

The reframe narrows the slice to recommendation and staging. It preserves the original promise as superseded history, keeps the manual queue as the safe fallback, and names the next field move. The value case then makes its assumptions visible: under the illustrative numbers, the forecast net value is only $320 per month. Small misses in adoption, acceptance, review effort, or support cost erase it.

The fixture runtime passes five committed deterministic cases, including authority and duplicate-safety paths. That is useful engineering evidence. It is not customer value, adoption, deployment approval, or handoff proof. The worked review therefore says: continue as a review-only shadow candidate; do not deploy.

That is what an honest gate looks like. Evidence can support another bounded step without supporting production.

## Use the full framework when needed

The definitive value framework is the [12 Factors of AI Value Engineering](https://davidahmann.github.io/applied-ai-field-guide/ai-value-engineering/). Use its [one-page scorecard](https://davidahmann.github.io/applied-ai-field-guide/ai-value-engineering-scorecard/) for a live decision. Then follow the [canonical lifecycle](https://davidahmann.github.io/applied-ai-field-guide/#from-idea-to-production) and retain the exact artifacts that justify each transition.

The Guide does not certify an initiative. It gives funders, delivery leaders, operators, and engineers a shared way to ask what has been proved, what remains assumed, who owns the next decision, and when the rational answer is to stop.