As I watch companies put AI into practice, I've gone back to see how the picture has changed. Eighteen reports on the state of AI from the last five years, from McKinsey, Microsoft, IBM, BCG, Deloitte, Stanford, MIT, Wharton and Menlo Ventures.
McKinsey's surveys show widespread AI use and a small group reporting substantial business value.
Across the reports, the recurring advice is to redesign the work and prepare the people before choosing the technology. On how to measure whether it worked, they are all over the place.
Redesign the work before you buy the technology
That first part matches my experience. The workflow redesign and the change management need to be in place before the technology. Decompose the work, decide what moves to AI and what stays with people, redesign the workflow around that, then pick the tools.
I'd add that we need to design around AI's current capabilities and limitations. An error in one step can affect the steps that follow. In my experience, the cost of those errors rises as they compound across a long workflow.
The investment needs to reflect the work involved in redesign and adoption. McKinsey's adoption paper describes one unit spent on the technology, about three on redesigning the process, and five on capability and adoption. The paper says most companies invert that allocation.
McKinsey's management playbook reports that only one in ten companies has adopted the distributed operating model used by the successful companies it studied. In that model, teams close to the business develop and improve solutions themselves.
Make or break sits with the middle layer
Every transformation I've worked through has depended on middle managers. As organizations become flatter, their role has to change too.
They need to take more direct responsibility for how the work gets done, making decisions, resolving problems and helping their teams adapt. Fewer layers will achieve little if managers still spend their time collecting updates and passing approvals up the chain. The redesign has to give them the authority to act and make them accountable for the results.
The reports also describe the pressure on middle managers. Executives describe a smoother experience than the people below them, and the strain shows up with middle managers.
What helps is familiar. An operating model people can find their own job in, ways of working they had a hand in agreeing, credible champions, real training. A credible champion is a colleague who uses the tools in their own work and can show others how to use them.
Give a senior business leader ownership of the outcome
IT and the Chief AI Officer share responsibility for making AI work across the business. IBM found that 76 percent of the organizations in its CEO survey had a Chief AI Officer, up from 26 percent a year earlier.
The outcome belongs to someone senior in the business. In the twenty companies McKinsey picked because they had already made AI pay, a senior business leader owned the outcome.
Measure it with the numbers you already have
A leader accountable for the outcome needs a clear measure of success. The reports use different measures, which makes their results difficult to compare.
Ask how often AI delivers a return and the reported success rates range from about one in twenty to three in four. Each report counts something different: a pilot, an initiative or a whole company, with its own threshold for success.
As an industry we don't have a shared way to measure this. Pick the unit you're judging and the threshold that counts as success, and the KPIs and OKRs you already have will do the job, the same as for any other initiative.
When deciding whether to build or buy, I would assess the return over the software's full life, including maintenance and day-to-day operations.
In my experience, that's where much of the cost of software lands, and producing a first version quickly tells you little about the cost of running and maintaining it over the next three years. If it's a one-off or a small capability, build it. If it's a platform the business will depend on, price the full life of it before you start, and use your new ability to build to get better terms from your providers.
What we do with the capacity
McKinsey's management playbook describes a path towards a flatter organization running on smaller teams. Some of the companies it looked at have been working on their transformation for more than five years.
In my view, productivity comes from redesigning the work around AI capabilities that are useful in practice, with their limitations accounted for. That frees people to do more with the time available.
Where that capacity goes is a separate decision, and it's the one I'd think hardest about. It can come back as a smaller cost base, the same work done by fewer people. Or it can go into work the business never had the people for, staffed by the people whose old tasks moved to AI.
What I'd avoid is cutting the team before redesigning the work. You risk losing the people you need to carry out that redesign and make a smaller team viable.