Three Scores Pick Your AI Use Cases. Conviction Decides Which Ones Ship.

95% of corporate AI pilots stall before they move the P&L. MIT counted. What the number does not say is that most of those pilots were not random experiments. They were the top workflows of a scored list. Somebody ranked them on value, feasibility, and risk, defended the ranking in a steering committee, and put the winners on a slide called the roadmap.

The scoring was fine. The list was still wrong. The three scores every advisory uses tell you which use cases deserve to ship. They do not tell you which ones actually will. That is a different question. A use case ships when a named operating leader puts their own name, their own metric, and their own calendar on it. Almost no advisory scores that, because the answers it turns up are ones nobody in the room wants to hear.

If you are a CEO with a prioritized AI list on your desk this fall, eight to twelve workflows, some green, some yellow, and you cannot tell which of them will be running in twelve months and which will be back on next year's deck, this is the test. It takes an afternoon. The list will come out shorter.

The three scores everyone uses, and the question nobody asks

Pull any AI prioritization deck from the last two years and the frame is the same.

1. Value. What the use case is worth.

2. Feasibility. Whether it can be built on today's data and systems.

3. Risk. What happens if it goes wrong.

Weight them however you like and you get a defensible list.

A round-up of 12 published AI use-case scoring frameworks, from the big strategy houses through the cloud and model vendors, found that none of them scores ownership. Not executive sponsorship, not a named owner, not anything like it. Sponsorship shows up, when it shows up at all, as a precondition wrapped inside "readiness" or "governance," a box checked once for the whole program rather than a score on each workflow. One practitioner guide states the ordering principle outright: "value first, feasibility second, governance always." Ownership is context. It is never the score.

The missing fourth lens is conviction. Specifically: a named operating leader, not the CEO and not "the executive team," who will personally see one use case through. Not the leader who got excited at the workshop. The leader whose quarterly goals are tied to it landing, who signs up for the outcome metric instead of the activity metric, who eats the redesign cost inside the function they run, and who will defend the workflow in their own staff meeting when the budget gets tight. ("Appetite" is the word trending this year for how much an organization believes in AI. Conviction is not appetite. Appetite is a mood. Conviction is a signature.)

Without that signature on a specific use case, the use case is not on a roadmap. It is on a slide.

Why a well-scored list stalls anyway

You can see the same ownership gap one level up, above the use case. In an April 2026 study of 505 Global 2000 executives by HFS Research, co-authored with the AI services firm Altimetrik:

Nearly 80% report unclear ownership of AI initiatives. (HFS Research and Altimetrik, 2026)

We see this with our own clients when we help them prioritize their workflows into a roadmap. At one company:

1. The workflow that shipped was a document-and-approval chain with manual review at every stage. The operating leader who owned it took the cycle-time number as theirs to hit, and pulled the redesign cost into their own staff meetings.

2. The workflow that stalled scored the same or better on value, feasibility, and risk. Nobody in the function that owned it would put their name on it. When the budget conversation tightened, nobody defended it. It did not get killed. It stayed on the list, which is worse, because it kept looking like a plan.

Across engagements, it usually comes down to the same thing. What separates workflows that ship from those that stall is rarely the score. It's whether a named operating leader in the owning function ties their own goals to the outcome.

This is where activity and outcome pull apart. McKinsey's 2026 State of AI survey found 80% of respondents reporting individual productivity gains from AI, against 37% reporting any effect on company-level earnings. Activity is everywhere. Outcomes belong to somebody, and in most companies nobody has been asked to be that somebody for a specific workflow on the list.

The test, one workflow at a time

Take the prioritized list. For each workflow, ask three questions and write the answers on the same page as the scores.

1. Who is the named operating leader? One person, in the function that owns the workflow. Contract approvals belong to whoever runs the approval chain. Contact-centre knowledge access belongs to whoever runs the contact centre. Customer-service escalations belong to the head of service, not the head of AI. If the answer is a committee, a program office, or "the executive team," the workflow has no owner.

2. What outcome metric have they signed for? Cycle time on the approval chain. Handle time and first-contact resolution in the centre. Days to close the month. An outcome the function already reports on, not a count of pilots launched or users onboarded.

3. What date will they defend it in their own staff meeting? Not the steering committee. Their staff meeting, where the redesign cost lands on their people, and where they will have to say the number out loud and explain it.

Run this on a twelve workflow list and you get a shorter list. Three workflows, maybe four, with a name, a number, and a date next to each. The rest is not a roadmap. It is the noise the company has been calling its AI strategy.

But avoid the trap:

A CEO who owns every item creates a list of one, maybe two. That is not a roadmap either.

The CEO owns the strategy. Each workflow still needs its own operating owner underneath, someone who can see it through without the CEO in the room.

What ownership looks like in public

The public examples I've chosen are all CEOs, because CEOs are who get written about. Look at what each of them actually did and it is operating-leader work on one named workflow.

In 2024 Sebastian Siemiatkowski put an AI agent in front of Klarna's customer-service queue and tied himself to it publicly: earnings calls, podcasts, the IPO story. By November 2025 the company said the agent was doing the work of more than 853 full-time agents and had saved $60 million.

But customers said the agent gave generic answers and could not handle the complicated cases, and in May 2025 Klarna reopened human support and started rehiring. The function's total cost went up: customer service and operations costs of $50 million in the third quarter of 2025, against $42 million a year earlier, alongside the claimed $60 million in savings. That is the version of the story most people stopped reading at.

Here's the useful part. In a February 2026 interview, Siemiatkowski said of the original position, "I had to pay a lot for saying that. People were very angry with me for saying that." Then he described a redesigned workflow, not a retreat from it. AI handles the simplest cases and supports the humans on the rest. Human support becomes the premium tier, "the VIP treatment." And Klarna is now recruiting its own most enthusiastic customers into the service team, because they already know how the product works.

Read that as our conviction test, not as a cautionary tale. The owner stayed on the outcome metric when it was embarrassing, absorbed the cost in public, and redesigned the workflow around what the metric said instead of quietly dropping the workflow. At Klarna the CEO chose to sit in the owner's chair for one workflow. In a company your size, that chair belongs to the head of service.

Two more, briefly:

1. Shopify, headcount requests. Tobi Lütke's public April 2025 memo made AI a baseline expectation and rewrote one specific workflow: teams must show why AI cannot do the work before asking for more people, and AI usage went into performance reviews. A named owner, a named workflow, a metric the function already reports on.

2. DBS Bank, customer data. Then-CEO Piyush Gupta, who stepped down in March 2025, personally led the cleanup of 80 million incomplete records that some 350 downstream use cases depended on, and took a 30% cut to his variable pay after a run of IT outages. Accepting the outcome metric, in the most literal way available.

The case for leaving the scorecard alone

There is a serious objection, and it comes in two parts.

• First, that ownership already lives inside "readiness", so a fourth lens double-counts, and that it cannot be scored on a one-to-five scale anyway, so adding it weakens the discipline that makes weighted scoring useful.

• Second, the data. The most recent survey of why AI initiatives stall, published in August 2026, lists data quality, security, system integration, adoption, and measuring business impact as the five causes. "No named owner" is not on it.

Both parts point the same way.

Conviction is not a fourth number to weight. It is pass or fail, which is exactly why it is cheap to apply and why it belongs before the scoring rather than beside it.

And the five causes on that list are where a stall shows up, not why it happens. Data gets cleaned, integrations get built, adoption gets driven, and impact gets measured when one person in the owning function has decided those things will happen and has a number to lose if they do not. Take the owner away and every one of the five is available as the reason.

Feasibility stopped separating the workflows

There is a reason this matters more in 2026 than it did two years ago. The cloud vendors now ship the hard parts as standard. Microsoft Foundry and Amazon's Bedrock AgentCore come with monitoring, evaluation, identity, and access control built in, and both run models from every major lab. Google's Gemini Enterprise ships its agent platform with single sign-on and per-user access control. For most workflows on a typical mid-market list, the feasibility scores now look very similar. Feasibility still separates the workflows that need specialized data or carry heavy compliance exposure. For the rest, it has stopped doing work.

When one of the three scores stops separating workflows, the list gets decided by whatever still does. In 2026 that is the question nobody puts on the page: who owns it.

The decision on your desk this fall

The 2027 budget cycle is in its fall planning window, and your AI workflows are likely on it. Some of them are commitments. Some are placeholders the board will be irritated by in twelve months, and today they look identical.

Pull the prioritized list. Walk the top three or four workflows with whoever built it and ask two questions for each:

1. Who is the named operating leader who will personally own this in their function,

2. What outcome metric have they signed for, with a date.

Write the answers next to the scores. If the answer is "we do not have one" for more than half the list, you do not have a roadmap yet. You have the scoring done and the deciding still ahead of you, which is a better position than it sounds, because now you know which workflows to fund.

At Architech we ask the ownership question before anything gets scored, and a workflow with no owner is out before its value is counted, however good that value score would have been. That is the first thing the AI Jumpstart does with a candidate list. Run the test yourself first. It takes an afternoon, and the shorter list is the one that ships.