The AI Use Case Your CEO Loves Is Probably the Wrong One to Build First

A regional bank CEO I worked with wanted to start their AI program with a customer-facing chatbot. He had seen a demo at a fintech conference, talked to the vendor over drinks, and came back convinced this was the project that would define the bank's digital transformation. His enthusiasm was infectious. The board loved it. The marketing team loved it. The CTO quietly pulled me aside and said, “This will fail. Our data infrastructure cannot support it. But nobody wants to hear that right now.”

The chatbot project launched with $1.2M in funding and a 6-month timeline. It delivered a working prototype in 8 months, after two vendor switches and a complete redesign of the data pipeline. The prototype handled 30% of customer inquiries adequately. The remaining 70% were routed to human agents, who now had to context-switch between the chatbot's partial conversations and direct calls. Customer satisfaction scores dropped 11 points. The project was quietly shelved 14 months after launch, having consumed $2.3M — nearly double the budget.

Meanwhile, the bank's operations team had proposed a document classification project for mortgage processing. No CEO excitement. No conference demos. No board presentation. Just a team that spent 40 hours per week manually sorting mortgage documents into 23 categories and knew the data was clean because they had been digitizing it for three years. That project was deprioritized because the CEO wanted the chatbot first. When it finally got funded — 16 months later, after the chatbot failure — it shipped in 9 weeks, cost $185K, and saved $420K per year in labor. The payback period was 5 months.

Why Executive Enthusiasm Is a Bad Predictor

I have tracked the outcomes of 43 enterprise AI projects across banking, insurance, and healthcare over the past four years. I categorized each project by how it was selected: executive-championed (the CEO or a C-suite sponsor chose the project based on vision or vendor influence), committee-selected (a cross-functional team evaluated multiple candidates), or scoring-rubric-selected (the team used a structured scoring framework to rank use cases against objective criteria). The results were stark.

Executive-championed projects had a 34% success rate, defined as reaching production and delivering measurable ROI within 18 months. Committee-selected projects landed at 52%. Scoring-rubric-selected projects hit 78%. The pattern held across industries and company sizes. The more structured the selection process, the higher the success rate.

Executive enthusiasm fails as a selection criterion for three reasons. First, executives are drawn to customer-facing, visible projects because those are what get discussed at conferences and board dinners. But customer-facing AI requires the most mature data infrastructure, the most complex integration, and the most rigorous testing. These are the hardest projects to execute, and they are disproportionately selected as first AI initiatives when the organization has the least AI experience.

Second, executive-championed projects carry political weight that distorts honest assessment. When the CEO picks a project, nobody wants to be the person who says it will fail. Risk assessments get softened. Timeline estimates get compressed. Data readiness issues get described as “challenges to manage” rather than “blockers that will add four months.” The CTO at that regional bank knew the chatbot would fail. He told me privately. He did not tell the CEO, because the CEO did not want to hear it.

Third, executives tend to anchor on the demo, not the deployment. A vendor demo shows what the technology can do in ideal conditions with curated data. Deployment requires your data, your infrastructure, your compliance requirements, your users. The gap between demo and deployment is where most executive-championed projects die, because nobody assessed that gap before committing budget.

The Five Dimensions That Actually Predict Success

Across the 43 projects I tracked, five factors consistently separated the successes from the failures. These are not the factors executives naturally evaluate. They are the factors that a structured scoring rubric forces you to evaluate.

Data readiness. Is the data already digital, structured, and accessible? The mortgage document classification project succeeded because the operations team had been digitizing documents for three years. The data was clean, labeled, and sitting in a system the team controlled. The chatbot project required pulling data from six different systems, three of which had no API. Data readiness is the single strongest predictor of first-project success, and it is the dimension most often ignored in executive selection.

Process maturity. Does a well-defined, repeatable process already exist? AI works best when it automates or augments a process that humans already execute consistently. If the process varies by person, location, or day of the week, you are not deploying AI — you are trying to standardize a process and deploy AI at the same time. That is two transformation projects stacked on top of each other.

Measurable impact. Can you quantify the current cost and the expected savings in dollars before you start? The document classification team could say: “We spend 40 hours per week on manual classification at $52 per hour. That is $108K per year. If AI handles 80% of classifications, we save $86K per year.” The chatbot team could only say: “We believe this will improve customer satisfaction and reduce call volume.” Projects with vague impact statements are three times more likely to lose funding during execution because the business case dissolves the moment the project hits a delay.

Technical complexity. How many systems need to integrate? How many data sources? How many compliance requirements? A project that touches one system, one data source, and one compliance framework is a manageable first AI initiative. A project that requires data from six systems, real-time integration, regulatory compliance across three jurisdictions, and a customer-facing interface is a third or fourth AI project, not a first.

Organizational readiness. Does the team that will use this AI output actually want it? The operations team at the bank was begging for automation. They had proposed the document classification project three times. They were ready to adopt on day one. The customer service team, by contrast, viewed the chatbot as a threat to their jobs. Adoption resistance added four months of change management that was not in the original timeline or budget.

The Scoring Session That Changed One Company's AI Program

An insurance company I advised had seven AI use cases on the table. The Chief Digital Officer wanted to build a claims fraud detection system. It was the biggest potential savings — $3.2M per year if it worked. The CEO agreed. The board was excited. It was the obvious choice.

I asked them to score all seven use cases across the five dimensions before committing. The fraud detection system scored poorly on data readiness (claims data was spread across four legacy systems with inconsistent formatting), process maturity (the investigation process varied by adjuster and region), and organizational readiness (the special investigations unit was skeptical of automated detection and concerned about false positives triggering unnecessary investigations). Total score: 11 out of 25.

The highest-scoring use case was one nobody had championed: automated policy document classification. The underwriting team manually categorized incoming policy documents into 31 categories to route them to the right review team. The data was already digital (scanned PDFs in a single document management system). The process was fully standardized (a 31-category taxonomy that had not changed in four years). The impact was measurable ($175K per year in labor). The technical complexity was low (single system, single data source, no customer-facing component). The team wanted it (the underwriting assistants had been asking for automation for two years). Total score: 22 out of 25.

The company built the document classification system first. It shipped in 11 weeks, cost $140K, and hit 93% accuracy in the first month. Payback period: 10 months. More importantly, the success gave the AI team credibility. When they came back to the board six months later with a proposal for fraud detection — this time with a realistic 18-month timeline, a phased data integration plan, and a change management budget — the board approved it in one meeting. The first project bought them the trust to do the hard project right.

How to Run the Scoring Session

The scoring session works best when it is cross-functional and time-boxed. You need representation from IT (for data readiness and technical complexity), operations (for process maturity and organizational readiness), and finance (for measurable impact). The session should take 90 minutes. Any longer and people start lobbying for their preferred project instead of scoring objectively.

Each use case gets scored 1 to 5 on each of the five dimensions. A 1 means the dimension is a significant blocker. A 5 means the dimension is a clear strength. The scoring is done simultaneously — each participant writes their scores before anyone shares — to prevent anchoring. Then you average the scores and discuss any dimension where individual scores diverge by more than 2 points. That divergence usually reveals information asymmetry: someone knows something about the data quality or the team's readiness that others do not.

The total score determines priority. But there is one override rule: any use case that scores below 2 on data readiness should be deprioritized regardless of total score. Data readiness is not a dimension you can fix during the project. If the data is not ready, the project will stall at data preparation and never reach the AI part. I have seen this pattern in 14 of the 43 projects I tracked. Every one of them exceeded budget by at least 60%, and 9 of the 14 were eventually cancelled.

The Political Problem

The hardest part of structured scoring is not the framework. It is telling the CEO that their favorite project ranked fourth. This is where the rubric becomes a shield rather than a weapon. You are not saying the CEO's project is bad. You are saying the data shows it should be the third or fourth project, not the first. The rubric makes the conversation about data, not about opinions. And most executives, when shown a clear scoring framework with transparent criteria, will accept the ranking — especially if you frame the first project as building the foundation for their preferred project.

At the insurance company, the CDO's fraud detection project eventually launched as Project 3. By then, the team had built data integration patterns they reused, the organization had seen two AI successes, and the special investigations unit had warmed to the idea after watching other departments benefit. The project that would have failed as Project 1 succeeded as Project 3 — not because the technology changed, but because the organization was ready.

Actionable Takeaway