How To Get Real Value From AI by Simon Russell

By Simon Russell, Head of AI and Data, Instant Impact

Ask most AI tools to summarise a report and the result is genuinely impressive. Ask the same tool to build you a financial model and it will hand back something that looks right, reads confidently, and is quietly wrong. Nothing has broken. You have simply asked a reasoning engine to do arithmetic, and arithmetic is the one thing it cannot do.

That gap, between what these tools appear capable of and what they are actually built to do, is where most organisations lose time, money and trust. So before any conversation about adoption, it is worth being precise about what you are dealing with.

What you are actually working with

A large language model is a statistical model. It is trained on an enormous body of text, most of the internet plus code and other sources, and then refined on how people converse so that it holds a dialogue naturally. That refinement is why different models have different personalities, and it is also the source of the illusion. Because the thing talks like a person, we assume it reasons like one.

It doesn't. It has no inherent concept of a number, and strictly speaking no concept of a word either. What it has is an extraordinary ability to predict what comes next, which manifests as fluent reasoning and, just as often, as fluent nonsense. It will fill gaps you never asked it to fill, and it will do so with total confidence.

Hold onto that, because almost every practical decision about AI flows from it. The tool is brilliant at language and hopeless at certainty. Match your tasks to that reality and it works. Ignore it and you get the confident financial model that is off by a mile.

Why the 'slow' organisations are winning

There is a pattern that surprises people. The organisations getting the most out of AI are frequently the large, heavily governed ones, the sort you would expect to move slowest. The reason is counterintuitive but simple.

To get consistent output from an AI agent, you have to hand it a task broken into small, clearly described steps, with a definition of what good and bad look like and where the boundaries sit. Large organisations did that work years ago. All that process documentation, the systematic mapping of who does what and how, the discipline that felt like pure overhead, turns out to be exactly the foundation these tools need. Better still, those organisations quietly collected evidence of good and bad outcomes at scale, and that evidence can be fed straight in.

The uncomfortable lesson for everyone else is that the advantage was never a clever prompt. It is knowing your own operation well enough to describe it precisely. If you cannot explain a process to a capable new starter in plain terms, you cannot hand it to an agent either.

Don't automate the human version

When teams begin, the instinct is to take a workflow that evolved for humans and ask an agent to copy it, step for step. It is the obvious move and usually the wrong one. Those workflows carry generations of human compromise, and mimicking them just automates the compromise.

The sharper question is what you would design if you wanted the outcome and had agents to achieve it. You quickly stop imitating the human version and start building something leaner. This matters because agents are trained on human language, which is full of ambiguity, and a large part of the real work is stripping that ambiguity out so the system behaves predictably.

We are already moving toward agents dealing directly with other agents. It is visible at both ends of recruitment, candidates using tools to sharpen applications while employers use tools to screen them, and it is coming fast to finance, where a mismatch between an invoice and a purchase order is on its way to becoming a machine-to-machine exchange. The honest question for every sector is when we stop having software imitate people talking to each other and design the exchange properly.

The part nobody can copy

Here is the reassurance and the challenge in one. You are not going to build these tools. You will buy them from vendors, and so will your competitors. The agent itself is not your edge, because everyone has the same one.

Two things remain genuinely yours. The first is the service wrapper, how you orchestrate delivery, the processes you define, the human relationships that still carry the work. The second sits underneath everything and is easy to miss. An agent is not a person. It is data and context. Left to run on generic context, it produces generic output, identical to whatever your competitor's identical agent produces.

The organisations that pull ahead will be the ones that capture what actually makes them distinctive, their culture, their tone, the hard-won judgement of their best people, and turn it into context an agent can use. This is a knowledge management problem that has been undervalued for a decade and suddenly has a point. I have started calling the discipline context governance. We are used to governing data and information. Now we govern the context agents operate inside: their goals, which drift, and their rules, which have to be maintained deliberately.

There is a subtlety worth naming. In a team, you get the behaviour you tolerate, and when a person drifts, time and a quiet word tend to correct it. An agent is software. It never forgets, and it never self-corrects. Deciding what it should retain and what it should let go is a live skill we do not yet have good tools for.

Where the value is heading

If you want to know which capabilities to invest in, I would point to three.

Process decomposition comes first, the ability to take a large, messy workflow and break it into discrete, describable, verifiable pieces. It is the scarce skill of this moment, and it is the gateway to everything else.

Knowledge management comes next. The data governance people many organisations underfunded for years are about to become some of the most valuable in the building, because they are the ones who can manage the context that makes agents effective rather than generic.

The third is the most human and the one I worry about most. A while ago I asked a tool to work out a margin and it confidently returned a markup. Those are very different numbers. I caught it in seconds, but only because I have made that exact mistake myself over fifteen years. Someone in their first year would have shipped it.

That is the real problem of this era, and it is not a technical one. Easy access to knowledge through AI is not the same as understanding, and knowledge without understanding is close to useless. Understanding is built the slow way, through curiosity, through getting things wrong repeatedly, through refusing to copy and paste past something that feels off. We now have to develop that instinct in people faster than experience alone used to provide it.

Which is the through-line of all of this. The organisations that thrive will not be the ones with the best technology, because the technology will be the same everywhere. They will be the ones who understood it well enough to know exactly what to ask of it, and what never to.

___

See how an AI-enabled tech stack fits inside your business: Plan your hiring approach.