Web Courses Bangkok · Brief
1.0 Summary
Forget the cleverer tool for a minute. What helps is some boring planning that fits on one sheet of paper, done before anybody opens a thing, and you will need a job, a finish line, some rules for the data, a named reader and a date to decide.
Three months is loads of time to waste on an AI pilot, and I reckon it is very easy to do. Somebody gets excited about a tool, half the team signs up, a few people poke at it for a week, and then nobody can say if it was worth the bother. Nothing broke. Nothing got decided either.
The cure is a bit of boring planning on one sheet of paper, done before anyone opens anything. So here’s what I’d put on that sheet, in the order I’d fill it in, with a few things to try on the way: a job checker, a pilot form, a “can this go in?” register and a decision stamp for the end.
The planning advice is my own teaching view. The numbers come from sources you can open, listed and numbered at the bottom.
| 1 job | not “AI for marketing”. Something that keeps coming round |
|---|---|
| 3 answers | on paper before you start: time now, quality now, what counts as a win |
| 4–6 weeks | is plenty for most jobs, with one or two people on real work |
| 3 outcomes | roll out, tweak and go round again, or stop. Stopping is fine |
2.0 Evidence
Four numbers on why a pilot needs a plan. Each figure links to where it comes from.
| Ref. | Finding | Figure | Source |
|---|---|---|---|
| E1 | of AI users at work bring their own AI tools. Your team may already be piloting | 78%1 | Microsoft & LinkedIn, 2024 |
| E2 | of leaders worry about quantifying the productivity gains from AI | 59%1 | Microsoft & LinkedIn, 2024 |
| E3 | average productivity gain for 5,179 support agents given an AI assistant. 34% for novices | 14%2 | NBER |
| E4 | of organisations reported using AI in 2024, up from 55% the year before | 78%3 | Stanford AI Index 2025 |
- Read the 14% with care: that was customer support at one company with one tool, and experienced agents gained little. It’s a reason to measure your own team’s job, not to borrow the number.2
- The Microsoft and LinkedIn figures come from a 2024 survey of about 31,000 knowledge workers in 31 markets,1 and the Stanford figure counts organisations reporting any AI use, which is a low bar.3
3.0 Pick one job
You can’t run a pilot called “AI for marketing”. It’s far too woolly, and you’d never know when it was over! So pick a job. Choose something that keeps coming round, which you could explain to your gran in a breath. The weekly customer update, say, or the support inbox, or turning meeting notes into a list of who is doing what.
How do you find the job? Ask the people doing the work what they hate doing every week. If you ask where AI could help instead, you only get wish lists. But ask about the annoying stuff and you will have a dozen answers in ten minutes. A few of those will be jobs people already hand to a chatbot on their own phone, which is good to know about.1
Test a job you have in mind. Tick what’s true.
4.0 Write down what a win looks like
Nearly everybody skips this step, and I would say it is the whole thing. Get three answers on paper before you start. How long does the job take now? How good is the result now? And what would make you say “right, that worked”?
Be very fussy here. “The weekly update takes me 90 minutes and needs two rounds of edits” is something you can measure, and “faster” isn’t. Then pick a finish line you’d honestly accept, say “under 45 minutes, and no more edits than now”. Do it today, because once things are going everybody will want to nudge the goalposts. You’re not alone, either. In a big survey, 59% of leaders said they worry about putting a number on what the tools save them,1 and a finish line written down in advance keeps you out of that club.
Fill in the form. It starts with the example from above. Swap in your own job.
Checks (done for you)
Nothing is saved or sent. It’s a worksheet to copy onto a real sheet of paper.
5.0 Settle the data rules on day one
So what can go into the tool, and what cannot? My rule would be that nothing private goes in. No customer names, no contracts, no passwords, nothing personal about your staff. Stick it on the wall where people can see it, and find out which account they’re really using, because a free personal login and a company one do not always treat what you type the same way.
It happens to the big companies too. Samsung banned its staff from using generative AI tools on company devices after, according to Bloomberg’s report, some of them uploaded internal source code to ChatGPT.6 Complete the register. Would you put it in?
Data registermark each line Permit or Refuse
- A list of customer names and emails
- Our public price list
- A signed client contract
- A password to the shared dashboard
- A rough outline for a public blog post
- Your notes from a one-to-one with a team member
Mark a side for each one.
Need something official to point at? The UK’s data protection regulator has a very long page on exactly this, though they say it’s under review after a change in the law.5 The American standards body also has a framework, and it’s voluntary, with a separate profile for generative tools.4 7 I’ll be honest, neither is a page-turner, but you will be glad of them the day somebody senior asks if anybody has thought about all this.
Laws are different from country to country, so if it’s anything legal, please ask a proper expert, not me! If you sell into Europe, our post on European AI regulations is a decent place to begin.
AI and data protectionAI Risk Management FrameworkEuropean AI regulations
6.0 Tell people what you are doing
Say it out loud, early. Some of your team will be worried that a trial like this is really a way of working out who can be replaced, and if you leave that unsaid it will fill the silence all by itself. So tell them what you are measuring (the time a job takes, not how good anybody is), why you picked this job, and that the point is to get rid of the boring bits.
What I’d say, in about 20 seconds
- “We’re measuring how long one job takes, not how good anyone is.”
- “We picked it because nobody likes doing it.”
- “Tell me what you think. You’ll spot the problems with my plan before I do.”
Then ask the people who do the job what they think. They will spot problems with your plan within about five minutes, and they will be much keener to try it if it was partly their idea. It’s worth knowing that the people you ask may be ahead of you: in Microsoft and LinkedIn’s survey, 78% of AI users said they bring their own tools to work.1
7.0 A human in the loop, kept small and short
Someone real should read what the AI wrote before it goes anywhere. Give them a name, so it is clearly their job to spot the howlers. I’d also have them keep a rough tally on a scrap of paper. How often was it fine as it was? How often did it need a small fix? How often was it rubbish? That tally will tell you more than anyone’s opinion, mine included.
Four to six weeks is plenty for most jobs. Have one or two people use the tool on real work, not the whole department. Once a week, spend ten minutes asking what went well and what drove them mad, and measure the time the same way you did for the baseline, otherwise you’re comparing apples with pears.
8.0 Time is up: put your numbers next to the finish line
Roll it out, tweak it and go round again, or stop. Stopping is a perfectly good result! A trial that says no after five weeks has done its job, and you have saved yourself a licence fee. Type in your tally and see which way the numbers lean.
The reviewer’s tally
roll it outgo round againstopping is fine
Write the decision down, with the reason. Six months from now, when someone suggests the same tool again, you’ll be very glad you did.
The calling rule here is mine, not a standard: roll out if you beat the finish line and under a fifth of outputs were rubbish, stop if you are missing it by a long way, otherwise tweak. Your own finish line, written on day one, matters more than my thresholds.
9.0 Actions: fill in the sheet before you open any tool
Tick them off as you go. Your progress stays on this device.
| Done | Action |
|---|---|
| Ask the team what they hate doing every week | |
| Pick one job and time it today | |
| Write the finish line you’d honestly accept | |
| Put the “nothing private goes in” rule on the wall | |
| Name the person who reads the output, and set the decision date |
So which job in your team would you hand over first?
10.0 How the pilot ends
It ends on a date, with a decision, and that’s all it ever needed to do. You pick the job and time it, you write the finish line down, and you put the rule about private stuff on the wall. A few weeks later somebody reads the tally and you say roll it out, tweak it, or stop. I reckon far fewer trials get that far than should, which is a shame.
Take the example from the form. The weekly update took 90 minutes and the finish line was 45, and the helper has it coming down to 50 on average, with six outputs fine, five needing a small fix and just the one rubbish. That’s a near miss, and the helper says tweak. You would go round again for a few weeks, with better instructions, and see if you can shave off the last five minutes. Those numbers are made up, remember, but that’s what a good ending looks like. It’s boring, and it has a number on it.
What you have at the end
You have a decision you can defend, and a number to go with it. Even if the answer is stop, you’ve saved a licence fee and a lot of meetings, and you know a bit more about how your people really work. My own guess, and it’s only my teaching view, is about a week to time the job and then four to six weeks of trial, so a month and a half from a standing start.
Don’t expect miracles. In the biggest study I know of, 5,179 customer support agents who were given an assistant got through 14% more issues an hour on average, and 34% more if they were new, but the experienced agents gained very little.2 That was one company and one tool, so it’s no promise for your team. What it does tell you is that the gains can be real and uneven, which is exactly why you measure your own job.
What to do next
This week, I’d ask your team what they hate doing and pick the job from that. Time it a few times if you can, and write the finish line down on paper. Don’t open any tool yet. It will take you an afternoon.
A month in, put the rule about private stuff on the wall, name the person who will read what the tool writes, and tell your people what you’re measuring and why, in those twenty seconds. Then get going with a couple of them on proper work, and keep up your weekly ten minutes.
After that, when the date comes, put the numbers beside the finish line and decide. Write it down with the reason. Then you can pick the next job and go again, a bit quicker, because the sheet is already done.
How people learn this
You learn it by going small and being wrong where it’s cheap. Your first finish line will be too vague and your first tally will be too kind. The reviewer gets bored by week three and starts waving things through (watch out for that one). The other common mistakes are picking a job so big that you can’t tell when it’s over, and letting the data rule slide because the tool is so handy.
If you’d like some reading, there’s a course from the University of Helsinki and MinnaLearn that’s made for people who aren’t experts, with no maths or programming, and you can take it at your own pace.9 NIST has a playbook to go with its framework, which isn’t a page-turner but is very handy when somebody senior asks.4 And our post on European AI regulations is a good start if you sell abroad.
The short version
- Pick one job that keeps coming round, and time it first.
- Write the finish line down before anyone opens a tool.
- Nothing private goes in, and a named person reads the output.
- On the date, decide. Stopping is a good result too.
Courses we’d suggest
The one this brief belongs to is AI for Managers & Product Leaders, and I made that course myself, so I’m a bit biased! It’s three evenings. You will go looking for chances in the work your team does, and then turn one of them into a proper pilot, and at the end you will leave with a 90-day plan that you can take back to the office.
If you haven’t used these tools much, Generative AI for Work is a gentler start, and over four evenings you will build a personal workflow for your job. When a pilot says roll it out, we’d send you to AI Automation & Agents next, where you build a working automation out of the work you do, and I think you will really enjoy it.
11.0 Sign-off
If you’d like some help
AI for Managers & Product Leaders
Funnily enough, this is what we go through in the course. I made it for people who run teams and want to know what really changes on Monday morning. Over those evenings you hunt for chances in the work your own team does, turn one into a proper trial, and finish with a 90-day plan to take back to the office.8
Approve: see the course| Length | 3 lessons of 3 hours, part-time evenings |
|---|---|
| When | Tue & Thu, 6pm to 9pm |
| Group | Small group, live in Bangkok or live online |
| Needs | No tech needed. Also runs as a private workshop for your company |
A References
- AI at Work Is Here. Now Comes the Hard Part. Microsoft and LinkedIn, Work Trend Index, 8 May 2024. 75% of knowledge workers use AI; 78% of AI users bring their own tools; 59% of leaders worry about quantifying gains. About 31,000 people, 31 markets, Feb to Mar 2024.
- Generative AI at Work. Brynjolfsson, Li and Raymond, NBER. 5,179 customer support agents; about 14% average gain, 34% for novices, little for experienced workers.
- The 2025 AI Index Report. Stanford HAI. 78% of organisations reported using AI in 2024, up from 55%.
- AI Risk Management Framework. NIST. Released 26 January 2023; voluntary; Govern, Map, Measure, Manage.
- Guidance on AI and data protection. UK ICO. Updated 15 March 2023; marked as under review after the Data (Use and Access) Act.
- Samsung bans staff from using generative AI after privacy gaffe. GSMArena, 2 May 2023, reporting Bloomberg (paywalled). “Reportedly” source code.
- NIST AI 600-1, Generative AI Profile. NIST, 26 July 2024. Companion to the AI RMF.
- AI for Managers & Product Leaders. Web Courses Bangkok course page: sessions, format, 90-day plan.
- Elements of AI. University of Helsinki and MinnaLearn. Free online courses for non-experts, self-paced, no complicated maths or programming required in the introduction.
