Software stopped being the bottleneck. Judgement didn’t.

For two years the same question has opened the board meeting: which AI tool should we buy. It is a procurement question standing in for an operating one, and it has the merit of admitting an answer — which is why it keeps being asked.

What has been happening for eighteen months is not the death of software. It is a relocation of value: it leaves the tool and settles somewhere nobody ever filed it, where almost nothing is written down.

What is moving

Enterprise software sold for thirty years on a simple promise: it held the rule so that people did not have to. A CRM imposed stages, an ERP a chart of accounts, a ticketing tool an approval path. The value never lay in the screen. It lay in the rule being executed without argument.

That part is becoming cheap. A model can read an invoice, match it to an order, draft the chaser. What it cannot know is what this particular business treats as an acceptable variance, above which amount someone picks up the phone instead of writing, and which client is never chased on a Friday.

The shift can be read in the vendors themselves. Salesforce repositioned around the agent rather than the interface. Docusign stopped selling somewhere to file a contract and started selling a reading of its obligations. Harvey is settling into law firms on a promise to draft and to review — and is worth, inside any given firm, exactly what that firm has managed to put into words about its own standards. In all three cases, what is sold is no longer the screen: it is the rule made executable.

This is the other face of a subject already covered here: what is left of a tool when nobody opens its interface. Seen from the vendor, it is a business-model problem. Seen from the customer, the question is simpler and more awkward: how much of your own trade can you actually write down?

Four levels of agent

The word agent covers four very different things, and conflating them is expensive. Most conversations about AI agents for business collapse all four into a single word. They are better taken for what they are: rungs.

• The assistant. It waits to be asked. It drafts, summarises, answers. The process stays with the person, who keeps both the control and the responsibility.
• The process agent. It owns a task end to end inside a frame that is already written: chasing a missing document, matching an invoice, routing an incoming request. The process exists; the agent runs it.
• The policy agent. It applies a policy that calls for appraisal: is this clause acceptable, does this file fall inside the stated case. It requires the rule to have been articulated, and its exceptions with it.
• The mandated agent. It acts for the business within a stated limit: it commits, it arbitrates, it answers. What it does binds the organisation, which moves the question from time saved to liability held.

These rungs are not climbed by changing model. They are climbed by writing. Each level asks for one more layer of judgement, and that is exactly where projects stop: not for want of a tool, but for never having set down in black and white what done well means here.

Which processes suit it

Not all of them do. Four criteria sort them, and they hold together.

• Repetition. The work comes round often enough to pay back the effort of writing it down. One case a quarter never will.
• Judgement. Something has to be appraised. If a simple rule were enough, a script would have done it long ago.
• Fast verification. An expert has to be able to spot a mistake in seconds. This is the deciding criterion.
• A reversible entry. One task can be handed over before the whole process is. A scope that cannot be narrowed is a scope that cannot be corrected.

The third is the one that decides, and the one most often forgotten. Delegating assumes the ability to judge what comes back: without that second half, the work has not been automated, it has merely stopped being looked at.

The question to ask

Choosing the tool is an execution decision. It is made, revised, replaced; exit costs are real, but they can be costed. It is not a strategy, and treating it as one confuses the vendor with the trade.

The strategic question sits elsewhere: which decisions does this business make every week on judgement nobody ever wrote down? Who holds it? What happens when that person is on leave, and what happens the day they leave for good?

That judgement always exists. It sits in a workshop manager’s head, in a bookkeeper’s habits, in the calls a salesperson makes without ever naming them. It is written nowhere, which is precisely why no agent can take it over — nor, for that matter, can any new recruit.

What an audit is there to produce

This is the point of an IT and data audit, and it is widely misunderstood. An audit is not there to recommend a tool: the market sees to that, free of charge and with enthusiasm. It is there to write down what the business already knows how to do without knowing it — the rules, the exceptions, the thresholds, the cases handled differently and the reasons they are handled differently.

That document has immediate value, before any artificial intelligence project: it makes the organisation transferable. It has a second value, which is to make agents possible at all. The order is not negotiable, and reversing it produces exactly what can be seen everywhere: pilots that work in a demonstration and that nothing allows anyone to generalise.

This is the work Maeliom Consulting does: an inventory and a formulation before it is ever a technology question — which is why IT audit services come before the choice of tool and not after, in the same way that mapping is not inventorying.

A report describes what ought to be done. A transformation writes down what the business already does, so that others — human or not — can take it over. We transform. We don’t just advise.

Common questions