Elder Research practitioners argued that AI projects stall on business basics, not models. They walked through four fixes: building leadership data and AI fluency, prioritizing projects by business value, replacing gut decisions with trusted data, and giving AI teams visibility into changes to their source data.
Key takeaways
- Failure is a business problem. Elder Research says AI success depends on business transformation, not just deploying a technical solution.
- Leaders need data fluency. Leaders who can ask the right questions define the right problems and win buy-in.
- Prioritize by value. Score ideas on impact, effort and feasibility, plot them on a matrix, ship quick wins and park the rest.
- Gut decisions signal missing trusted data. A clear process, a better intake form and one source of truth cleared a year-long data access backlog.
- Tell AI teams when data changes. SLAs and simple communication habits keep upstream changes from breaking models.
Most AI projects that fail don't fail on the model. In the Data Science Connect webinar "Not Your Typical AI Webinar: 4 Practical Tips for Real Results," a team from Elder Research argued that AI results depend on four business basics: leadership fluency, project prioritization, trusted data and clear communication about data changes.
The session, sponsored by Elder Research, was facilitated by Leah Severance, Director of Commercial Business Development at the data science and analytics consultancy. She was joined by four Elder Research colleagues: learning engagement manager Ramsay Zotti, data scientist Shaylee, and technical business analysts Josh Fairile and Stephen Assink. Brian Mink, co-founder and president of Data Science Connect, hosted.
Why do so many AI projects fail?
Severance said the commonly cited 80% failure rate for data science projects is "quite conservative" based on recent studies. She listed familiar outcomes: multi-year projects that never reach deployment, solutions the business won't act on, blown budgets spent on the wrong problem, and one-size-fits-all tools that give a company the same answers as its competitors.
Elder Research had its own share of failed projects in its first decade, she said. The lesson: "data and AI success is not just about deploying a technical solution. It's really about business transformation." The firm built a four-part framework of business strategy, resource allocation, project execution and strategic deliverables, and Severance said projects that follow it have a greater than 90% success rate with its clients.
Do leaders have the data and AI fluency to ask the right questions?
Zotti opened the business strategy piece by asking whether leaders can ask the right questions, propose the right projects and empower their people. Without that fluency, he said, it is hard to define the correct problem, win buy-in or explain model performance to non-technical colleagues.
He described a customized AI training program for a Fortune 100 consumer packaged goods company, which he called "CPG" to protect its privacy. Built with the VP of the company's enterprise data science organization, it used the company's own data and use cases and covered the machine learning lifecycle, model validation, feature selection, time series forecasting, clustering, neural nets and generative AI.
In the Q&A, Zotti said the course runs 14 weeks and includes a midterm, final exam, project evaluation and presentation, and a small percentage of participants fail. Participants can bring their own work as the capstone project. He said several projects went into production and saved the company millions, in some cases tens of millions of dollars a year.
How should teams prioritize AI and analytics projects?
"Your analytics backlog can have 30 project ideas, but maybe you can only staff about five of them," Shaylee said. Choosing by gut feel, she warned, can leave strong people on low-impact work while better opportunities "collect dust."
She described a successful customer satisfaction model that spun off ideas for churn prediction, real-time alerting and new teams. The project team fragmented and the original work slipped. The fix was a half-day session with the whole cross-functional team that put every idea on the table, then asked three questions: what is the impact, what effort will it take, and is it feasible with current resources and timelines?
The ideas were plotted on an impact-versus-effort matrix. The team shipped several quick wins within weeks, scoped one major customer prediction project with dedicated resources and parked the rest. A transparent framework, she said, means "no politics, no guessing, no squeaky wheel gets greased."
Asked what signals a project shouldn't be prioritized even when it sounds exciting, Shaylee pointed to data availability. Teams have to decide whether to work with the data they have or wait for the right data.
Where are teams still making gut decisions?
Fairile said decision paralysis usually comes from a lack of trusted data, and "project execution cannot be sustained by gut level decision-making."
His example was an enterprise analytics team at a quick-service restaurant company that managed legally sensitive labor data such as employee PII, hourly wages and sick leave. When Elder Research arrived, the team had a backlog of data access requests going back more than a year. Fairile found three problems: no defined process, an outdated request form that didn't capture what reviewers needed, and information scattered across tools including Miro, Confluence, Excel, email and Teams. "Data related to their project was not accessible anywhere because it was available everywhere," he said.
The fixes were a process map, a new intake form that made legally risky use cases easier to flag, and a Power Automate flow that sent requests to a Jira board as a single source of truth.
Do AI teams have visibility into their source of truth?
Assink, who works on a client's customer data product teams, said the data feeding AI models changes. Tables are updated or retired, and upstream changes break pipelines or leave models stale.
He described a call with an AI team planning LLM-based classification work on a dataset that was about to be completely rewritten in less than a month. No one had told them. Because "without good data the most sophisticated models are irrelevant," he recommended service level agreements between AI teams and data owners covering data timeliness, quality, support channels and change notices, plus simple communication habits such as dedicated channels, regular stakeholder meetings and announcements of planned changes.
When ML teams don't trust data, he added, they build their own siloed datasets, creating tech debt and "data hoarding."
Learn from enterprise AI practitioners
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions.
Frequently asked questions
Why do most AI projects fail?
According to Elder Research's Leah Severance, projects often fail because teams act before confirming they are solving the right business problem, and because success requires business transformation rather than only a technical solution.
How do you prioritize AI and analytics projects?
Elder Research's Shaylee recommended putting every idea on the table, asking about impact, effort and feasibility, and plotting ideas on an impact-versus-effort matrix to pick quick wins and one major project.
What should a data SLA for AI teams include?
Stephen Assink said it should set expectations for data timeliness and quality, name support channels, and commit to informing affected teams about updates or changes to the data.
What happens when teams don't trust their data?
Panelists said teams fall back on gut decisions or decision paralysis, and ML teams build their own siloed datasets, creating tech debt and data hoarding.
Watch: Not Your Typical AI Webinar: 4 Practical Tips for Real Results on YouTube
Event: Webinars · More coverage: Webinars insights
