Studies report that up to 95% of GenAI pilot projects never scale beyond the lab . In fact, nearly 8 in 10 companies using GenAI have seen no significant impact on their bottom line . These stats underscore a hard truth: simply deploying a flashy AI agent is no guarantee of value.
Aadit Sheth recently shared great insights from McKinsey on why so many AI agents fail and how to avoid it. We couldn’t agree more with his key points, which include: focusing on workflow redesign over shiny tech, applying agents only to the right high-variance tasks (not every problem needs an agent), avoiding “AI slop” by treating agents like new team members (with clear roles, onboarding, feedback loops, and benchmarks), tracking every step for transparency (so silent failures don’t hide), building modular components instead of one-off agents, and remembering that humans remain essential in the loop. These are golden rules – follow them and you won’t “burn millions” on an AI science project with no ROI.
To build on those lessons, we want to zoom out and get back to basics. Before writing a single line of agent code, we should start with a thorough systems assessment. In our experience, the success of any AI solution is 90% preparation and understanding the problem, and 10% the AI magic. Here are some fundamental questions and steps to consider upfront:
Who is the user (and what do they need)? Identify your end-users or customers and their pain points. A solution should be user-centric, solving a real problem that matters in their daily work.
What is the current workflow? Map out the tasks and the process as it exists today, step by step. What are people actually doing? This helps spot inefficiencies and determine where (if anywhere) an AI agent could help.
Where is the data? Understand the information sources the workflow relies on – databases, documents, APIs, human knowledge, etc. Is the data accessible, and is it clean and reliable? An AI agent is only as good as the data and context it can leverage.
How long and costly is the process now? Establish a baseline. Measure how much time, money, or resources the current workflow consumes, and where the bottlenecks or error rates are. You need this to define what improvement counts as success (e.g. 50% faster processing, 30% cost reduction, or higher accuracy).
What are the proposed changes? Given the above, brainstorm how to improve the process. Sometimes the fix might be non-technical (streamlining steps or clarifying roles). Other times, an AI or automation could speed things up. Always tie improvements to reducing the baseline costs or improving quality identified earlier.
What’s the right human↔AI mix for the solution? Not every scenario calls for a fully autonomous agent. Determine where your solution falls on the spectrum:
By answering these questions, we ensure that any AI agent we introduce is grounded in real business needs and integrated into the fabric of the workflow. This is exactly what McKinsey emphasizes: ROI comes from strong intent – define the outcomes, embed agents deep in core workflows, and redesign processes around them, paired with clear strategy and tight feedback loops . In other words, we succeed by treating AI adoption as a holistic transformation, not a plug-and-play gadget.
Ultimately, the future of work is human + agent, together. If we do the homework on our systems and apply AI thoughtfully, we can turn that 95% failure rate on its head. By redesigning workflows with purpose and keeping humans in the loop, we turn novelty into measurable value – delivering results rather than regrets. 🚀