The Hardest Part of AI Is the Change, Not the Technology

On 23 September, Microsoft wrote to its customers and said something software companies rarely say about their own market: "The hardest part of this isn't the technology. It's the change."

For most businesses, that settles the question of which AI tool to buy. You have probably bought one already. What decides whether it pays off is whether your team's real work has moved onto it, and no licence does that part for you.

What did Microsoft say, and is it right?

The letter, Building the system for AI at work, is signed by Jared Spataro, Microsoft's Chief Marketing Officer, AI at Work. In short: AI can now do real, valuable work, and getting it to do that takes three things. A platform that does not tie you to one AI model. A way of paying that separates per person subscriptions from usage billing. And partnership, because the hard part is "rewiring how work gets done, and getting people to trust it", which is "not something you can buy."

We think the diagnosis is sound, and the independent evidence supports it: in Singapore's national survey, only 3.8% of firms have built AI into their core processes. The letter's most useful idea is that an AI system is only starting when it goes live, and it is the one fewest businesses act on. Separating seats from usage makes sense, as long as usage spend is tied to work you have measured. Where the letter stops short is the remedy. Engineers placed alongside your team do not come with a licence, so in most businesses someone still has to do that work. The rest of this article is about how.

The evidence: access has arrived, the work has not moved

Microsoft is spending on this, not just writing about it. In July it set up the Microsoft Frontier Company, a $2.5 billion unit of more than 6,000 people, to embed engineers inside customers and build AI systems alongside them (GeekWire, July 2026).

The clearest independent numbers come from Singapore's Ministry of Manpower, which surveyed 2,560 firms between January and March 2026, with a 97% response rate (MOM, April 2026). 28.5% had started using AI. Only 3.8% had built it into their core business processes. This is what the firms said stood in the way:

What stands in the way of AI adoption	Share of firms
High implementation costs	44.9%
Lack of skilled personnel or expertise	42.4%
Integration with existing systems or processes	37.4%
Lack of clear strategy or understanding of AI's potential	32.3%
Resistance to change from employees or low trust in AI	30.6%

None of these says the software is not good enough. The top barrier is the cost of implementing it, not the price of the licence. The rest are expertise, fitting AI into how the business already works, knowing where to start, and trust. That is Microsoft's "change", measured. The survey also shows where firms are putting their effort: those with fewer than 200 staff were mostly focused on training and getting AI tools in, while those with more than 500 had moved on to governance and redesigning workflows.

What the change looks like in practice

The tools are already in the building. Microsoft 365, Google Workspace, ChatGPT and Claude can all work with your documents and email. What usually has not happened is that someone sat down and decided, for one piece of recurring work:

• Who does what now: where the AI does the first pass, and where a person takes over.
• Where a person approves: anything that leaves the business or commits you to something.
• What happens when it is wrong: it will be, and the plan for that matters more than the accuracy rate.
• What stops: the old way is switched off, or the new one becomes an optional extra nobody uses.

Take chasing overdue invoices. The AI drafts each reminder from the ageing report, in the tone your team already uses. Someone in finance reviews the batch before it goes out, in a fraction of the time drafting used to take. When the AI cannot tell whether a payment has landed, it flags the account instead of sending. And the spreadsheet the team used to track who had been chased is retired, so there is one record instead of two.

That is a few hours of decisions per workflow, not a transformation programme. Trust follows when people watch the system handle their own cases, including the awkward ones. And ask one question before any of it: is AI the right answer here at all? Sometimes a template or one rule in a tool you already have does the job for nothing.

Launch day is the start, not the finish

This is the line in the letter most worth acting on: "The organizations pulling ahead decide what good work looks like, measure against it, and improve a little at a time until the agent outperforms anything they could have specified up front." You do not need a new product to do that. For one piece of work:

1. Write down what good looks like, in a way you can check: the right figures, sent by 10am on Monday, with no more than one correction.
2. Keep ten to twenty real examples, including the awkward ones, and run them again whenever anything changes.
3. Log every correction a person makes. After a month, that list is your improvement plan.
4. Review every fortnight: the time taken, how often a person steps in, and where the freed hours went.

Say your account managers use AI to draft the weekly update each client receives, and note every fix before sending. The list keeps showing the same two problems: last week's figures pulled in instead of this week's, and a tone too formal for long-standing clients. Fix those two things, run the examples again, and the next fortnight's list is half the length. Nobody had to guess what to improve.

This is also why a project that ends on launch day disappoints. Someone inside the business has to own the loop, and the implementation guide covers who that should be.

Stop chasing the newest tool, and stop waiting for a better one

The letter opens on the noise: "New models, record benchmarks, louder claims every week." Both of the usual reactions cost you. Chasing the newest tool means your team never gets good at any of them. Waiting for the right one means another quarter of the same work done by hand. For drafting, summarising, chasing and reporting, today's tools are already good enough. What is missing is the change, and no release does that part for you.

Starting now is also the best preparation for the next model. The written standard, the examples, the approval points and your team's habits all carry over. A business that starts this quarter can test a better model on its own cases in an afternoon and switch if it wins. A business that waited starts from nothing, just with a newer tool. Keep your instructions, templates and examples in files you own, and switching stays a small job.

Watch the second AI bill

AI now comes with two costs, and each needs judging differently.

• Seats are a flat monthly fee per person. Judge them by use: check the usage report in your admin console at 30 and 60 days, and move seats nobody opens to people who would use them.
• Usage charges for the work AI agents actually do. Judge it per finished piece of work: divide a month's usage spend by the pieces of work finished, and compare that with the hours each piece used to take, multiplied by what those hours cost you. A task that is cheap each time can still add up over thousands of runs, and an expensive one can still be a bargain.
• Set limits before the first bill. Whichever vendor you use, find the spending limit setting and set it before usage starts, not after the invoice arrives.

What good change work looks like, whoever does it

Microsoft's answer to the change is engineers who sit with your team. Whether that is someone inside your business or someone you bring in, the work looks the same. Use this as a checklist for your own plan, or for anyone pitching you:

• Start with how the work actually happens: who does what, in which tools, before choosing a product.
• Be willing to conclude that AI is the wrong answer. Some tasks need a rule, a template or a process change instead.
• Put a number on each opportunity, in hours or money, so you know what comes first.
• Set it up inside the tools you already pay for, and add new software only when those cannot do the job.
• Train each person on their own work. It is not finished until people use it.
• Stay for the improvement loop. Someone has to own it after launch.

Anyone who starts with a new platform for you to buy has it the wrong way round.

What to do this quarter

1. Check which AI licences you hold and who used them last month.
2. Pick one piece of recurring work, not a tool. The method is in how to find the AI use cases worth doing.
3. Ask whether AI is the right answer for it.
4. Write down what good looks like, and collect ten to twenty real examples.
5. Name an owner and an approval point before it goes live.
6. Set spending limits on any usage billing.
7. Review it every fortnight.

Common questions

We use Google Workspace, ChatGPT or Claude. Does this apply? All of it. The change work, the improvement loop and the two costs are the same with any vendor.

Does AI adoption mean fewer people? Mostly not, so far. Among Singapore firms already using AI, 6.2% reported reduced headcount, while 18.9% redesigned roles and 13.9% created new AI-related jobs.

We bought licences and nobody uses them. Now what? Do not buy more. Pick one piece of work and one team, and run the steps above with the licences you have. Usage follows a working example faster than it follows a training session.

Where to start

Pick one piece of work your team repeats every week, write down what good looks like, and move just that onto the tools you already have. Within a month you will know whether it is working, because you will have defined what working means.

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.