Today, a RAND Corporation report notes estimates that more than 80% of AI projects fail, roughly twice the failure rate of comparable IT projects. Abandonment is another issue on the rise: Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027.
What is failing is not the technology at hand. It's the same operational pattern that catches many businesses off guard, which happens when businesses move fast without putting the right systems and infrastructure in place.
The tell is always the same
When I look at why AI initiatives stall, the root causes are similar to the many hurdles businesses face when they look to implement a new solution with no clear definition of what success means, and no foundational operational capabilities in place. The excitement with AI tools supersedes the necessary steps required for a thought-out strategy. And with AI tools, they don't stay still once deployed, they drift, requiring a more robust system to just monitor them. As with anything in business, a disciplined approach to implementing new strategies is critical for success.
JPMorganChase offers a useful example of scale paired with operating discipline. The firm has reported more than 500 AI use cases in production and access to LLM Suite for 200,000 employees. Its public materials emphasize controlled access, shared infrastructure, business-aligned workflows, and measurable efficiency gains. The lesson is not that every company should copy its scale. It is that adoption is treated as an operating system, not a collection of isolated tools.
The AI Curation Loop
So what does an effective operating structure and governance look like? This is the framework and questions I ask clients pursuing AI tools in their business.
Choose the pain, not the tool.
Before any build conversation, three things need answers:
- What's the actual business pain, in dollars or hours, not in excitement?
- Is the data and workflow infrastructure underneath even ready?
- How contained is the failure mode? A wrong AI-generated summary gets caught by a human. A wrong AI decision touching client money or patient care doesn't.
Build the one measurable proof.
Start with one workflow and measurable baseline defined before going live. Map out the real current, human-driven process and identify the critical pathways for improvement to feed it as a requirement for the agentic workflow. Then design the solution and deploy.
Additionally, embed a human reviewer to conduct quality assurance before the tool gets any autonomy. Set up the gates/stages with real measurable metrics rather than a vague sense that it's "going well." You are proving the concept at this stage, ensuring the process for deployment works just as the tool or workflow itself works. This is critical for adding more use cases in the business.
Adopt the use and trust.
Treat adoption in two buckets: use and trust. Adoption of use means people are actually opening the tool. Adoption of trust means people believe what it tells them enough to act on it, rather than clicking through it because it's now sitting in their workflow. High usage with low trust looks like a win on a dashboard and is actually a quality problem. You can't manufacture either one by announcing the tool is exciting. People adopt what visibly makes a specific part of their day easier, and that has to be demonstrated, not declared.
Operate the solution for sustainability.
The stage where it's sometimes an afterthought for businesses, and where the expensive surprises live. Set three systems for maintaining the deployed models:
- Measure the drift:
- Cost drift, where the tool still works but compute or labor spend quietly outpaces the value it delivers
- Performance drift, where the model's accuracy degrades against its original baseline
- Value drift, where the tool still works fine, but the business problem it solved stopped mattering because the organization moved on or strategies shifted
- Staff appropriately. Engineers who build and engineers who monitor drift should be separate. Businesses can create internal AI education and champions to offset some of the operational or administrative work that will arise with these models.
Evolve with your business, not with the hype.
Scale what's proven. Shut down what isn't. Have a decommissioning process to effectively retire models that are outgrown or have not shown value. And most importantly, scale gradually. Just because the models are providing real business impact does not mean you need to add many more tools for the sake of having them. Business problems and desired outcomes should drive the solutioning, not the other way around.
The real question isn't "should we use AI"
It's what does success look like in six months, in a number you can point to, and who owns watching it after the launch excitement fades? If you can't answer that yet, that's not a reason to wait. It's a reason to build the operating structure first, the same way you would for any new hire, process, or system entering the business. AI is just the newest thing that's shown up needing one.