The Business Case for Less AI

A small-business owner I know recently began investing in AI tools for his company.

Most customers found the business through Instagram. They wanted faster responses, clearer information, and reliable service. None had asked for an AI assistant. AI was only one possible way to improve that experience.

But he had started with the wrong question. He was asking, "How can I add AI?" before asking, "Which part of my business needs to improve?"

That distinction matters.

The problem is not that businesses are adopting AI. The problem is that many treat AI adoption as a company-wide objective rather than identifying the specific division, workflow, or bottleneck where automation can create measurable value.

AI does not need to be everywhere. It needs to be useful somewhere.

Start With the Business, Not the Technology

Every business is a collection of different activities. Sales, customer service, finance, marketing, operations, procurement, and management do not have the same problems. They also do not offer the same opportunities for automation.

One division may spend hundreds of hours processing repetitive documents. Another may depend on judgment, trust, or personal relationships. A customer-service team may benefit from faster retrieval of approved information. An account-management team may lose value if personal communication is replaced with generic automation.

The correct question is not:

Where can we use AI?

It is:

Which business process is creating a measurable constraint, and can AI improve it better than a simpler alternative?

That question changes the sequence of investment.

Instead of buying tools and searching for uses afterward, the organization first studies how work is performed. It identifies delays, errors, duplicated effort, capacity limits, customer friction, or lost revenue. Only then does it decide whether AI is the right intervention.

Sometimes it will be. Sometimes a clearer policy, a shorter form, better data, conventional software, or removal of an unnecessary approval step will solve the problem more cheaply.

Optimize the System, Not the Step

Operations management has long distinguished between improving one activity and improving the performance of the whole system.

A faster individual step does not necessarily increase total output. It may simply move work more quickly toward the next delay. A highly capable AI model applied to a non-constraint may produce an impressive demonstration while leaving the business no better off.

AI should therefore be evaluated by its effect on the end-to-end process. Did it improve throughput, quality, cycle time, operating cost, or customer experience? Or did it make one activity look more efficient while shifting cost and work somewhere else?

This system-level view is important because local efficiency is easy to demonstrate. Business value is harder.

Automation Magnifies the Process It Receives

This is not a new management lesson.

In The Road Ahead, published in 1995, Bill Gates wrote:

"The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency."

AI makes this principle more important, not less. It can process information, generate content, classify requests, recommend actions, and coordinate work at greater speed.

But speed does not correct a poorly designed process. It can make the waste move faster.

Elon Musk reached a similar conclusion after Tesla struggled to automate Model 3 production. In an April 2018 statement, he wrote:

"Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated."

Tesla's experience does not mean automation is a mistake. Tesla remains one of the world's most automated manufacturers.

The lesson is narrower: automation should be applied where it improves the production system, not where it adds complexity simply because automation is available.

The same discipline should guide AI.

Select One Area and Establish the Baseline

Before applying AI, a business should define the current performance of the selected process.

Without a baseline, the organization cannot know whether AI created improvement. Prompts, generated documents, user seats, and completed tasks are activity measures. They do not establish business value.

The organization should then introduce AI within a defined boundary.

A customer-service pilot might cover one category of repetitive questions. A finance pilot might classify invoices before human approval. A sales pilot might summarize meetings without allowing the system to contact customers independently.

The purpose of the pilot is not to prove that AI works in general. It is to test whether a specific application improves a specific business outcome.

Measure the Result in Economic Terms

Chamath Palihapitiya has framed the unresolved AI investment question in direct financial terms. During a May 2026 All-In Podcast discussion, he said companies eventually need to demonstrate:

"I spent X and I made Y, where Y is now greater than X."

That is the basic discipline missing from many AI programs.

The calculation should include more than the subscription price or model bill. Full AI cost includes subscriptions, usage, integration, testing, human review, rework, controls, and the cost of failures or incidents.

The return may appear as lower operating cost, shorter cycle time, additional capacity, fewer errors, better customer retention, incremental gross profit, or avoided loss.

But the organization should define the expected benefit before the pilot begins.

A simple ROI calculation is:

ROI = (Measured financial benefit over the period - Full AI cost over the period) / Full AI cost over the period

Not every benefit will be immediately financial. A business may run a pilot to learn, reduce future risk, or build a capability it expects to need.

That can be legitimate. But the learning objective, budget, duration, and stopping condition should still be explicit.

The scarcest resource is often not model capability. It is management attention. Every AI initiative competes with other operational improvements for the same budget, talent, and organizational focus.

When the Economics Change Faster Than the Roadmap

Recent METR research estimates that the 50%-success time horizon on its mostly software and research tasks has doubled roughly every six to seven months.

METR defines this as the amount of serial human labor a model can replace at a 50% success rate, not how long the model can work independently. The estimate is imprecise, varies sharply across domains, and does not mean that every task below the threshold can be safely delegated.

For a growing set of tasks, the question is no longer only whether AI can perform the work. It is whether AI can perform it at a lower total cost while meeting the required standards for quality, reliability, and control.

That calculation will change. Model prices, capabilities, and implementation costs continue to move. A process that fails an ROI test today may become attractive six months from now. A process that looks valuable today may lose its advantage when the same capability becomes widely available at a lower cost.

That is why every AI business case should include an expiration or revalidation date.

Adoption Is Not the Same as Value

The pressure to adopt AI is real. Leaders do not want competitors to learn faster or operate more efficiently. Waiting indefinitely has its own cost.

But adoption is not the same as value.

In PwC's 2026 Global CEO Survey, only 12% of CEOs said AI had delivered both cost and revenue benefits. Fifty-six percent reported no significant financial benefit from AI.

These are self-reported results, not independently measured returns. The finding does not prove that AI lacks value. It shows that access to AI does not automatically produce it.

The missing step is often operational selection: choosing where AI fits, defining the expected improvement, and measuring the result against the previous process.

AI Should Earn Its Place

An AI strategy should not be a declaration that the whole company is becoming AI-driven. It should be a portfolio of specific applications, each connected to a real business need.

Some applications will reduce cost. Some will increase capacity. Some will improve quality or customer experience. Others will fail to justify their expense.

That is normal.

The purpose of measurement is not to guarantee that every experiment succeeds. It is to prevent experimentation from becoming permanent spending without demonstrated value.

The central management question is simple:

Which part of the business can AI improve, and what evidence will prove that it did?

If leadership cannot name the process, the baseline, the expected benefit, and the evidence that would prove success, the organization is not ready to automate.

It is ready to experiment — with a budget, a deadline, and a stopping condition.

References