The most expensive AI mistake: automating the unnecessary to perfection

In brief: Deloitte's autumn-2025 framework for choosing the right AI-agent use cases is genuinely useful, especially its "no" list of where agents don't belong. But a year of live projects reveals three blind spots: implementation — not selection — is the real risk; "agentic" vendors are often rebadged automation; and treating the complexity ladder as a to-do list traps you in trivial wins. The framework's core still holds — just add three 2026 warnings before you build.

Last autumn, Deloitte put out a short, practical piece on how companies should choose the right application area for AI agents (the study's title: Unlocking the right agentic AI use cases). The approach is telling: it doesn't promise that agents are good for everything, but rather that the hard part is the selection. In 2026, that's a relief — because the bulk of the market is preoccupied with what AI can do, when the fate of the investment actually hinges on whether you know what you're allowed to entrust to it.

The framework is good, but there are a few points where 2026's operational experience has already overwritten the 2025 theory. Let's look first at what it offers, then at where it's worth going beyond it.

What Deloitte gets right: the "no" list

Most consulting AI material is about what agents are good for. Deloitte's is valuable precisely because it takes the reverse seriously. It sets up a diagnostic filter of eight criteria, and for each one gives an "ideal" and a "not ideal" example. The point isn't the eight lines, but the underlying logic: an agent is justified when the task calls for reasoning, context-dependent decisions, and multi-step, goal-driven action. If it doesn't, a cheaper and more reliable tool will do.

Three of the eight dividing lines are the most cutting; they're worth highlighting:

An agent is not an analyst. Deloitte's explicit example: an agent can't answer "why is our revenue stagnating?". An agent is meant for goal-driven action, not causal explanation. This is a more important distinction than it first appears, because a great deal of executive expectation falls into exactly this trap: people expect explanations and connections from an agent, when the agent executes.

Where plain automation wins. If the task is sending a trigger-based, pre-written template message or entering data into a CRM, there's no point in building an agent for it. For a workflow that consists of predetermined, well-defined steps, simple automation runs more cheaply and more predictably than an agent. In the industry, this is becoming a cliché: where the process consists of fixed steps, traditional automation delivers a better result at lower complexity.

Where it's brilliant, on the other hand. Where you have to reason from unstructured data, where context matters, and where you have to navigate between multiple systems (ERP, OCR, external portals) with decisions. Complex invoice processing or IT incident management are Deloitte's textbook cases.

To this, the firm pairs a prioritization matrix. It supplements the classic "impact versus feasibility" picture with a third dimension, the Differentiability Index: how critical AI is to the given task, and how much of an advantage it provides that a competitor can't simply buy tomorrow as off-the-shelf software. This is a genuinely useful angle, because many companies pour money into an agent that anyone will be able to purchase as a finished product within six months.

So much for the framework, and its logic holds up. But the study was written in the autumn of 2025, and since then some of the first major agent projects have run their course, and the initial hype has subsided too. Seen from this vantage point, the model's blind spots are already visible. Three of them are worth highlighting.

Blind spot 1: selection isn't the biggest risk. Implementation is.

The implicit assumption of the Deloitte material is that once you've chosen the use case well, the hard part is behind you. The 2026 data show otherwise. According to Gartner's forecast, more than 40% of agentic AI projects will be scrapped by the end of 2027 — and according to the Gartner analyst, not because the technology doesn't work, but because the projects are early-stage, hype-driven experiments with poor strategy and missing controls.

In other words, the pitfall typically isn't at the use-case selection stage, but afterward. Several analyses conclude that projects fail because companies automate broken processes instead of fixing them first. Deloitte's filter tells you which process is suitable, but it doesn't protect you from a suitable process being implemented with a poor data backbone and no controls. McKinsey calls this the "gen AI paradox": nearly eight in ten companies have adopted some form of generative AI, and roughly the same number report that it hasn't meaningfully moved profit. The paradox, then, is that widespread adoption and actual business benefit have come apart: lots of experimentation, few measurable results.

So the Deloitte framework is worth supplementing with a question zero before we agentize: is this process even in good shape to begin with? A bad process will only get bad faster with an agent.

Blind spot 2: "agentizable" doesn't equal "a real agent"

Deloitte assumes that if a vendor offers an agent, it's an agent. In 2026, that's a dangerous assumption. Gartner has given the phenomenon its own name: agent washing — when a vendor rebrands an existing chatbot or automation as an agent. Their estimate is sobering: of the several thousand vendors marketing themselves as "agentic," roughly 130 provide genuine autonomous capability.

This directly affects the Differentiability Index. Deloitte's matrix measures how unique and hard-to-copy a use case is. But if the market is full of off-the-shelf automation disguised as "agents," then the question of competitive advantage is joined by a question of credibility: is what you're buying actually an agent, or just plain automation with an agentic price tag? The framework needs to be topped off with this filter when deciding between in-house development and buying a finished product.

Blind spot 3: the complexity ladder isn't a strategy in itself

Deloitte gives three complexity levels: low (e.g., CV screening — a single domain, rule-based), medium (e.g., legal review of bank guarantees, with three collaborating sub-agents), and high (e.g., end-to-end invoice processing, 5+ agents, unstructured data, continuous learning). The advice is logical: start with quick wins, build trust, then move up.

There's nothing wrong with that — right up until "start small" turns into the "get stuck at the trivial" trap. McKinsey's observation is sharp: the first wave (copilots, chatbots, summaries) delivered hard-to-demonstrate benefit precisely because it automated the easy, visible surface of knowledge work, not the process itself. A low-complexity quick win is good for building trust, but ROI mostly comes from redesigning the process, not from an agent dragged onto an existing one.

In other words, the complexity ladder shouldn't be read as a linear timetable ("first all the low ones, then the medium ones"). Better to read it as a portfolio: a few quick wins for learning and credibility, but in parallel at least one highly differentiated process where the actual transformation happens. Anyone who only fiddles at the bottom of the ladder is a sure victim of the gen AI paradox.

What remains: the core of the framework holds in 2026 too

The criticism doesn't mean Deloitte is wrong. Quite the opposite: its main message holds up this year as well — in fact, it's more urgent than ever. Don't adopt an agent because it's new and trendy. Define the problem and the expected outcome, justify why the agent is better than traditional automation, and only dive in if the task genuinely makes use of the agent's unique strengths: reasoning, adaptation, learning.

To this, 2026 adds three warnings. First: selecting the use case is necessary but not sufficient, because implementation (data, process, controls) is decided where the Deloitte framework can no longer see. Second: verify that what you're buying as an "agent" really is one. Third: the quick win is the entry point, not the goal; the return lies in rethinking the process.

The best line on this isn't even from 2026, but from Henry Ford, quoted in a related Deloitte piece: many people labor to do better something that shouldn't be done at all. Doing a useless thing better is not progress. The most expensive mistake in agentic AI in 2026 is exactly this: automating something to perfection that shouldn't exist.

Sources

Deloitte: Unlocking the right agentic AI use cases (September 2025)

Deloitte Insights: Agentic AI strategy (Tech Trends 2026)

Gartner / MarTech: 40% of agentic AI projects will fail (2026) — source for the 40% attrition figure and the "agent washing" data

After McKinsey: Why 2026 Is the Year the Pilot Phase Has to End — the "gen AI paradox"

Cloud Latitude: Why agentic AI projects fail and how to scale

This analysis is based on the cited sources; the market forecasts are estimates, not guarantees.