•Exponential Partners

AI change is not a software rollout

Most organisations know how to roll out software. You choose a vendor, configure it, migrate the data, train the users, switch it on and close the project. The plan has an end date, and once the licences are assigned the work is largely done.

AI does not behave like that. Treat it as if it does and you can end up with a tool everyone has and a way of working nobody changed.

The gains sit in the workflow, not the licence

In her annual lecture at the LSE on 10 November, the Cabinet Secretary, Dame Antonia Romeo, said that "pure automation of specific tasks absent a rethinking of the entire workflow, or indeed the whole operating model, will not fully capture the underlying gains from technology".[1] Her example was MHCLG using AI to analyse planning applications, which has helped halve processing times for routine householder applications.[1] She also said the starting point must be streamlining underlying business processes "which have calcified over time".[1]

A rollout asks whether people can use the tool. A change asks which piece of work should now run differently, who owns the new version and what happens to the steps either side of it. If the honest answer is "the same team, doing the same steps, with an assistant open in another tab", expect modest gains.

Go-live is not the end of the project

A conventional rollout has a configuration you test and sign off. AI products keep changing underneath you. From 9 December, Microsoft will enable Anthropic's Claude models for UK Copilot tenants whose administrators have not already chosen a setting, and The Register reports that Anthropic processes the relevant data in the US.[2] On 19 November, Claude Data turns on by default for Anthropic's Enterprise plans.[3] Neither change needs a project inside your organisation to happen. Both need someone to decide whether they should.

Model choice moves too. When Anthropic launched Claude Fable 5.2 on 9 November, it recommended starting most coding and knowledge tasks on Opus 5.5, then switching up or down depending on the scope and complexity of the work.[4] Which model handles which step is now an operating decision you will keep coming back to.

That argues for an owner after go-live, with a standing review of new defaults, new models, changes to where data is processed and anything the vendor switched on that should be switched off again.

The plumbing usually comes first

On 10 November Qantas announced a multi-year agreement with Accenture to "modernise its back-end systems, processes and data, as it steps up investment in technology and AI", according to The Guardian, which reports that the airline is preparing to cut 850 Australian jobs, with the first cuts aimed at finance, procurement, human resources and marketing staff.[5] Whatever you think of the decision, the sequence is worth noticing: the AI investment arrives bundled with a clean-up of systems and processes, and the effect on people lands early.

Most organisations will not make changes at that scale, but the same dependency shows up in smaller projects. An assistant that drafts supplier responses is only as good as the supplier records it reads. An agent that reports on revenue needs to know which revenue. Launching an always-on version of its Genie agent on 11 November, Databricks' chief executive Ali Ghodsi said Genie can run in the background and come back only when it needs a decision because "it knows how our company works: what we mean by revenue, which numbers are official, and who can see what".[6] That knowledge does not come with the model. Someone in your organisation has to write it down and keep it current.

Be straight with people about their work

Software rollouts come with training plans. AI change also needs an honest account of which tasks are moving, which roles will change and what people are expected to do with the time that comes back. The Cabinet Secretary was direct about it: "Yes, jobs will change."[1] In the same lecture she said civil servants need to be equipped "with the most relevant frontier models for their work".[1] Describe the change in terms of the work people actually do, and say it early.

Measure the work itself

Adoption dashboards count users. They do not tell you whether the planning application, the invoice or the customer reply moved faster or came out more accurate. Before you scale, pick the measure that fits the workflow you changed, such as elapsed time, rework, errors caught or cost per completed task. Take it before the change and again after.

Five questions before you call it done

Before signing off an AI programme as delivered, ask:

  • Which workflow changed, and who owns the new version?
  • Who reviews vendor defaults, model changes and data-processing terms, and how often?
  • Where are the business definitions the system relies on written down?
  • What have you told the people whose work changed?
  • What did the measure of the work show, before and after?

If those answers are vague, the software may be live, but the change has not happened yet.

Sources

Related Insights

·Riz Pabani

Salesforce in Claude puts your sales process on trial

Salesforce in Claude brings live CRM data and 37 sales skills into Claude. Why your sales process now decides the results, and what to fix first.

·Riz Pabani

What Agentic AI in Wealth Management Actually Looks Like When You're the One Building It

Deloitte says agentic AI could add $35tn of capacity to wealth management. Here's what it looks like from inside a real build: the spreadsheets, the 90/10 rule, and the compliance work nobody pitches.

·Riz Pabani

The CFO Question That AI Has Finally Answered

C.H. Robinson went from quoting 60% of inbound requests in 15–45 minutes to 100% in seconds — and the stock rose ~20%. What Q3 2024 proves about where AI ROI actually lives.

Back to all insights