After the AI licences go out, most organisations follow a familiar script.
Training gets booked. A hackathon lands on the calendar. Someone spins up a channel so people can post their wins.
None of that is wasted effort. It is also only the opening move. Sooner or later someone asks the uncomfortable follow-up: are we actually better at anything that matters?
That is the question I sat with Richard Geary, Founder of FloHQ, on this week's podcast.
Richard came up through the early SaaS wave at Salesforce, helped SeatGeek's B2B side grow toward unicorn status, and has since built FloHQ around training several thousand people on AI.
The classic mistake
Richard does not treat AI as a special case compared with every other transformational technology he has worked through. What has changed is how cheap and everywhere it feels. Flip it on quickly, and it is easy to believe the hard part is done.
His warning:
"What you end up with is a shiny tool looking for a problem to solve."
I have watched the same pattern. A CEO spends a weekend amazed by what the models can do, then asks someone on the team to "make our business faster with AI."
Richard's order of operations is almost stubbornly old-school, rooted in his years as a solutions consultant: people, process, technology. In that sequence.
Be clear on the experience you want to create. Do you have the right people, trained the right way? Are the processes written down, and does anyone know how to work around them when they break? Only then choose the technology that serves those two.
Buying and switching a tool on is not the same as adopting it with intended use defined, documented, trained, supported, and actually executed.
Old-school diagnosis is back
So where do the highest-impact AI opportunities actually sit?
Richard's answer took me straight back to building our first startup: get people in a room, map the process, cover the windows in Post-it notes. Look at the three V's of data: volume, veracity, variety. Keep asking the five whys until you hit the real break, not the symptom on the shelf.
That lands close to something I keep seeing. Encouraging AI experimentation is healthy. Experimentation without a clear objective, though, produces a lot of noise: agents, workflows, and automations that look impressive in a demo and quietly fail to earn their place in day-to-day work.
The hard part is alignment in two directions at once. Horizontally, you need the business strategy. Vertically, you need enough depth at team and individual level to understand how the work really runs.
"There's always a Sally in accounts or a Brian in the stores who truly knows why things don't work or why they do work when they work."
Leadership can give you mission, strategy, and goals. Sally and Brian give you the operating truth. Skip that mix, top to bottom and across departments, and you will build against intent that never sticks, because it is not how people actually work.
The useful side effect of proper diagnosis is that people understand their own business better by the end of it. That feels slow when AI can produce an answer in seconds. Rushing past it is exactly how organisations get stuck in permanent pilot mode.
From there comes an opportunities audit. After the mapping work, ideas get scored. The exact framework shifts with the business, but the pattern is familiar: potential ROI, feasibility (technical complexity, data readiness, governance and compliance), plus, as Richard stresses, sponsorship and strategic fit.
An idea can look strong on ROI and still die if nobody in leadership will own it and it does not lock to a real business objective.
Training is not capability
This is where Richard and I were properly aligned.
One-off tool training is a certificate for a moment in time. With AI, new models arrive in the same week: you can go to bed feeling expert and wake up behind. Train for that one moment and you are mostly ticking a compliance box. You are not staying good.
Richard's preferred word is enablement. Get people to a shared standard first. Then keep them improving and current.
His driving analogy is the cleanest version of the point. You learn enough to pass the test. Once you have the licence, you still have to learn to drive.
Being a good driver has very little to do with knowing every button in the car. Lessons cover the clutch, the three-point turn, parallel parking. (Or, if you are me, you are still negotiating with that last one.)
Access to an AI tool works the same way. You can learn the features. You can learn some prompting. The real skills sit elsewhere:
- Thinking through processes end to end
- Judging whether AI is even the right tool for the task
- Knowing how to judge a good output
- Using the tools safely and responsibly
Models keep getting better. Prompt craft matters less than it used to. That leaves us surrounded by plausible, often wordy output, and it is still our job to judge it, manage it, and steer it toward the outcome we actually want. That gap is the difference between training and capability.
By this stage of the conversation I was deep into Richard's analogies, and the next one stuck. Because AI sounds so confident, we assume it knows more than it does.
Parents will recognise this immediately. It is 8pm. You tell the kids to get ready for bed. In your head that means a clear sequence: change, shower, teeth, and so on. The instruction felt obvious. How often do you then walk upstairs and find them doing anything but that?
We make the same mistake with AI. We assume context. We skip steps. We expect the blanks to fill themselves. Richard's line to clients:
"AI is the world's most well-read intern. It's read everything that's ever been digitised. But does it know you in that moment in your context with the task you have in mind?"
Brief it like an intern you actually care about. Over-explain the process. Spell out the steps. State the goal. That is the human work up front. With a clear brief, the model will usually go and do the job. Humans still own the critical review at the end. Richard calls that the "human sandwich." Awful label. Useful idea. A tanker one degree off course is a long way down the coast by the time it reaches Africa. Without review and course correction, AI will take you with it.
His practical frame for the brief is GCSE: goal, context, sources, expectations. Put the thinking in at the start so you are not repairing a bad prompt for twenty turns. He is not chasing perfect output. Get to roughly the 80th percentile of usable, then reverse-engineer the pattern so the next run is faster.
Personal wins are not organisational wins
The last thread is the one I cannot stop turning over. Outside a handful of strong pilots, I still rarely see a clean answer for how personal AI wins become organisational ones.
Hackathons create personal productivity. People return to their desks energised. Slides rebuild in minutes. Email gets faster. Those are real individual gains. They will not become organisational leverage unless the business is deliberate about what happens next.
"I don't think we've got a good way of measuring value... without a heavy amount of manual overhead."
Tokens will not tell you who is doing the best work with AI. Neither will logins. Someone who summarises every email can burn through usage while creating little value. Someone who runs one high-leverage workflow a month can look quiet and still move the needle.
The organisations that win will be the ones that translate personal productivity into organisational efficiency and real operating leverage.
So what do you say to "we need to do AI"?
When a leader opens with that line, Richard's answer is deliberately short.
Ask why.
If they cannot answer, or the answer is "the board wants it" or "we want to cut costs," go deeper. The useful version sounds like: we want to do X, stop Y, or offer something differently from today. That is a conversation you can build on.
I agree with him. Tool-first mandates produce exactly the shallow usage pattern we have been living with. Problem first. People and process next. Technology last. That is how adoption turns into capability.
Watch the full episode.
