Agentic AI Is Heading for a Reckoning. After 26 Years Building These Systems, I Can Tell You It Won't Be the Model's Fault.
- Dimple Thakkar
- Aug 6
- 4 min read
I've spent twenty-six years in data and AI — not observing it from the sidelines, but building it. I came up the hard way, one shipped product and one hard lesson at a time, from the engineering and data trenches into the rooms where strategy gets set. I've been the product executive deciding what we build and why it matters, and the CTO who's answerable for whether it actually holds up in production. That double vantage point — builder and owner — is exactly why the current agentic AI wave gives me pause, and why I'm confident about how it plays out.
Because the numbers have started to turn, and they're pointing at the wrong culprit.
The reckoning already has a date on it
In June 2025, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 — and named the causes: escalating costs, unclear business value, and inadequate risk controls.
Read that list again. Model capability isn't on it.
And the direction of travel is only accelerating. Gartner also projects that by 2028, at least 15% of day-to-day business decisions will be made autonomously by AI agents — up from essentially zero in 2024. We are about to hand real decisions to systems we can barely explain, at the exact moment most organizations can't yet prove those systems are right.
The ROI problem is really a trust problem
The failure isn't hypothetical. MIT's 2025 State of AI in Business study — the one that landed on every board's desk this year — found that roughly 95% of enterprise generative-AI pilots delivered no measurable impact on the P&L. And the researchers were explicit: the gap wasn't explained by model quality. It came down to integration, trust, and whether these systems could be made dependable inside real workflows.
I've lived that reason from the inside. As a CTO, the model is rarely the hard part anymore. The hard part is standing in front of a board, a regulator, or your own risk committee and answering one question: can you prove this output was right — and show me exactly why? For a single model, that's difficult. For an agent , a system that chains model calls together to take an action, where step three feeds step four feeds step ten , it becomes nearly impossible unless you engineered for it from the start. Most teams didn't.
The governance gap is now measurable - and it's enormous
Here's what should focus every executive reading this. Adoption has sprinted ahead of oversight. In 2025, AI use became near-universal across enterprises, while only single-digit percentages could claim a comprehensive governance framework to sit underneath it. Two numbers make the exposure concrete:
63% — IBM's 2025 Cost of a Data Breach report found that nearly two-thirds of breached organizations either had no AI governance policy or were still building one.
48% — Grant Thornton found that almost half of boards have not set any AI governance expectations at all.
This is the real story behind the cancellations. Companies didn't deploy models that couldn't think. They deployed systems they couldn't govern, couldn't trace, and ultimately couldn't defend and when the value got questioned or the risk got real, the projects died.
What building taught me and what the survivors will do
The agentic projects that live through 2027 won't be the ones running the biggest models. They'll be the ones that can attach a name to the decision and a trace to the outcome where, when something goes wrong, you can reconstruct exactly which step broke and why.
The industry's favorite governance question — is your AI fair? — tells you the model didn't discriminate. It does not tell you the answer was right. In an agentic system, "right" isn't a single judgment; it's a chain of them, and every link has to be traceable back to the data field that produced it. The bar has moved from monitored to provable.
That gap between AI that looks governed and AI that can prove it was right — is the entire reason I built PradimeAI and Veridex. A layer that validates AI outputs against ground truth and traces every result back to the field that caused it, so that when the board, the regulator, or your own conscience says prove it, you can. We spent months mapping what's landing now - the EU AI Act, NIST, ISO 42001, SR 11-7 -back to what an agent actually has to be able to show. The uncomfortable finding: most autonomous systems being piloted today couldn't produce that record if asked.
I'm not arguing against agents. I've built enough technology to respect what real automation does for an organization drowning in manual work. I'm arguing that after twenty-six years, I've watched this exact pattern before: capability races ahead, discipline lags, and the projects that skipped the discipline are the ones that don't survive.
So before you let an agent act on your behalf, make sure you can answer the only question that has ever mattered:
Can you prove it did the right thing?
If that's the question you're sitting with, I'd value the conversation -
it's exactly what we're building to solve, and I'm looking for a small number of regulated enterprises to solve it alongside.
— Dimple Thakkar, Founder, PradimeAI
Comments