AI is intelligence,
not software.

This is not AGI, and it is not human. But it is a genuine and highly capable form of intelligence - and for the first time it can be delivered the way we deliver power, water and information: piped into the organisation, available wherever the work happens.

That framing does most of the work. If AI is software, you buy it, switch it on in one function, and wait for someone else's roadmap. If AI is a resource, the question changes: how do we get this intelligence into every part of the business, and what do we have to build so that it arrives with enough context to be useful?

Two properties make this different from earlier technology waves. It is digital, so it replicates and distributes at almost no marginal cost - one good agent becomes a hundred. And it improves faster than any organic intelligence can. The capability you evaluated eighteen months ago and found wanting is not the capability sitting in front of you today.

Everyone has access to the same frontier models. What separates a useful agent from an impressive demo is context: what it can see, what it remembers, and what it is allowed to act on.

Context does not live in one system. It lives in the CRM and the ERP, in Slack threads and email chains, in supplier PDFs, in a photograph someone took on a site visit, and in the head of the person who has been running the process for eleven years. An agent embedded inside a single application can only ever see that application's slice of it.

So agents should sit across your systems rather than inside them - as connected and as context-aware as the people they work alongside. This is also the plainest explanation for why the AI feature bundled into a tool you already own tends to underdeliver: it is architecturally blind to most of your business.

The practical consequence is a rebalancing: which parts of a workflow are done by human intelligence, which by digital intelligence, and where the handoffs sit. It applies to operations, to problem solving, and increasingly to creative work.

Rebalancing is not the same as replacement. Most workflows contain work that people only ever did because there was no alternative - rekeying, chasing, collating, reconciling, first-pass reading. The gains show up when you re-cut the work rather than bolt an assistant onto the existing shape of it.

Mid-sized companies have a real advantage here, for now. The workflows are still legible to a single person, the stack is not yet fossilised, and the decision to change how something is done can be made in a room rather than a committee. That advantage is temporary, and it belongs to whoever uses it first.

Both are legitimate, and they are not the same investment. Applying third-party AI to your current processes buys productivity: real, fast, and worth having. But it is rented and symmetrical - your competitor can buy the same seats on the same Tuesday.

Out of the box, a general assistant pointed at your drive or your inbox puts you in a probabilistic state. You are not guaranteed to get exactly what you asked for, or the full set of what exists. This is not a defect in the model; it is what happens when several imperfect steps compound.

The number matters less than the shape of the failure. You rarely see the missing quarter. It surfaces as a confident answer built on the documents that happened to be retrieved - fine for a first draft, not fine for a client quote, a compliance response or a board number.

Determinism is a direction of travel rather than an absolute. Moving that way means restructuring and indexing data on the way in rather than hoping to find it later, running multiple passes across relational, vector and graph stores, building deliberate redundancy so completeness can be checked, verifying answers against source, and placing a human exactly where being wrong is expensive.

The same work buys you two things at once: accuracy you can defend, and a system shaped around your workflow rather than a generic one.

Put the two frameworks together and something accumulates. The data you have shaped, the agents you have built, the context you control and the workflows you have redesigned are not a product anyone sells. They are the record of how your company decided to work with this technology, and how that balance was tuned to your business over time.

That layer is the asset. It is also the reason the pace question matters: to move only at the speed of a third-party product is to spend your transition period enriching someone else's IP instead of building your own. We are in the middle of that transition now, and the gap between companies that see it and companies that do not is widening rather than closing.