Saturdays are for thinking. They are for stepping back from the weekly rush and asking a basic question: are we planning the work, or are we just budgeting for optimism? Today, let’s take a look at the question: Are your agents are “assistants” next to the work, systems embedded in the work – and where do you want to be? This is a fundamental question in identifying your approach. Your answer to this question may be subtly different from your colleagues, or the person signing off the budget for the work.
There are plenty of great ideas in enterprise AI, and the enthusiasm can result in poor planning dressed up as ambition. Organisations approve pilots, fund proofs of concept, and announce targets before they have addressed the plumbing: data engineering, governance, integration, ownership, and measurement. The result is familiar. The demo works. The business case does not.
That is the planning fallacy in action. Teams underestimate cost, time, and complexity. They overestimate readiness, adoption, and speed to value. AI without a plan is not strategy. It is a fallacy.
Most pilots are disconnected from the actual workflow, the source systems, and the economics of the business. To bridge this gap, organisations need an architecture-first approach and a practical path toward Agentic AI for Enterprise.
Without data fluency, teams cannot assess whether an AI output is useful, risky, incomplete, or impossible to put into operation. As I have discussed in our guide on bridging the AI execution gap, strong data foundations are what make or break enterprise AI.
Sources:
A 10-Year Plan Versus 34 Years of Delay
Some of you will know that I do volunteer work on a Sundays, which brings me into contact with people from many different walks of life. A 90-year-old volunteer recently explained her 10-year plan to me with more clarity than many enterprise AI programs. The point is not age here; the point is precision. She knew what mattered, what was realistic, and what she wanted to contribute over time. Now compare that with HS2, Britain’s latest effort to be the most self-sabotaging nation on Earth. The UK high-speed rail project was first formally proposed in 2009 and then spent years in redesigns, cost escalation, scope changes, and political reversals. The estimated cost had risen sharply to an eye-watering £102.7 billion pounds or $138,028,803,081 in US dollars. HS2 was announced in 2009 and will not carry passengers until the mid-2030s, and it won’t be ready until 2043 – that’s 34 years! The National Audit Office reported repeated issues with cost control, schedule certainty, and scope management. It is clear nobody in Government knows their Voltaire, because the work continues with a Panglossian attitude. That contrast between the 90-year-old volunteer and HS2 matters. Unfortunately, enterprise AI often follows the HS2 pattern. Leaders approve a big vision, and assume execution will sort itself out at the expense of long hours donated by employees. They will then discover that the hard parts were ignored at the start, such as duplicate storage and compute costs due to overlapping data sources. As psychologists Daniel Kahneman and Amos Tverskey put it, the planning fallacy is the tendency to underestimate “the costs, completion times, and risks of future actions while overestimating the benefits.” It is a standard enterprise delivery problem, rather than some abstract behavioural science concept.Generative AI pilots need measurable P&L impact
An investigation regarding generative AI report from MIT and NANDA indicates that approximately 95% of generative AI pilots show zero measurable P&L impact. While nearly 80% of organisations have experimented with tools like ChatGPT or Copilot, only about 5% have deployed AI systems that deliver material value to the bottom line.
Most pilots are disconnected from the actual workflow, the source systems, and the economics of the business. To bridge this gap, organisations need an architecture-first approach and a practical path toward Agentic AI for Enterprise.
What the Planning Fallacy Looks Like in Enterprise AI
The planning fallacy is the tendency to assume a project will take less time, cost less money, and face fewer problems than the evidence suggests. Optimism bias is the related habit of believing your organisation will avoid the issues that affected everyone else. It’s important to reach an agreement about where your agents are right now, or what your starting point will be. Will they be “assistants” next to the work or systems embedded in the work? The distinction will not always be discerned, and it could lead to poor planning which usually shows up in five ways in enterprise AI:- The pilot is approved before the data is usable.
- The model is selected before the process is mapped.
- The business case is written before the controls are defined.
- The budget covers experimentation but not integration.
- The timeline assumes adoption without change management.
The House Metaphor: You Need an Architect Who Understands the Plumbing
A useful way to think about enterprise AI is to think about building a home. Most buyers do not want to inspect every pipe, valve, and junction. They want to know the house is safe, functional, and built to last. That is why you need an architect who understands the plumbing. In AI terms, that plumbing is data engineering. If the underlying pipes are wrong, the house looks finished until the taps stop working. If the data pipelines are weak, undocumented, siloed, or poor quality, the AI system may look polished in a workshop and fail in production. That is where data fluency matters. Data fluency is not a training slogan. It is the operational ability to understand what data exists, where it comes from, how it should be used, and what business decision it supports.
Without data fluency, teams cannot assess whether an AI output is useful, risky, incomplete, or impossible to put into operation. As I have discussed in our guide on bridging the AI execution gap, strong data foundations are what make or break enterprise AI.
SatNav Versus Map: Why Rigid Planning Fails
Many organisations still approach AI with the equivalent of a printed road atlas. The route is fixed, the assumptions are locked, and any deviation looks like failure. That is the wrong planning model. AI delivery works better like a SatNav. You still need a destination. You still need constraints, checkpoints, and governance. But the route adjusts as road conditions change. New data issues appear. Business priorities shift. Integration takes longer than expected. Regulation changes. Teams learn what users will actually adopt. A map says, “Follow this route exactly.” A SatNav says, “Recalculate based on reality.” That is the difference between rigid planning and agile planning. The goal is not to abandon structure. The goal is to replace fantasy schedules with evidence-based adaptation.From “Chat” to “Work”
The industry is moving past the novelty of chat interfaces. Enterprise impact is found in the transition from conversational AI to agentic AI. In a “chat” model, a human asks a question and receives an answer. The human still does the work. In an “agentic” model, the AI is granted the ability to use tools, access specific datasets, and perform multi-step actions within defined guardrails. An agent reconciles that invoice against the contract, checks the warehouse receipt, and prepares the payment in the ERP. This is where the 5% of successful companies are finding ROI. They are focusing on back-office processes such as finance, operations, and compliance, where process efficiency is measurable and governance is non-negotiable. It really does free people up to do more fulfilling work.The Strategy Pathway
Moving an AI pilot into production requires less grand planning and more honest planning. The practical path usually comes down to three areas:- Architecture Before Automation: Define the operating model, source systems, controls, and data flows before expanding the use case.
- Integration Before Scale: Connect AI to CRM, ERP, document stores, and internal databases in a governed way. Integration and change management usually cost more than the model.
- Measurement Before Celebration: If you cannot trace the work to reduced cycle time, improved margin, lower vendor spend, or stronger compliance, the pilot is not delivering enterprise value.
Why This Matters Today
The current learning gap is about planning discipline. Tools are not adapted to the organisation, and the organisation is not adapted to the tools. The issues show up when pilots stall, leading to budget cuts. In the meantime, shadow AI spreads. By creating a centralised, governed, and integrated AI strategy, organisations reduce the risk that employees use consumer tools to solve enterprise problems with no oversight. They create a safer environment for experimentation and a clearer route to production value.Your Next Move
If your AI roadmap is starting to look more like HS2 than a credible business program, the answer is not more hype. The answer is better architecture. Like a SatNav, organsations need someone to tell you when you are going the wrong way – the earlier the better! I help organisations design AI strategies that are grounded in delivery reality. This is potentially the ‘boring’ items that people don’t want to think about, such as data foundations, governance, measurable outcomes. I think about these things so I can help build systems that fit the business. If you need an architect who understands the plumbing, or an AI ‘SatNav’, this is the moment to get in touch. Start with our guide to Agentic AI for Enterprise, then contact me for practical architectural guidance on how to move from AI ambition to AI execution.
Sources:
- MIT NANDA, The GenAI Divide: State of AI in Business 2025.
- Fortune, “95% of generative AI pilots are failing to show P&L impact.”
- National Audit Office, HS2 Euston and related HS2 reporting.
- Daniel Kahneman and Amos Tversky, Intuitive Prediction.
- Morgan Housel, The Psychology of Money
