The Models Have Never Been Better. So Why Do AI Initiatives Still Fail?

Let’s be honest. Today’s AI models are remarkable: they draft, analyse, code and reason at a level that looked implausible three years ago, and each generation arrives more capable than the last. Whatever else is said about AI, the raw capability is real.

Which makes the pattern inside organisations harder to explain. Initiatives built on these models still stall. Pilots go nowhere. The value that seems obviously within reach keeps failing to materialise. When that happens, the instinct is to blame the model. But, as we’ve said before, the model is rarely the problem. The industry’s own vocabulary has been pointing at the real answer for years.

The AI industry has a habit of cycling through fashionable phrases. A few years ago, the phrase to know was prompt engineering: asking a model the right question. Then it was context engineering: choosing what information the model sees before it answers. A more recent term is harness engineering.

It is tempting to dismiss each phrase as yet another moment of hype. However, it’s more useful to read the sequence as a record of what the industry has learned, because each term marks a real evolution in understanding.

What the sequence is telling us

Prompt engineering treated the model as an oracle. It answered one question at a time, so the wording of the question was where the leverage lived. Whole job titles appeared around phrasing the ask well.

Context engineering came next. Teams discovered that an answer is only as good as what the model can see, and most of what matters - your customers, your data, your definitions, your history - is not in the model at all. The leverage moved from the wording of the question to the information supplied with it.

Harness engineering marks a bigger shift. Models stopped just answering and started doing: using tools, taking a step, checking the result, and deciding what to do next. Once a model can act, the question changes from how to get a good answer to how to let it act safely on your behalf, and the leverage moves outside the model entirely.

The industry has worked through three vocabularies to arrive at one conclusion: the model was never the whole system.

What a harness is and why it carries so much weight

The harness is that system. It determines what the model can see, what it is allowed to do, how its work is checked, and when it must stop and hand it over to a person.

Let’s go back to our familiar talented new hire comparison. No leadership team expects a brilliant hire to create value through raw talent alone; value comes from their role and responsibilities, access to the tools and systems they need to work in, and clarity around how they work with others. While a capable model is the hire, the harness is everything else.

That’s why the harness matters as much as the model, and in live production systems, often more. Swapping one capable model for another is a small change, often a single line of configuration, while building the harness is where the engineering time goes.

It is also why pilots so often look brilliant and then stall on the way to production. A pilot only has to show that the model can do the work; the model is the part you can prompt in an afternoon. Putting the same capability into production means building the harness around it: connecting it to your systems, getting security and permissions right, deciding what gets checked and by whom. That takes knowledge of tooling and security as well as models, and it is the stretch of work where many initiatives stall. When an AI system disappoints in live use, the cause is more often a missing piece of harness than a shortfall in the model.

Where most organisations actually are

The gap I find more interesting than any definition is this: while the industry debates harnesses, most conversations inside organisations are still one or two stages back. Which model should we buy? How should people prompt it? How much context is too little or too much?

That’s not a failing; it’s just part of the journey. It is also why the sequence is useful, because it maps the path and the order:

• Prompting is the entry point: cheap, immediate, and mostly a matter of practice and realistic expectations.
• Context is usually the biggest step: connecting models to your own data and knowledge so the output stops being generic. For most organisations, most of the missed value sits here.
• The harness becomes essential the moment you want AI to act rather than only advise: to touch systems, move work along, and make routine calls on its own.

The order matters because each layer depends on the one before it. An agent given tools but no context acts confidently on the wrong information. A harness built before anyone can prompt well or judge output quality automates a process nobody understands. Skipping ahead is how organisations end up with impressive demonstrations and nothing dependable behind them.

The questions worth asking

So when an AI initiative reaches your leadership table, the revealing questions are where on this path the organisation actually is, and what the next layer of work looks like.

I would be interested to hear if other leaders are seeing this evolution within their own organisations.