Every week, someone tells me their company has launched an AI transformation and expects to need a third fewer people by this time next year. Before that transformation reaches a real customer, changes the P&L or reshapes someone’s job, make sure you have: > a problem customers actually notice, rather than a mandate to “use more AI” > one workflow to start with, instead of forty use cases from a workshop > a number the business already cares about, like cycle time, cost per case or cash collected > a baseline measured before anything changes, so you can agree later on what improved > someone inside the business who owns each workflow, rather than leaving it to the vendor or the board > a CEO who takes responsibility for the priority, because a program left entirely to IT will stay an IT program > money allocated in this year’s budget, beyond a mention in the board deck > an upfront decision about what you’ll do with the time you free up: deliver more or spend less > something that gets better for the customer, beyond reducing your own costs > a look at the tools you already pay for, because the capability you need might be a setting nobody has turned on > a clear reason to use AI instead of a rules engine, a report or a better form > an honest list of what you won’t automate > the full workflow mapped by the people who do the work, including what happens when things don’t go to plan > a process that works well enough to automate, because a broken one will just go wrong faster > a written list of exceptions and who handles each one today > the business rules written down, along with why each one exists > reasonably clean data, because duplicates and typos can find their way into the AI’s answers > an agreed system of record for each thing that matters, so two dashboards don’t give you two answers > table and field names that someone in the company can explain > access to the data within days, without having to wait through a procurement cycle > vendor contracts that protect your ownership of the data, so you can leave without losing your history > an honest conversation with the team about how their jobs will change > new responsibilities for the people whose manual work disappears, especially because they know the exceptions best > training built around people’s actual work, rather than a generic webinar > time set aside to learn, because people can’t fit all of this around an already full week > people in every team who can help colleagues adopt it, beyond the same four enthusiasts who try everything > usage data checked against what leadership believes, because “everyone uses it” can turn out to mean four people > managers held responsible for adoption in their teams, beyond showing up at the kickoff > a plan to involve the experienced staff who’ve done it the old way for fifteen years and know where the AI is likely to get it wrong > a hiring plan that reflects the new workflow, so you aren’t automatically backfilling roles it no longer needs > an inventory of every AI tool and agent in use, including the ones employees have set up themselves > an approved tools list, so people have a clear alternative to finding their own > clear rules about which data can go into which tool > human approval before AI moves money, sends formal correspondence or closes an account > a named person who can turn any agent off and knows exactly how to do it > an AI policy that sets the limits, backed by systems that enforce them > logs your CISO can understand > answers for the auditor about what the AI did, who approved it and which version was running > legal review of what AI is allowed to commit to on the company’s behalf > disclosure to customers wherever the law requires it > signed data agreements before real customer or patient records go anywhere near a model > applicable data residency requirements addressed for your EU customers > an incident plan for when AI gets something badly wrong in front of a customer > vendors chosen for what they’ve put into production, with more to show than a good demo > references you can actually call > code, prompts, accounts and IP owned by the company > an exit plan agreed at kickoff, while everyone is still getting along > the ability to change model providers > model contracts that cover price changes and retirement dates > company-owned API accounts, instead of charges going on a developer’s personal credit card > a partner willing to tell you when you don’t need them > cost per completed task measured before rollout; on one logistics rebuild, ours came to about $10 of AI spend per merged pull request > a monthly operating budget for each workflow, rather than budgeting only by seat > hard spending limits that stop the work when you reach them > AI spending and AI savings in the same monthly report > a plan for the token bill as usage grows > a CFO who has reviewed the numbers before anything gets built > a month spent proving it on real data before committing to a twelve-month roadmap > the new workflow running alongside the old one before you switch over > a gradual rollout, starting with one team > evaluations built from real cases by the people who know the work > someone reviewing actual outputs every week, alongside the dashboard > a rollback plan for every change > one workflow running reliably before you start the second > a definition of done that requires the system to be in production and in use > the same business metric measured every month after launch > rework included in the numbers, because faster output means little if people spend longer fixing it > adoption measured by actual usage > a quarterly review where stopping a project is a legitimate outcome > board updates that show business results, rather than counting AI initiatives > evidence a buyer’s diligence team can understand, if the company might ever be sold > skills and knowledge that stay with the team after the outside help leaves And once you’ve done all of that, you’ll still need many of the people you were planning to replace: > the people doing the work today, to explain how it actually happens > those same people, to tell you when the AI gets it wrong > a team lead to decide which cases the AI should never handle > experienced staff to take over when an AI-assisted case goes wrong halfway through > managers to keep adoption going long after the launch email > finance to work out whether anything actually got cheaper > security and compliance to explain to auditors what happened and why > engineers who can connect AI to your systems and keep its access under control > someone senior enough to tell the board that phase two needs to wait until phase one works > the person who still remembers why a rule exists, long after everyone else has forgotten Good luck. 🫡

Want to grow as a leader? Put personal development on your calendar, not just professional goals. I believe in making growth practical. In your next meeting, listen without planning your reply. Before the day ends, ask yourself what you could have handled better. Small acts of self-awareness give you something concrete to work on. Don’t just measure what you accomplish. Pay attention to who you’re becoming while you do it.