Trending Articles
Trending Topics
Trending
Enterprises are getting better at turning generative AI (GenAI) proof of concept into reality. But half of projects continue to fail, according to Gartner. It cites poor data quality, unclear business value and insufficient guardrails as common reasons why plans founder. But there are many more.
If, as Michael Dell claims, GenAI is a “generational opportunity for productivity and growth”, the stakes couldn’t be higher. So how can you make sure your organization’s projects succeed? To find out, we distil the latest research from Gartner, Prosci and Forrester, and speak to a subject expert from PwC.
No one disputes that AI is already offering value to organizations. In fact, we’d go further: from customer service chatbots to coding assistants, GenAI has the potential to transform entire industries.
But every use case generates different cost, benefit and risk calculations. It may sound obvious, but the first point to consider is whether a proposed project will actually deliver meaningful value for your business.
As Gartner argues, a lack of business value “is the most fundamental failure mode”. So first prioritize projects according to whether they will improve core metrics like revenue, cost, speed and customer experience. Then assess their feasibility, by looking at data quality, model suitability, workflow integration, people and platform.
The following are foundational elements that will help to define the success or failure of a project.
It all starts with data. Poor quality data means you’re building your GenAI project on sand. This will rapidly erode trust by producing unreliable outputs and models that can’t be fine-tuned effectively. Instead, ensure your data is reliable, consistent, well governed, and aligned with your specific AI use-base requirements.
Counter-intuitively, Gartner suggests that “high-quality” data does not necessarily make it AI-ready. “When thinking about data in the context of analytics, for example, it’s expected to remove the outliers or cleanse the data to support the expectations of the humans,” it explains. “Yet, when training an algorithm, the algorithm will need representative data. This may include poor-quality data, too.”
The bottom line is, if your organization treats data as a strategic asset, and prioritizes data governance, it will be on the right path. If you don’t have GenAI data management experts in your organization, consider training teams up with these specialized skills.
Next, is the AI model you’re going to deploy suitable for the use case you’ve selected? Fail to get this right and costs and risk could soon spiral. Consider whether the model is aligned with your sovereignty/residency requirements. Can the provider guarantee they won’t use your data to train the underlying model? Can it demonstrate high levels of accuracy in your specific domain/industry/use case?
“Leaders should ask whether the model is truly suited for the task and whether its limitations are understood,” says PwC’s US and Global Chief AI Engineering Officer, Scott Likens. “Model readiness isn’t just about accuracy in testing. It’s about reliability, explainability, and performance in real-world conditions.”
Perhaps just as importantly, what are the model vendor T&Cs like? How easy or disruptive will it be to jump ship if things don’t work out?
Another important factor determining a model’s suitability is how well it will fit with your existing enterprise workflows. Models can work fine in isolation but still fail to create the impact you’d like once operational.
This can lead to wasted time, money and effort, and the risk of shadow AI tools proliferating, as users seek more intuitive alternatives. According to Microsoft, 71% of UK employees have used unapproved consumer AI tools at work, and 51% continue to do so every week.
PwC’s Likens says workflow integration means more than plugging your GenAI into existing systems via APIs. “It requires aligning decision rights, updating processes, and making sure outputs trigger the right downstream actions automatically and consistently,” he argues.
“In some cases, it means redesigning parts of the workflow altogether. It also requires visibility, transparency and accountability for AI itself as it takes action.”
The real difference between success and failure is down to your people.
Research from Prosci reveals that nearly two-thirds (63%) of AI implementation challenges are down to human factors, versus just 16% ascribed to technical issues. Forrester agrees that people are “often the most critical and hardest to master” of potential AI success factors.
C-level commitment is the first thing to secure. Informed leadership will set the tone, but ownership is also important at a workflow level, PwC’s Likens told us. Not everyone needs to have deep AI expertise, but decision makers should have the fluency to understand what questions to ask.
In fact, a diverse set of complementary skills will best serve the project. The key is integrating these across teams that may hail from various functions – IT, engineering, operations and the business. Ensure stakeholders from these disparate teams are heard from the start.
AI skills are in high demand, so effort must go into attracting and retaining talent, especially in areas like data science, engineering and analytics. Experts agree that culture is also important. Forrester argues that “inquiry and continuous learning” are vital to drive long-term success, while well-structured change management helps to shift mindset and behavior.
Finally, it’s time to think about the technology infrastructure supporting your GenAI project. There is no shortage of options on the market. But that makes it more important to know the right questions to ask of prospective vendors.
How do they support critical functions like security, governance and observability? Do they monitor in real-time for hallucinations and data leakage? Does RAG infrastructure support your existing business software stack?
For many organizations, it may be preferable to keep infrastructure on-premises – taking advantage of the superior customization, flexibility and control it offers. If you have strict regulatory and data security requirements, this could also be the option for you. Open-source software will help to avoid vendor lock-in without sacrificing quality.
If you don’t have the appetite or resources to build everything from scratch, consider starting with off-the-shelf GenAI tooling that can be fine-tuned with your data.
Before you start, remember to work out how to measure success. Evaluation is often an afterthought but shouldn’t be, argues Forrester Principal Analyst, Boris Evelson. “This is a complex process and few organizations today implement a structured approach,” he explains.
To get measurement right, start by documenting strategic-level KPIs for every business function, linked to each GenAI initiative. Use observability tools to understand how the AI is being used, and put in place a consistent measurement process.
Finally, use evidence-based ROI to demonstrate impact to senior executives. Fail in this crucial final stage, and your project may be short lived.
