AI automation for business and government

Every organisation running an AI pilot can tell you which tool it bought. Far fewer can tell you what happened to the work. That is the harder question, and it decides whether anything changes.

Three kinds of task, and only one of them moves

Pull apart any role and you find three kinds of work. The first is routine transfer: reading something from one place and typing it into another, checking a field against a rule, filing a document where it belongs. Shellharbour City Council described this on our own site before they changed it, roughly 800 development applications a year moved by hand between their system and the NSW Planning Portal.

The second is judgement. Deciding whether a proposal meets a control written for a different decade. Weighing a neighbour’s objection against a planning instrument. Telling a customer something they do not want to hear, and keeping them.

The third is relationship. The conversation that stops a complaint becoming a legal matter. The supplier who takes your call on a Friday afternoon.

AI moves the first kind. It assists with the second. It cannot touch the third. Projects fail when someone buys a tool for routine transfer and sells it upward as though it replaces judgement.

The work nobody budgets for

Automating a task does not delete it. It changes who does it and when.

Somebody decides what the system does when it is unsure. Somebody reads the exceptions, and exceptions are by definition the awkward cases. Somebody answers when a resident asks why a decision went the way it did, which means the reasoning has to be recoverable months later. Somebody owns the thing when the rules change.

That is real work, and it lands on the same team whose task you just automated. Budget for it, or the automation quietly becomes someone’s unpaid second job.

Where the people go

The honest version of redeployment is not a slide about freeing people for strategic work. It is specific, and it usually looks like one of three things.

Some people move up the same process. The officer who spent mornings re-keying applications spends them assessing instead, which is the work they were hired and trained for. This is the easiest move and the most under-used.

Some people move to the backlog. Most government teams and plenty of enterprise ones carry work they have never had capacity to start. Released capacity can go there, and it is the only version of this that shows a visible result inside a year.

Some people move to the exceptions and become the quality function for the system that replaced the routine task. That role is more senior than the one it replaced. Say so out loud, and pay it accordingly.

What does not work is leaving it unsaid. Automate a task, announce no plan, and the team will assume the plan is redundancy. They will be right to assume it until told otherwise.

Enterprise and government are not the same argument

Enterprise buys this on margin. Cost to serve falls, the same team handles more volume, and the saving lands in a line somebody owns.

Government rarely buys it on headcount, because headcount is usually not the thing it is allowed to change. It buys it on backlog and on statutory timeframes. A council that must determine an application inside a legislated period is not trying to employ fewer planners. It is trying to stop breaching the clock, and to stop paying for the appeals and complaints that follow when it does.

One business case written for both produces something true of neither.

Why this is being bought now

In January 2026 Accenture agreed to acquire Faculty, a London AI firm, in a deal reported at around a billion dollars, roughly fifteen times revenue on one analyst’s estimate. Faculty was not an implementation partner. It owned outcomes in two domains, government and health, and it had built its own product rather than configuring somebody else’s.

That is what is valuable now: knowing a domain well enough to change how its work is done, and having built something that proves it.

Where to start

Pick one process where you can name the volume. Not a department, not a strategy. One process with a number against it. Work out which of the three kinds of task it contains. Automate the routine transfer, decide who owns the exceptions, and agree where the released time goes before you release it.

Then do the next one.