If you’ve felt the pace of AI innovation is surpassing your ability to keep up, you’re not alone. A recent survey of 2,000 CIOs and CTOs across 33 countries found that 70% of respondents felt their teams were deploying AI faster than IT could track.
These numbers are often read as evidence that GenAI technology is not ready for the enterprise. Foundation models improve almost quarterly, resolving previous issues each time. What's inhibiting organizations' ability to scale GenAI implementations is their ability to build systems around frontier models, including governance and data strategy. Additionally, the teams’ understanding of the technology determines whether a deployment delivers value or becomes a sunk cost.
Ironclad shares some of the recurring mistakes organizations make when trying to use GenAI and the implementations to fix them.
1. Treating foundation models as universal tools
The instinct to route every task through the largest available model is understandable, given its capabilities. But conflating capability with efficiency is a costly mistake. A recent Gartner survey found that only 28%of AI infrastructure projects deliver the promised return, and one in five failed outright because they were “overly ambitious or poorly scoped.” Matching a model to a task’s complexity and risk profile is a critical step for delivering better ROI.
Extraction of a contract’s data from a complex legal risk analysis are two applications that require different models because of the associated work, risk, and validation. Could a powerful model deliver on both tasks? Probably, but that’s the wrong question to ask. The right question: "Which model delivers the best ROI for a given task?”
2. Underspecified objectives, i.e.what AI should and shouldn't do
Generative AI operates probabilistically and must be used accordingly. When you give a model an open-ended objective without constraints, it will likely take the path of least resistance to satisfy it, and that path may diverge significantly from what you intended.
Research shows that detailed prompts with structured guidelines will generate more accurate results. When tuning models using methods like Reinforcement Learning, where AI systems improve through iterative feedback, a model's improvement is tied to clear goals, as well as what not to do. For complex enterprise work, building production-ready systems should include reference documents, boundary conditions, and explicit constraints that give models the necessary context.
3. Building feedback loops around the wrong data
Unlike previous software, GenAI tools are shaped through interaction data: Every prompt, correction, and accepted output teaches the system what "good" looks like. But that learning loop only works if the data collected is representative of the full range of intended applications.
If an enterprise’s finance team uses an agentic expense management tool primarily for one region's spending patterns, the system learns specific tax codes, vendor classifications, and reporting requirements that differ significantly from those used in global applications. Unaware of this training, the team inadvertently builds a compounding deficiency with the tool’s applicability.
It’s critical to assess AI tools against the full scope of work your team actually does. To ensure feedback loops share the right data, it's important to monitor inputs and regularly evaluate the tool’s outputs.
4. Relying on one model for both creation and verification
As AI-generated code and content scale beyond what any team can manually review, engineering organizations are building verification models to check the output of generation models. This layered approach is architecturally sound, but only when the verification model is genuinely independent of the generation model.
Using the same provider's model to generate and verify output is the equivalent of seeking a second medical opinion from the same doctor. Models from different providers, trained on different datasets with different methodologies, carry different biases. Cross-provider verification introduces the randomization needed to surface errors that single-model confidence will mask.
5. Ignoring hidden errors in the technology
The wide gap between what these models can do and how they’re applied is certainly a factor for enterprises’ failed attempts, as the Gartner report illustrated. When developing software for past mobile or cloud technologies, an incorrectly structured push notification or cloud deployment generated a failure notice, but that's not the case with GenAI. Even poorly framed inputs will yield seemingly plausible outputs, hiding the underlying structural issues.
Closing the gap requires teams across an enterprise to have operational fluency with the technology. Designers, engineers, and product managers should all have equal footing. Otherwise, engineers, agentically coding, are constantly seeking new tasks, while designers and product managers inefficiently wire together new features.
What these mistakes share
These five mistakes all stem from treating GenAI as a technology problem, but it’s actually a systems problem. The models are the easiest component to improve. Governance, data discipline, and organizational adaptation are where value is actually created or destroyed. Successfully scaling GenAI usage starts with teaching and building the understanding of how to use the technology, because without it, better technology only scales the same mistakes faster.
This story was produced by Ironclad and reviewed and distributed by Stacker.
(0) comments
Welcome to the discussion.
Log In
Keep it Clean. Please avoid obscene, vulgar, lewd, racist or sexually-oriented language.
PLEASE TURN OFF YOUR CAPS LOCK.
Don't Threaten. Threats of harming another person will not be tolerated.
Be Truthful. Don't knowingly lie about anyone or anything.
Be Nice. No racism, sexism or any sort of -ism that is degrading to another person.
Be Proactive. Use the 'Report' link on each comment to let us know of abusive posts.
Share with Us. We'd love to hear eyewitness accounts, the history behind an article.