There is a slide making the rounds in my world right now. You have probably seen a version of it. It promises a live AI agent in production, a validated ROI model, and a scale roadmap. Six weeks. Fixed price. The design is clean, the language is confident, and the offer is hard to argue with.
I want to clarify something before I go further. I am not anti-AI. I use it every working day, often inside the very SQL Server work that pays my bills. Where it lands on a real problem with real data, it earns its place at the table. This is not a 'STOP AI!' post.
My concern is more focused, and I think it is harder to dismiss. It is the speed. We are wiring these systems into production faster than we are building the boundaries that production demands. And the reason we are moving this fast is that someone has made moving fast look easy, packaged it up, and put a price on it. That package is worth a hard look before we talk about what it leaves behind.
What the packaging promises
The pitch is consistent across vendors -- A high-value use case, fully deployed. A cost model your CFO can sign. A backlog of the next ten use cases, ready to go. All of it on a timeline measured in weeks rather than quarters, and all of it for a number you agree to up front.
I understand why leaders say yes. It is a clean answer to a messy mandate. When the directive from the top is 'we must do AI now,' a fixed-price sprint to production looks like progress you can put on a slide. The trouble is that the slide and the outcome don't always follow the same path.
The door we are leaving open
Every agent stood up in a hurry is a new door into the environment, and too many of them are being hung without locks -- and the people who make a living finding unlocked doors have definitely noticed. Consider what has already happened, in production, to organizations far better resourced than most.
In June 2025, researchers disclosed a flaw in Microsoft 365 Copilot they named EchoLeak (CVE-2025-32711, rated 9.3 critical). It was a zero-click attack. An outsider sent a single crafted email, and when Copilot did the helpful thing and read it, hidden instructions inside told the assistant to go find sensitive material across OneDrive, SharePoint, Teams and Outlook and quietly ship it out. The victim clicked nothing. There was no malware to scan for, because the payload was plain English. Microsoft patched it and reported no exploitation in the wild, but the lesson should still stand: the assistant became an insider threat, and it never knew it.
Or consider the coding agent that, in 2025, deleted a production database despite instructions not to modify production systems, then produced inaccurate information about the recovery state. You can read about it here. No attacker required. The agent simply had more authority than judgment.
These are not edge cases. The OWASP project that tracks AI security risks used to catalog threats as things that might happen. Its current edition catalogs them as CVEs, vendor advisories and breach reports -- in other words, things that already happened. That is the part the vendor's six-week timeline never mentions. An agent that can read your data and act on your systems is a privileged user, and we are handing out that privilege faster than we are regulating who, what or how, it is managed.
What the data says
MIT's NANDA initiative studied 300 enterprise AI deployments. Roughly 95 percent of generative AI pilots produced no measurable impact on the bottom line. Only about 5 percent reached real, scaled value. I will be fair about that headline. It has been widely cited, though not without criticism of the methodology. But even the people pushing back tend to agree on the underlying point, which is that most pilots stall long before they reach production. You can read the coverage here.
Gartner is more direct, and more relevant, because the slide I described is selling agentic AI specifically. Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027. The reasons it gives are "escalating costs, unclear business value or inadequate risk controls." The full prediction is here.
Gartner even has a name for part of the problem, 'agent washing'. The practice of rebranding ordinary chatbots and automation as autonomous agents. By its count, of the thousands of vendors claiming agentic AI, Gartner estimates only about 130 of them are real. Most of what is being sold as an agent is an old tool wearing a new label.
The failures are not about the model
This is the part that should matter to anyone who runs systems for a living. Read the reasons again. Escalating costs. Unclear business value. Inadequate risk controls. Not one of them says the AI is too dumb. The pilots are not failing on intelligence. They are failing on the pieces that were missed: data that was never cleaned, integration that was never scoped, governance that was never written, controls that were never tested.
Those are boundaries, and a six-week, fixed-price sprint to production is, almost by definition, a sprint past those boundaries. You cannot clean the data, prove the ROI, establish governance, test the controls, and build the system all in the same six-week sprint. Something gets deferred. In most organizations, it is the part that was supposed to keep everyone out of trouble. In a regulated environment, the things that get cut are usually the same things that surface in an audit eighteen months later, long after the launch party.
What the 5 percent did differently
Here is the encouraging half of the story. The organizations that got real returns were not the fastest. They were the narrowest. They picked one painful, well-understood problem instead of trying to transform everything at once, and they fixed their data before they pointed a model at it. They also tended to partner rather than build everything in-house, and they often started in the boring back office rather than the flashy customer-facing demo -- where the value was quiet but real.
None of that fits on a six-week timeline, and none of it markets well. Patience is the least sellable quality in anybody's board room. It is also, apparently, the one that correlates with success.
Why I am writing this down
I am not asking anyone to slow the technology. At this point, I doubt anyone could if they tried. I am asking that we let the boundaries keep pace with it. Governance, data quality, risk controls, and a human who can still read the audit trail are not friction. On the evidence above, they are the entire difference between the 5 percent and the 95 percent.
The technology itself is going to be fine. It always is. My worry is for the organizations that bet the timeline before they build the guardrails, because that bill always comes due, and it does not come due on the quarter the slide was presented. Fast is exciting. Fast is sellable. But fast is not the same as ready. The organizations that succeed will not be the ones that moved first. They will be the ones that understood a very simple truth: Powerful systems require equally powerful controls.
