What AI Will Not Do for You Yet

We build AI systems for a living, so this list costs us something to publish. It is worth it, because the fastest way to waste a year is to point this at the wrong problem, and the second fastest is to conclude from that failure that none of it works.

MIT’s Project NANDA reported in July 2025 that against an estimated $30 to $40 billion of enterprise investment, roughly 95% of generative AI projects produced no measurable return. Their study covers large organisations rather than companies of your size, and “no measurable return” often means nobody set up a way to measure. But the number is a fair warning about aim. Most of those projects failed on selection, not capability.

So here is the boundary, honestly drawn.

It will not fix a process nobody has agreed on

If two experienced people describe the same process differently, automating it produces a system that is wrong for one of them and probably both.

This is the most common failure we see and it is not a technology problem. Sit in a room for an hour and settle what the process actually is. Frequently that hour delivers most of the value the project was going to, and occasionally it makes the project unnecessary.

It will not carry accountability

A system can make a recommendation, produce a draft, flag an exception and route a decision. It cannot be the party responsible when the decision was wrong.

That matters most in regulated work. In lending, insurance, health and legal advice, a named person signs. Build the AI step to prepare, evidence and speed up that signature. Do not build it to replace the signature, because the liability does not move and the regulator does not care how good the model is.

It will not know things nobody wrote down

The client whose invoice always needs a purchase order number. The supplier who says two weeks and means four. The site that requires a different induction. If it lives only in somebody’s head, no system can use it.

This one has a straightforward fix that most businesses skip: get it written down. That is a genuine project of a few weeks, and it is worth doing regardless of whether you automate anything, because it is also your answer to what happens when that person leaves.

It will not replace the relationship

It can make sure the right conversation happens with the right person at the right moment, having read everything that came before. It cannot be the conversation.

Anywhere trust is the product, the human contact is the product. What AI does there is remove the reasons you currently miss those moments, which is a large gain and a different one from replacement.

It will not save you from a bad offer

If the proposition is not compelling, better targeting, faster follow up and more personalised messaging will move a bad number slightly. The problem is upstream and no amount of system fixes it.

Worth saying because we are asked to fix pipelines that are not broken. The mechanics work and nobody wants what is being sold at the price it is being sold at. That is a positioning conversation, and it is cheaper than anything we would build.

It will not run itself forever

Systems drift. Your business changes, the tools you connect to change, and a system nobody owns degrades quietly.

Budget for ownership from the start: a named person, a monthly look, and alerting when something stops. Most abandoned automation was not badly built. It was unowned.

Now the other side of the line

Everything above is scope, not failure, and the space inside the boundary is very large. What is genuinely, reliably solvable today:

• Reading anything. Documents in any format, emails, photographs of paperwork, handwriting. Extracting what matters and writing it into a system.

• Judgement at volume. Which of four hundred enquiries deserves attention first, which records are the same client, which contract has an unusual clause.

• Personalised contact with a whole book. Not the twenty accounts somebody remembers. All of them, written from what is actually in each file.

• Watching for what changed. The client who has gone quiet, the job that is drifting late, the number that moved for a reason nobody has noticed.

• Everything that currently waits on one person being available to do something routine.

Each of those was blocked for the same historical reason: the step needed judgement, judgement meant a person, and a person could not do it at volume. That constraint is genuinely gone, and it is the reason the results now available are large rather than marginal.

Two things people expect it to do badly, and it does not

Worth correcting, because both stop projects that would have worked.

“It cannot handle our messy data.” This was true and is not. Messy input is the case AI handles best, and structured input is where conventional software already worked. If somebody tells you to clean everything up before you start, be sceptical: the reading is frequently the easiest part.

“We are too small.” The opposite. A smaller company has one owner, one process and one decision, so the distance between something working and it being how you operate is a conversation. That distance is precisely where large organisations lose their projects.

How to use this list

Take whatever you are considering and check it against the six limits above. If it clears all six, it is very likely to work and you should move quickly.

If it fails one, that is not a no. It is a smaller, cheaper job to do first, and it is usually a job worth doing on its own terms.

Sources

• Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT Project NANDA, July 2025.

AI Optimize will tell you in the first conversation when the honest answer is that a process should be fixed rather than automated. That is why the things we do build tend to still be running a year later. Start with Workflow Automation.