The Rate-Limiting Step That Will Slow AI

The rollout of AI has a built-in, organic rate-limiting step that will naturally throttle its rollout, dramatically increase its cost of use, and, in the end, make that rollout far more thoughtful.

The rate-limiting step that will slow AI.

Everyone – from companies to cities to countries to individuals – is going all in on AI at a breakneck pace. That pace, if it accelerates, is what is going to slow down the release of AI, increase its cost of rollout, and lead to more thoughtful AI applications.

I studied chemical thermodynamics. It should have just been called thermodynamics, but I am sure my school added the chemical part for marketing reasons, since there is only thermodynamics. Then again, there is physical chemistry, and even there I think thermodynamics would do, though many would disagree.

In chemistry, we have the concept of the rate-limiting step. I will simplify this to make the main point. To produce a compound, the reaction must move through a sequence of steps — call them Step A, then B, then C, then D. The product only comes out the far end once all four have happened, in order.

Now imagine I am in a lab, trying to run this reaction faster. This would require imagination since I avoided lab work at all costs. Steps B, C and D are quick. Step A is painfully slow. So, I heat the vat. I stir harder. I pour in more raw material. B, C and D race along. And still, barely any product comes out the other side.

Welcome to the rate-limiting step. The whole reaction can only go as fast as its slowest step. It does not matter how fast B, C and D are — everything is gated by A. Speed up all of them except A, and you have sped up nothing.

Again, I am simplifying this, but the point is you cannot make the reaction go faster no matter how much you heat, stir, or feed it, because in operations terms, A is the bottleneck.

Even if I bought my ingredients at the cheapest possible prices and built the world’s largest mixing chambers, if I did not speed up Step A, I would not produce the compound any faster.

And now assume Step A is the one you cannot speed up. The stubborn one. The step no amount of money, heat or pressure will hurry — the one the whole world is stuck on. Hold that thought.

So, what is AI’s rate-limiting step? What is the one thing that will not allow you to create more value with greater use of AI, no matter how many AI tokens, tools, and budgets — or how much effort — you throw at it?

“Humans mess up everything,” to paraphrase Quark in Star Trek: Deep Space Nine

Let’s take the average user of AI in a professional setting, like a consultant, lawyer or corporate employee. They are almost giddy with excitement. There is now a tool that produces a reasonable facsimile of their work. Everyone assumes that today they do not need to master the art of proper evidence-based, deductive, inductive and/or hypothesis-based research.

In their minds, why study basic long division, trigonometry or calculus when you can use a scientific calculator? Similarly, why bother doing research, writing a report or doing any foundational work when their AI tool of choice will do it all for them?

Why stress about learning a new sector or company over the weekend? AI will tell them what to focus on.

Why worry about turning down work in a new area? AI will make you an expert.

Why stress about summarizing years of financial statements? AI will do it in minutes.

Why worry about writing a thoughtful business case, memo report or storyboard? AI will do it.

Why worry about anything, like even responding to emails? AI will do it.

Why bother getting out of bed? A string of agents will do everything for you.

There has been an explosion of corporate, consulting, legal, audit, financial, R&D, etc., AI content/reports/recommendations/actions produced by professionals.

It’s such a big tsunami that it’s the envy of surfers worldwide. Now, that would be a cool video game.

So, what’s the problem?

There are two problems.

Problem when you are the expert

Let’s assume you are one of the world’s, or your company’s, expert on the pricing of soft drinks in unstable emerging economies. It’s for a different piece, but there is a difference between having some/lots of experience in a field and being an expert. It does matter.

It typically takes you three weeks to offer your guidance on whether your company should change its pricing in a country. You believe, as inconclusive as the evidence may be, that if you just had more exposure to the heads of EMEA and LATAM, you could have more career-promotion options. Alas, you serve the Eastern European region, which is only experiencing declining consumption due to a hollowing out of the population.

You know the EMEA and LATAM regions need pricing experts, so you decide to offer your services. And the reason you can offer your services is because you can use AI to do most of the work to complete the EMEA and LATAM pricing recommendations.

Over the three-week period, you think your value to the company went up 3X, because you are now producing three reports versus one.

Why stop there? You probably used AI to produce the three reports in 1.5 weeks. So, if the time halves, your value went up 6X.

There is one part that has not changed. The speed at which you can analyze, check and modify the reports has not changed. If it takes you two weeks to analyze and check a report that you wrote, does it take one week, two weeks or four weeks to analyze a report you did not write, where you don’t know the assumptions and input data?

You cannot speed up the checking process, simply because you cannot work faster than you can.

And using one AI tool to check a report that another AI tool wrote simply compounds the checking effort you need to make. It does not solve the problem.

It’s even worse if you use the same AI tool to check the report said AI tool wrote. Which is the most common checking process.

You can produce as many AI-driven reports as you would like in the field where you are an expert, but you cannot improve the rate at which you can personally check the work.

There is a caveat, and some hope, here. An AI system like Michael AI is singularly built for a specialized process, and will understand what is needed to produce high-quality work to solve a business problem to raise the value of a company and enhance your career. It is a dual objective. That is because it is designed by experts for a specific task. Those experts bring their decades of expertise, as opposed to just experience.

Unfortunately, practically no one uses specialized AI tools. They use general-purpose AI tools like Gemini, which know very little about a lot and are not trained on only expert material. They are trained on everything.

Yet, even a specialist AI tool needs to know your unique context, because you will be executing the recommendation, and the recommendation needs to take your limitations and strengths into consideration. Michael AI singularly tells you if a project could hurt your career, and what the mitigation steps are.

Problem when you are not the expert

Let’s take the example above but change it. Now you are producing three reports versus the one, but the additional two are in fields you know little about. Let’s assume you do not see a future in pricing or are not excited by a long-term career in pricing soft drinks, which is essentially trading futures on diabetes. So, you focus on internal strategy and AI within the soft drink conglomerate.

Knowing very little about either subject, you are truly dazzled by the AI reports. They sound and look so polished. They look like impressive reports. You learn so much by reading them. The arguments are logical and persuasive.

Yet, they are just facsimiles. They look like the real thing, but they are not.

Let’s go even further and assume you used a specialist AI model.

If you know little about strategy or AI, how can you assess the quality of the AI-driven reports?

Further, how can you even know that the reports must take into consideration your ability to do the work — and tell you if the recommendations you are making may be good for the company but bad for your career?

Assuming you are asked to implement your strategy or AI recommendation, how will you ever do it?

AI cannot make you an expert overnight. Even if it could, are you going to refuse all meetings in person so you can run everything through AI before responding? The remnants of the COVID era mean some will certainly try this.

Here is the rate-limiting step.

AI is code. It’s software and, somewhat, hardware. When you roll out code, or the output of the code, either without checking it or without the ability — the will, capacity, capability or budget — to check the code/output, you allow mistakes in.

I could have said you make mistakes, but I am trying to be diplomatic.

Some mistakes are small, but I am sure very significant to someone. Like refusing a refund of $100 to someone making $300 a week. Or dinging the credit score of someone through no fault of their own.

Some are very large, like the incorrect precedents in a legal brief, which could change a person’s life or a company’s balance sheet.

Some, like the UK Post Office scandal, led to destroyed careers, destroyed families, imprisonment and suicide.

Others, like using AI to distinguish between armed combatants and civilians, will have horrifying consequences. Have we simultaneously increased the budgets and numbers of the military lawyers vetting the surge in AI-identified combatants? And that’s assuming that country even has this step in place. Most will not, because AI is meant to cut labor costs.

The rate-limiting step is your ability to check the work. The real-world consequence that will slow AI, raise its cost, and lead to a more thoughtful rollout is this: the very real scandals that will come from the misuse of AI.

The seeds for those scandals have already been planted. We just must wait for them to bloom.

Kris, my colleague, on the other hand has great examples where maturity and proven technologies where not implemented well or at all, leading to scandals. That is for a different article.

This is not the fault of AI

Imagine there is a race across some treacherous terrain. Teams spend years mastering the terrain and countless hours modifying their cars to endure it. Engine modifications are needed, the entire suspension must be custom-built, drivers spend months conditioning themselves, and an entire support staff of caterers, nutritionists and emergency rescue staff are on hand.

Peter, a looksmaxxing influencer, sees this once-in-a-lifetime opportunity. To promote the race, the organizers of this year’s edition are going to invite fans with the largest followings to join. Peter sees his chance. He already owns a good car for off-road racing. He does a little offroad driving. He flies in the weekend before, sees the local sights, hits the gym, hydrates and gets to the race on Monday. To save time, he pre-bought everything he needed on Amazon and drop-shipped it to the location, where an assistant he found on Upwork set up everything.

The outcome here is predictable. Peter is going to have a very bad experience, get seriously hurt or become a fatality.

Peter misused the technologies at his disposal.

A commercial off-road vehicle is not the same as one designed for that specific course.

Finding a support team on Upwork is not the same as finding a race crew.

The convenience of Amazon is not the same as the inconvenience of tested supplies.

If you are using general-purpose commercial AI to do anything other than general-purpose, directionally reasonable work, you are misusing the tool.

If you are taking that output from the commercial tool and using it for a highly technical recommendation, you are misusing AI.

If you are not using AI tools developed by domain specialists in your field, then you are misusing AI.

And here is the final argument. And the one that matters. When you produce three reports, versus the one, from a specialist AI tool — one that takes your abilities into consideration and is built by domain experts rather than those with just experience — you can create value, because the base quality is higher. The higher the base, the far less you have to do.

You can even submit that report as a discussion point with your team versus as a finished product.

If you do not do that, all your time savings are spent on QA, and some bad reports will get through. Setting the entire rate-limiting step in motion, and the scandal that is surely to come.

And this lands where it always lands: corporate strategy. The only reason to run AI at scale is to create value — to widen the spread between ROIC and WACC, and to grow. Time saved that you cannot check is not value. It is a liability you have not booked yet. A stack of faster reports is vapor unless it raises the worth of the company.

So, the real rate-limiting step is not compute, or budget, or talent. It is your ability to check. Until that scales, more AI does not create more value. It manufactures more risk.

The lifetime value of using AI can only be measured if no scandals, eventually, arise from its use.

P.S. How has this changed the way you are using AI?

Related reading

• AI Strategy Is Not a Strategy — why AI must serve the corporate strategy, not the other way round.
• A Perfect Corporate Sleight of Hand Becomes Tougher With AI — advice that is finally measured against banked value.