The numbers are in, and they are not flattering.
According to MIT’s State of AI in Business 2025 report, roughly 95% of generative AI pilots fail to produce measurable business impact.
Not “underperform against expectations.” Not “need more time.” Fail. Ninety-five percent.
If your organisation is one of the many that has spent the last two years running experiments, building prototypes, and presenting demos to the board (and you’re still struggling to point to outcomes that matter) you are not alone.
But you are running out of time to pretend this is a technology problem that the next model release will solve.
It isn’t. And the sooner leadership teams understand that, the sooner they can build something that actually works.
The Problem Was Never the Model
Here is the uncomfortable truth that most AI vendors won’t tell you: the gap between a working AI demo and a deployed AI system that delivers business value has almost nothing to do with artificial intelligence.
Urgent, energetic, full of genuine innovation, Zapier ran a company-wide sprint in 2023. Teams built AI-powered workflows overnight. The energy, by all accounts, was extraordinary.
And then most of it quietly disappeared.
The models worked in isolation. They struggled to survive the complexity of the organisation around them: the existing tools, the data sources, the approval flows, the human handoffs.
But get this, the AI was fine. It’s the orchestration that wasn’t.
This is the pattern playing out across enterprises everywhere right now. Projects don’t stall because the technology fails, but because no one mapped how it would actually integrate with the systems and people it needed to work with. Governance showed up as a surprise blocker in the final weeks. Data was messier in production than in the sandbox. No single person owned the outcome. Just the experiment.
The hard part of AI is not the AI. It’s the organisational infrastructure that surrounds it.
What Leaders Are Actually Being Asked to Do
As AI agents become a serious part of the enterprise conversation, the strategic demand on leadership is shifting in three important ways.
From model performance to workflow performance. The wrong question is “how accurate is the model?” The right question is “is the business outcome improving?” These are different measurements that require different governance disciplines. If your current AI reporting focuses on technical benchmarks rather than workflow-level impact, your strategy is still calibrated for the lab, not the business.
From experimentation to production ownership. The era of unlimited pilots should be over. Every initiative needs a named owner accountable for production outcomes, not for the prototype, not for the demo, not for the monthly status update. For results. That means building integration planning, security review, and data readiness into the project architecture from day one, not treating them as hurdles to clear at the end.
From enthusiasm to governance. The arrival of AI agents introduces a governance challenge that many organisations are not yet equipped to handle. When Anthropic released its Claude Cowork plugins earlier this year, markets briefly panicked, $300 billion in SaaS valuations evaporated in a fortnight, as investors feared agentic AI would render enterprise software obsolete overnight. The panic subsided. But the underlying question it exposed is entirely legitimate.
Agents can now operate inside enterprise systems, execute tasks, communicate with other agents, and produce outputs that affect real business processes.
The idea that a CIO might soon have thousands of agents running across an environment without clear visibility into what each one is doing is not hypothetical: it is the direction the technology is heading.
J&J’s CIO Jim Swanson put it simply: “I’m not going to have 10,000 agents running in the environment that I don’t know what they’re doing.” That is not conservatism. That is accountability.
The SaaS Question Nobody Is Answering Correctly
The SaaSpocalypse narrative is a distraction from a more important conversation.
Enterprise software is not dying. SaaS vendors are already embedding agents inside their own applications. The future is not “agents instead of software”, it is agents alongside software, on top of software, inside software, all operating simultaneously across a stack that was already complicated before any of this began.
What that creates is a software complexity challenge of a scale most IT roadmaps are not designed to accommodate.
More tools, more integrations, more agent-to-agent dependencies, more opportunities for technical debt to compound quietly in the background. Bristol Myers Squibb seized the moment to rebuild a critical forecasting system from scratch using AI, cutting errors by 50%.
That kind of clean-sheet thinking is available to organisations willing to take it. But it requires architectural courage, not just AI enthusiasm.
The leaders who will navigate this well are not the ones asking “should we be using agents?” They are asking “what is our agent governance model, how do we discover and orchestrate agent work, and how do we ensure humans are reviewing outputs at the right checkpoints without becoming a bottleneck?”
A Framework for What Comes Next
Drawing on the hard-won lessons from organisations that have pushed past the pilot stage, there are four disciplines that separate AI strategies that compound in value from those that quietly collapse.
Start with the workflow, not the model. Before evaluating any AI tool or agent, map the end-to-end workflow it needs to operate in. Where are the manual handoffs? Where does data move between systems? Where are the approval dependencies? The orchestration complexity will tell you more about project risk than any model benchmark.
Build governance before you need it. Security reviews, compliance checks, and data governance should be architectural inputs, not late-stage obstacles. Every week a project spends waiting on a security review that could have been scoped at the start is a week of momentum lost and a signal to the organisation that AI is hard to operationalise.
Know when to stop. Some AI initiatives should be killed, and killing them is a strategic decision, not a failure. If a project has been perpetually “almost ready for production” across multiple planning cycles without a clear and resolving blocker, it is consuming resources that belong elsewhere. The speed at which AI capabilities are advancing means that a project designed twelve months ago may be fundamentally obsolete: not because the team failed, but because the landscape moved. Evaluate your portfolio with that reality in mind.
Treat the human layer as the architecture. AI agents will increasingly handle the execution layer, the tasks, the tickets, the data processing. Human employees are shifting toward governance: reviewing agent outputs, correcting agent behaviour, making judgment calls that agents cannot.
Designing your operating model around that shift (not retrofitting it after agents are already deployed) is the difference between an agentic capability that scales and one that creates chaos.
The Moment You’re Actually In
Enterprise leaders are sitting at an inflection point that is easy to misread in both directions. The hype side says everything is about to change overnight: SaaS is dead, agents will run your business, transformation is one deployment away. The cynical side says this is another technology cycle, the failure rates prove it doesn’t work, and the sensible move is to wait.
Both are wrong.
What the evidence actually shows is that AI delivers real, material outcomes in the hands of organisations that approach it with the same rigour they bring to any significant business transformation, clear ownership, integrated design, governance built in, and a willingness to stop things that aren’t working.
Bristol Myers Squibb didn’t get a 50% reduction in forecasting errors by running an inspired pilot. They rebuilt a system from scratch with a clean brief and a production mindset.
That is the posture the moment requires. Not more experiments. More production.
The technology is ready for the business. The question is whether the business is ready for the technology.
Teraflow.ai helps organisations move from AI experimentation to AI at scale. We design the orchestration infrastructure, governance models, and production workflows that turn promising pilots into measurable business outcomes. Talk to us.
