Prioritizing Where to Start

The surface from Chapter 5 contains fifteen to thirty AI candidates, each anchored to the strategy. The next question is which two or three to start with.

This chapter explains how to make that choice. The work is comparison, not generation — and the discipline is to make a defensible call without producing more analysis than the candidates warrant.

Why two or three

The number is deliberate. It is not five, and it is not one.

An organization at the start of an AI adoption effort has no track record running AI projects, no policy layer to support them, and no muscle memory for the kinds of trade-offs the work demands. Trying to run five projects in this state means running all five badly — capacity dilutes, governance attention fragments, and the projects compete for the same scarce resources without anyone making explicit calls.

Running just one project produces the opposite problem. There is no comparison to learn from. The single project becomes the entire reference set for what AI in this organization looks like, which biases everything that follows.

Two or three is the right shape. Enough variety to compare across — different categories of opportunity, different readiness profiles, different risk types — and few enough that each project gets the attention it needs to actually run. The portfolio thinking that the framework defers (covered in Chapter 14) starts to become relevant once three or four projects are in motion. Two or three is the entry point.

The three criteria

Selection at this stage uses three lenses. None of them are heavy. Each is applied with the same lightness as the candidates themselves.

Time-to-value. How quickly will this project produce evidence the organization can read? Some opportunities take a long time to show whether they work; others can produce signal much sooner. Early in an adoption effort, the value of a project is not just whether it eventually succeeds — it is whether it produces evidence fast enough to inform the next round of decisions. A project that will not produce signal for a long time contributes nothing to the organization’s learning in the meantime. At least one of the starting projects should be fast-feedback. This is how the framework’s bias toward motion expresses itself in selection.

Learning yield. What will running this teach the organization that it does not currently know? Different projects teach different things. An operational efficiency project teaches you about your data infrastructure and your governance speed. A customer delight project teaches you about your brand reflexes and what your culture will and will not deploy. A novel capabilities project teaches you about your fluency, your imagination, and your risk appetite for unfamiliar terrain. The first batch should be chosen partly to teach across the dimensions the organization is going to need to know about — not all of them at once, but a deliberate spread.

Risk diversity. Do not start three projects that all live in the same risk bucket. If all three depend on the same data source, one stalled data agreement freezes the whole portfolio. If all three are customer-facing, the organization has no operational or capability evidence to compare against. If all three lean on the same person’s bandwidth, you have one bottleneck and three projects waiting on it. Diversity at the start hedges learning across multiple risk types, so that whatever comes back, something is moving.

These three criteria interact. A project with high time-to-value and high learning yield in a category nothing else in the batch occupies is an obvious yes. A project with low time-to-value, modest learning yield, and a risk profile already covered by another candidate is an obvious no. The interesting work happens in the middle — comparing candidates that score differently on different dimensions and deciding what mix the organization needs first.

What not to optimize for

Some criteria look reasonable and produce bad selections.

Most exciting use of AI. A novel application of the technology is not the same as a valuable one. The strategy-first principle from Chapter 2 still applies here: the question is not whether AI does something interesting, but whether what it does advances the strategy.

Easiest to sell internally. The political path of least resistance often produces the lowest-leverage projects. A candidate that gets fast approval because it threatens nothing usually accomplishes very little.

Highest projected ROI. ROI projections on novel AI work are speculative at best. Treating them as decisive turns the prioritization conversation into an exercise in spreadsheet creativity rather than strategic judgment. Use ROI estimates as one signal among several, not as the deciding factor.

Lowest risk. A starting batch composed entirely of low-risk projects produces a thin portfolio that teaches the organization very little. Risk is not the enemy at this stage; the enemy is risk the organization cannot afford or learn from.

Making the call

The comparison happens with the same lightness as the surface. Candidates are short. They get laid out side by side, scored loosely against the three criteria, and discussed against the strategy. The conversation does not require a deep analytical phase before it can begin — that is exactly the over-planning the framework was built to prevent. A candidate that has been described in three sentences has been examined enough to compare. A candidate that has been turned into a fifteen-page brief before the comparison has accidentally become a commitment.

The output is two or three projects that move forward. Each gets a project canvas, and the work shifts to Part III. The remaining candidates stay on the surface. They did not lose. They simply did not win this round. As evidence comes back from the starting batch, some of them will move into the active set; some will become irrelevant; some will look very different in light of what the first projects revealed.

Holding the surface forward

The surface is not discarded after prioritization. It travels forward as a working artifact, available to the next round of selection when capacity opens up. New candidates can be added to it as the organization learns. Old candidates can be removed as they become irrelevant or are absorbed by other work.

In an organization that has been running this framework for some time, the surface is a living document — always present, always being revised, never quite finished. It is the visible inventory of where AI might still advance the strategy, and it is the input to every prioritization conversation the organization has from this point forward.

The next chapter takes up the question that sits above all of this work: when does the strategy itself need to be revisited?