Can't we just build it ourselves?

The best service I ever used, I left in four hours. Not because it got worse. Because leaving got cheap.

Somebody above you is about to ask why you still pay for that. The CRM, the vendor, the agency, the programme on month fourteen. The question will sound like "can't we just build it ourselves now?" and it deserves a better answer than "it's complicated".

I have one, because I asked it of myself and acted on it. The answer was not what I expected.

The afternoon I found out what was holding me

In September I opened my own website during a repositioning and felt embarrassed. Not because it was ugly. Because it described someone I no longer was. I had moved; the site hadn't. That is the hardest work there is, reading your own label from inside the jar, and it is why I do it for clients faster than I could ever do it for myself.

The site ran on Showit, which I will say on the record was the best service I have ever paid for. Beautiful tool, human support that fixes the thing. I recommended it for years.

On 22 September I opened Claude Code, pointed it at my site and asked it to rebuild everything on my own hosting. Four hours later it was live. I cancelled.

Nothing about Showit changed that day. If anyone had surveyed me, I was a nine. The only thing that changed was the price of leaving.

Your satisfaction score matters least where customers are most locked in

That is not a provocation; it is the largest review of the question I could find. A meta-analysis in the Journal of Retailing covering 47 samples looked at what happens to the link between satisfaction and repurchase when switching costs are high. The most common finding: the link weakens. When leaving is painful, how happy customers are predicts less and less whether they stay. The literature is split, and some studies find the opposite, but the largest cluster says what I found on myself: a lot of loyalty is friction with a satisfaction score on top.

What finally made me leave was the renewal email. I had procrastinated all summer. Then "your subscription renews on" put a decision in front of a satisfied customer who wasn't deciding anything. The one touchpoint designed to keep me was the one that made me go.

I am not an exception. Three Stanford economists used transaction data from a payment card network covering about a third of US subscribers and looked at the month a card gets replaced, when the customer is forced to re-enter details. Cancellations in that month ran at four times the normal rate. They estimate that for the average subscription, revenue is about 85 percent higher than it would be if customers made an active decision each month, and they suggest some subscription businesses would not be viable if their customers were paying attention.

If that sounds like a subscription-app problem, look at an energy bill. In May 2024 CREG, Belgium's federal regulator, reported more than 1.2 million households paying up to €1,000 a year too much for gas and electricity, with the ten most expensive contracts holding about half the market and 70 percent of customers unable to name the product they are on. The regulator's own observation: even during the price crisis, most people did not question their contract. The UK ran the full investigation in 2016: 70 percent of the big suppliers' customers on the default tariff, £1.4 billion a year overpaid, and a remedy that forced suppliers to hand over the list of customers stuck for three years so that competitors could write to them. A regulator building the card-replacement moment by hand.

I am not claiming anyone decides to keep customers asleep. I am saying every number above is consistent with a business that does better when customers don't look, and that the enterprise version of the card replacement has familiar names: contract renewal, budget cycle, a new CFO.

So a satisfaction score is the sum of three different things: customers who would rebuild you, customers who can't be bothered to leave, and customers who have already left and haven't noticed. Only the first group is yours.

Klarna is the same story with more zeros

Klarna told investors it had shut down Salesforce and would shut down Workday within weeks. The headlines said "replaced by AI". The detail is more interesting: Klarna had been running most of HR on Deel already and stitched its CRM back together from other tools and in-house builds. It left because leaving had stopped being hard. Salesforce's CEO asked, in public, how that was even possible. That is the sound of a vendor measuring its moat for the first time and finding it was procedural, not emotional.

What the four hours sat on

Here is where I part ways with half my feed. The lesson people take from Klarna, and from stories like mine, is "the tools are cheap now, build it yourself". The four hours are real, but they sat on something.

Before I touched my own site I had built sites for clients with Claude Code, and because clients don't want surprises I had built a method around it: the agents, the checks, the way you get quality out of a tool that will produce garbage at speed. By September, starting over inside a page builder looked like more hassle than running my own setup. The method had been paid for, on other people's projects, before I benefited from it. Without it, the honest number is not four hours. It is a weekend, a broken contact form and a site that looks right and does nothing.

And then the part that still gets me. The rebuilt site was perfect. Every page, every form, every colour. An exact copy of a site that said nothing about what I actually do. I had built, flawlessly, the thing that wasn't working. The platform was never the problem; what the site was for was the problem, and the tool reproduced it with total fidelity, because that is what tools do. They obey.

The site that finally said what I do took two weeks. The positioning work before it, the actual deciding, I did not time, because nobody does. It doesn't look like work. It looks like thinking. No tool did a minute of it.

Four hours to move. Two weeks to say it. Longer than both to know it.

The danger isn't building something bad. It's building something true

Aalto University researchers had people do hard reasoning tasks with ChatGPT. Everyone overestimated their performance, and those who rated themselves most AI-literate overestimated the most, even when paid to judge themselves accurately. I used to read that as "people think they can code". I now read the illusion differently: not "I can build it" but "I know what I want". I had the skills and the method and still had the illusion, for about a day.

Look at Klarna again with that lens. Everyone asked whether it could rebuild Workday. Nobody asked what it would rebuild. If the answer is "our current workflows, in-house", it paid to automate what it already had.

My clients are not trying to become developers. They have no time for it and no wish for it, and they are right. What they need is not someone who understands the tool but someone who refuses to build until three questions are answered. What is this for. Who has to change for it to work. What would we see in three months if it landed. That question used to hide inside the vendor's timeline; six months of implementation gave everyone time to stumble into it. Now the build takes an afternoon and the question has nowhere to hide.

The switching cost didn't vanish. It moved, from the vendor's migration to your own clarity. People who have it can leave anyone in a day. People who don't will rebuild their problem at AI speed.

Three things to do before Friday

• The sentence for the ExCo. "Yes, we can build it in an afternoon. The question is whether we'd build what we have or what we need, and nobody in this room has answered that yet." It turns the AI question from a threat into your agenda.
• The ten-minute test for any build request. If we rebuilt this tomorrow with no constraints, would it be the same thing? If yes, you are about to automate the problem. If no, write down the difference. That difference is the programme; everything else is a build.
• The retention split. Separate the customers who would rebuild you, the ones who can't be bothered to leave, and the ones waiting for the renewal email to notice they have gone. A nine on a scale does not tell you which is which. Ask them to describe their alternatives in their own words, and look at who stayed through the last price increase and the last contract renewal.

Then the test I now run on myself and on every client. Take your best customer. Imagine they can rebuild what you give them in an afternoon, because soon some of them can. Would they build the same thing, or a better one? If the same thing, you were selling a mirror. If a better one, ask why they needed you to see it. That is what you sell. Not the system. The seeing.

If someone above you is asking why the programme takes so long, that is your answer: the build was never the slow part. The slow part is the part you can't hand to a tool, and you shouldn't want to.

What are you still being paid for that your customer could now build?

Sources

• Switching costs weaken the link between satisfaction and repurchase (47 samples): Pick and Eisend 2014, Journal of Retailing
• 60 to 80 percent of defectors called themselves satisfied (Reichheld, 1990s): Perspective, satisfaction vs loyalty
• Reichheld and Sasser's 5 percent / 25 to 85 percent claim was a thought experiment: Byron Sharp, Loyalty myths
• Card replacement: cancellations 4x; revenue 85 percent higher than with active monthly choice; 14 percent to over 200 percent: NBER w31547, NBER Digest, SIEPR
• Some subscription businesses would not be viable without inattention: PYMNTS
• CREG 2024: 1.2 million households, up to €1,000 too much; 70 percent can't name their product: Brussels Times, VRT
• CREG 2023: ten most expensive contracts hold about half the market: VRT
• Belgians didn't question bills even during the crisis: Belga
• CMA 2016: 70 percent on default tariff, £1.4bn a year, disengaged-customer database remedy: gov.uk, CMA sets out changes, gov.uk, final reforms
• Klarna shut down Salesforce and Workday; replaced with Deel and other tools plus in-house builds: CX Today, OnlyCFO, ITPro, Benioff's reaction
• Aalto: everyone overestimates performance with ChatGPT; most AI-literate overestimate most; paid for accuracy and still couldn't: Aalto University, arXiv preprint