AI is necessary but not sufficient

In 2000, at the height of the boom in big enterprise software, Goldratt wrote a book with Eli Schragenheim and Carol Ptak called Necessary But Not Sufficient, and its argument was awkward for an industry that was busy selling companies very expensive systems with the promise that they’d transform the business. Goldratt’s view was that the technology was necessary, since you couldn’t get the step change without it, but that it was nowhere near sufficient, because most companies installed the new system and then carried on running exactly the same rules they’d had before, so all they got was the old way of working with a bigger invoice.

I think that book is the most relevant thing anyone could read about AI right now, and I’m slightly amazed that nobody seems to talk about it, because it describes almost perfectly what I see happening in companies that have rolled AI out to everyone and are still wondering where the results are.

Goldratt’s four questions

The heart of the argument is four questions that Goldratt said you should ask about any new technology before you adopt it, and they’re so simple that it’s easy to skim past how much work they do.

• What is the power of the technology? What can it do that we couldn’t do before, put as plainly as possible?
• What limitation does it diminish? What was holding us back that this technology now removes or shrinks?
• What rules helped us accommodate that limitation? What habits, processes and policies did we build because the limitation was there?
• What rules should we use now? If the limitation is gone, what should we do instead?

The trick is in the third question, because rules built around a limitation don’t disappear when the limitation does; they hang around, because everyone’s used to them and nobody remembers why they were created, which is just another form of the inertia Goldratt warned about in the fifth of the five focusing steps.

Now ask them about AI

The power of AI, for a company, is that it can read, write, summarise, remember and ask questions at almost no cost, at any hour, for everyone at once, which means the work of knowing what’s going on no longer has to be done by people.

The limitation it diminishes is the one that has shaped how companies are run for a hundred years, which is that gathering information from across a business, moving it to the people who need it and keeping it up to date used to be slow, expensive and done by hand, mostly by managers whose job was largely carrying updates up and down the org chart.

The rules we built to accommodate that limitation are everywhere once you look: the weekly status meeting where everyone reports in turn, the monthly report that takes three days to write, the quarterly planning cycle because gathering the facts more often was too expensive, the cascade that passes goals down one level at a time, and the scorecard that reduces everything to a percentage because nobody had time to read the reasons. I made a version of this argument in Information is cheap now. Decisions aren’t., but Goldratt’s questions make it sharper, because they show these aren’t just old-fashioned habits; they’re rules built around a limitation that has now largely gone.

The rules we should use now follow naturally, which is that people update the plan as things happen, in a sentence, rather than saving it up for a meeting; that the time the team spends together goes on decisions rather than reporting; that the plan is a living loop rather than a document redrawn every quarter; and that everyone, from the founder to the board, sees the same current reality without asking anyone for it.

Faster at the wrong things

Here’s what I actually see, though, which is that most companies are using AI to make the old rules quicker rather than replacing them, so AI writes the status report, AI summarises the status meeting, AI drafts the quarterly review and AI turns the scorecard into a nicer slide. Every one of those saves somebody an hour, and every one of them keeps alive a rule that only existed because of a limitation AI has just removed, which is precisely the mistake Goldratt watched companies make with enterprise software twenty-five years ago.

It’s also the scout hike from The Goal all over again, because speeding up the parts of a company that aren’t the constraint doesn’t make the company move faster; the line just stretches further while the kids at the front feel very productive. If writing the update wasn’t what was holding your goal back, then writing it ten times faster won’t bring the goal any closer, and you can see this in companies that have given every person an AI assistant and are moving no quicker than before, which I wrote about in Built for the whole company.

The constraint has moved

The interesting thing about applying the Theory of Constraints to AI is that when a technology removes a limitation, the constraint doesn’t vanish; it moves, and I think it has moved somewhere most companies aren’t looking. When information was expensive, the constraint was usually knowing what was going on, but now that knowing is nearly free, the constraint is the part that only people can do, which is agreeing on what’s really in the way, thinking it through properly and making the call.

That has two consequences that I feel very strongly about. The first is that AI should be subordinated to that human constraint, in exactly Goldratt’s sense, meaning its whole job is to feed the decision, by gathering the facts, asking the right questions and laying it all out so the team can decide, which is the line I drew in The machine asks, people decide. The second is that an AI which does the thinking for people doesn’t lift the new constraint at all; it weakens it, because judgement is a muscle and, as the research in AI isn’t rotting your brain shows, it fades when you stop using it, which is why I’d always want AI to be a mech suit, not a replacement.

The new rules, built in

This is the thinking behind how AJ, the AI inside Clearahead, works, because we tried to build the new rules into the product rather than leaving teams to invent them. Instead of a status meeting, people tell AJ what happened in a sentence when it happens, and AJ updates the right items and keeps what they said as evidence; instead of a monthly report, This week from AJ lands every Monday with what moved, what you learned and up to three things that need a decision; and instead of a quarterly reset, the plan is a loop that finds the next constraint the moment one clears. The weekly rhythm, the summary emails and notifications fill in the rest, and none of them asks anyone to report.

If you’re planning how your company should adopt AI, I’d start with Goldratt’s questions rather than with a list of tools, and the AI adoption planning page shows how a leadership team can run that as a goal with constraints of its own. If you’re wondering whether a general assistant could do all this, I’ve set out the honest pros and cons in ChatGPT or Claude vs Clearahead, and the short version is that individual AI is necessary but, for a whole company, it isn’t sufficient. The last piece in this series walks through every part of Clearahead’s model and the idea it came from.

Questions people ask

What does “necessary but not sufficient” mean in the Theory of Constraints?

It’s the title of a 2000 book by Eliyahu Goldratt, Eli Schragenheim and Carol Ptak about business software. Their argument was that new technology is necessary for a step change in results, but it isn’t sufficient on its own: the benefit only arrives when a company also changes the rules it had built to cope with the limitation the technology removes.

What are Goldratt’s four questions about technology?

What is the power of the new technology? What limitation does it diminish? What rules helped us accommodate that limitation before it was diminished? What rules should we use now? Some versions add questions about how to change the technology and how to cause the change, but those four are the core.

How should a company adopt AI?

Start by asking which limitation AI removes for your business and which of your habits only exist because of that limitation, then change the habits, not just the tools. For most companies the limitation is the cost of gathering and moving information, so the rules to change are status meetings, reports and long planning cycles.

Can’t we just use ChatGPT or Claude to run our goals?

A general AI assistant is very good at the work an individual does, but it doesn’t hold a shared plan the whole team works from, and it forgets the context between conversations. Running a company goal needs a shared model, a record of what changed and why, and an AI that works from both.