Why AI Training Does Not Stick, and the 30 Days That Make It Stick

Most AI training does not stick because it is built around the tool rather than the work. A morning on prompts, exercises on somebody else's documents, nobody deciding what the saved time is for, and nothing after the day. Within a fortnight the team is back to the old way, and the next time someone suggests training, the answer is "we tried that".

The numbers agree. Of Singapore workers who attended AI training in the past year, 62.8% say they learned about AI and 47.6% feel more confident, but only 43.3% say they could apply it in their work (IMDA, Singapore Digital Economy Report 2026). More than half came back no more useful than they left.

We sell this kind of training, so read what follows as a position rather than a neutral survey. The position: the fix is not better slides. It is thirty days built around your own tools and your own recurring work, with the manager in the room and a check at the end. First the why, then how we run the thirty days.

Is training even the right first step?

Yes, and owners already know it. Singapore's Ministry of Manpower surveyed 2,560 firms early this year (MOM, April 2026). Among firms with fewer than 200 staff that have started with AI, the most common step is internal training (46.6%), ahead of giving people tools (41.1%); external courses come a distant third (26.9%). Firms that have started are not short of software. They are trying to get the people using it.

The workers want it too: more than two-thirds say they need AI training, but only 36.7% had any in the past twelve months, and the top reason is that their employer did not send them (34.9%) (IMDA, same report).

So the instinct is right. The question is why so much of it fades.

Why does it fade?

Four things go wrong, usually together. Each is a rule in how we run the thirty days, so read them as a design brief.

It is practised on the wrong material. A public course has to work for everyone in the room, so the exercises are generic: summarise this article, draft an email to a made-up client. Your people learn that the tool can do things. They do not learn that it can do their thing, the quotation rebuilt every Tuesday or the client update written every month, because nobody opened that in the room. The 43.3% above is this gap measured.

It is taught on a tool the team does not have. Say your people spend a day on an AI assistant, then find that the version in their own email is a different product, a different plan, or behind a login they do not have. The habit never starts because the first attempt fails. The assistant bundled with a business email plan is not the paid add-on, and a personal ChatGPT account is not the company's.

Nobody is in charge after the day. The biggest one, and the research agrees. In Gallup's tracking of US employees, people whose manager actively supports their use of AI are 1.7 times as likely to use it a few times a week or more (Gallup AI indicator, 2026), and Gallup's own conclusion is that "training alone is unlikely to be enough". BCG's survey of 11,749 workers in 14 markets found that clear direction from leadership adds roughly 25 points to the measured impact of AI, against about 5 for better tools (BCG, June 2026). A manager who expects the new way, asks about it and clears the time for it is what carries the change. We wrote about that side of it in the hardest part of AI adoption is the change.

Nobody measures, so nobody notices the fade. Non-technical people try once, fail quietly and stop, so use tends to drop after the first week and again after the second unless something brings it back. If nobody looks at who is using the tool, the fade is invisible until the owner asks "are we using that thing?" and the honest answer is "not really".

The quieter fifth reason: when the tool does save someone an hour, nobody has decided what the hour is for, so it disappears into the day and the case for continuing never builds.

How do we tell whether it stuck?

We settle this before the session, because it changes what the session is for.

"Stuck" means three things: the same people use the tool week after week, they use it on the recurring work they named, and that work is now done the new way by default. The first two are visible. Every business plan from the major vendors has a usage report in its admin settings showing who used the assistant and how often: Microsoft 365 for Copilot Chat (Microsoft Learn), Google Workspace in its Gemini reports (Google Workspace Admin Help), ChatGPT Business and Claude Team in theirs. We open it with you before any training and note who is using the tool now. That is the baseline. We open it again on day thirty.

The third we ask. Thirty days after the session, each person answers three questions: what they use the tool for every week, what they tried once and dropped, and what they would want next. The "tried once and dropped" list is where the next month's work is.

How do we run the thirty days?

One team at a time, on the tool the business already pays for. One team, because the work overlaps: the people who write the quotations also chase them, confirm the orders and answer the same supplier questions, so each person trains on the tasks on their own desk, and what one of them sets up, the next can use. Sales admin, client service, finance: the shape is the same. For a small team that is everyone in one sitting; for a very small one, it is us beside each person in turn, at their own screen, rather than a session in a room.

It starts with a conversation about what your team has tried and where it stalled, and we say plainly whether this fits. If it does, the thirty days look like this.

Before day one: a week with the owner

What we do with you	Why
Check which assistant comes with your plan, and that everyone can sign in with their work account	The second failure above
Open the usage report and note who is using it now	The baseline
Ask each person what they used AI for last week and what they tried once and dropped, two columns on a page	Where the team is; the dropped list is where the session starts
Ask each person for their three most repetitive tasks, with rough minutes a week, and one line on what worries them about AI	The tasks are what the session runs on; the worries are what it answers first
Agree one sentence on where the saved time goes	"Hours saved on quotations go into chasing quotes that have gone quiet." No sentence, no saving
Agree what may go into the tool and what may not, on one page	A green list (your templates, public information) and a red list (customer personal data, anything under NDA). Singapore's privacy regulator advises exactly this (PDPC guidelines, July 2026)

Day one: the session

Short, and on their work. Ninety minutes to half a day, one team, each person at their own screen, in the tool they will use tomorrow morning. We keep it basic on purpose: most people have used these tools as a search engine, and an hour that ends with the tool set up on their own work is already a leap.

• We answer the three things people are privately thinking before showing anything: it will replace me, it will make the work less human, using it feels like cheating. Said plainly, with the manager in the room, they are answered before they become the reason nobody tries.
• One live demonstration on a task from the list: the request made bare, then the same request with the context a colleague would need, side by side. The second answer is the lesson. We include a mistake the tool makes and how a person catches it. The rule the session hangs on, said to each person: the AI drafts, you check, you own it.
• Each person works their own task on a safe version of their own material (a made-up quotation in your format, never a live customer's). Pairs check each other's output.
• What worked is saved as a shared assistant for the team, not twelve personal ones, so every draft sounds like the company.
• Ten minutes on the one-page list of what goes in and what stays out. Not longer.
• One commitment each to close: the task they will do the new way for thirty days. We name a champion with you, and the commitments go where the manager can see them.

The manager is in the room for all of it, to say in front of the team that this is how the work is done from now on.

Days two to five: one small step a day

One step a day for a working week, with a reminder from us each morning:

1. Set up the assistant with who you are and what you do, so it stops asking.
2. Point it at the documents it is allowed to see, within the green list.
3. Join the team's shared space, which the champion sets up with us on day two: the company's writing guidelines, email rules and the one-page data list inside it, so every draft starts from the same brand and the same rules.
4. Save the instructions that worked on day one as shared assistants the whole team uses: one for the quotation, one for the follow-up, one for the month-end note.
5. Automate the smallest repeated step, or build the template the task will reuse.

The steps are the same in every tool; only the names differ. Ten minutes each. The aim by Friday is a working setup for each person's own tasks inside a shared space the whole team uses, which is already more than a morning on prompts leaves behind.

Weeks two to four: the part everyone skips

• Within a day of the session, we send the notes, the saved assistants and written answers to every question asked in it.
• An hour of office hours a week for four weeks, where anyone brings the thing that did not work, plus a chat channel for the questions in between. The champion sits in, to carry it on afterwards.
• The manager asks about it in the normal weekly meeting. "How is the quotation draft working?" is enough. The question is the support.
• On day fourteen, we open the usage report. If the names match the baseline, something went wrong on day one, and it is cheaper to fix now.
• When someone finds a better way, it replaces the saved assistant for everyone. One version of the task, not one per person.

Day thirty: the check

We open the usage report next to the baseline and put the three questions to each person. You get the answers in writing, with the usage change and our recommendation: more of the team's work, a second team, or stop because it is not paying. Stopping is a fine answer when the numbers say so; never finding out is not. Run the same check yourself at three months. The habit still there then is the one you can build on.

What does it need from the owner?

Less than you might expect, and most of it from the owner or the manager rather than the team.

• An hour before day one, spread over a week: the six items in the table.
• The manager in the room on day one, and the weekly question for four weeks after. Without that the rest is a day out, and the Gallup and BCG figures above are why we do not run it otherwise.

In return, nothing to buy: it runs inside the software you already pay for, and if the first recommendation you hear from anyone is a new licence, ask what the one you have cannot do. The rules we hold ourselves to are the failures above turned around: your tool and your own work, everyone at a screen doing their own task, a shared setup so the whole team writes in one voice, and a baseline before with a check after, in writing, with a recommended next step. No numbers and no follow-up is a day out, not a change.

Common questions

Our people already use ChatGPT on their own accounts. Is that not enough? It is a start and a good sign. It is also the untrained state: everyone in their own way, on their own login, with your documents going wherever that account sends them. The thirty days move the same habit onto the company's own tool, where the setup is shared, the data rules are yours and you can see who is using it.

Which tool should it run on? Whichever one your business already pays for, with everyone signed in on company accounts. The steps are the same in each; the names differ.

What about client data? Settled on one page before day one: what goes in, what stays out, and the rule that anything leaving the business is checked by a person. The business plans of the major vendors commit in writing not to train their models on what your people type; personal accounts vary.

How long before we know? Thirty days for the signal, three months for the proof. If the day-thirty report looks like the baseline, we change the tasks or the setup before anyone changes the training.

Where to start

Whether or not you bring us in, do this on Monday: ask each person what they used AI for last week and what they tried once and dropped, two columns on a page. Open the usage report and see who is on the tool now. Pick one team, and start with the items on its dropped list that cost the most hours. That page is the first thing we ask for, so you lose nothing by starting without us.

If you would like a second opinion on where to start, book a free call. Thirty minutes, no obligation, and if the tools you already pay for can do it without us, we will say so.