# AI: become a good evaluator

- Published: 2025-06-01
- URL: https://notes.webutvikling.org/how-to-excel-in-the-age-of-ai
- Tags: ai

Generating decent content gets cheap. The work that remains is taste, judgment, and deciding what to make.

ChatGPT released 2.5 years ago, and I feel now that it's now quite clear (even to me) that AI will change our lives quite drastically. I say _will_, because I think the changes ahead is quite a lot larger than those behind us.

This text is a braindump of how I guess things will change in the next 10 years, in the context of digital product development. Basically, its summary is that

> **Generating (decent) content becomes cheap** and we should therefore instead focus more on **architecture**, **feedback** and **verification** of work. Also, the value of **excellent** content increases, so hone your skills

This will **reduce** demand for _doers_ who clock in and out, and **increase** the demand of critical, managerial and architectural skills and people who _excel_ at what they do.

## The product workflow

I think product development mainly consists of the steps below. Build something new often require all steps, while smaller things may go straight from idea to implementation.

I've marked them with color to indicate where AI will 🟩 reduce or 🟥 increase labor/focus/costs. I'll give reasons to this later.

1. ⬜️ **Idea**: A problem that exists, and an idea for how it can be solved.
2. ⬜️ **Idea solution**: Service design (or high-level technical architecture).
3. 🟨 **Implementation design**: UX design, excl. graphical design (or high-level coding patterns/architecture).
4. 🟩 **Implementation specs**: Graphical design (or low-level technical design).
5. 🟩 **Implementation**: Writing code, or performing configurations.
   - For more complex implementations, you may have to go back to step 1 (and follow the parenthesis texts)
6. 🟥 **Quality Assurance**: Testing, security, reviews, metrics.
7. 🟩 **Support**: Documentation, articles, support requests.

### Idea ⬜️

Somewhere, someone says

> You know what would be a good idea?

It can be for a new product/company on a whiteboard, a new feature from customer feedback or a bug found in Sentry.

We do not need a larger _quantity_ of ideas, and AI won't excel at the _quality_ of ideas. At best, AI can help bring more objectiveness to _prioritizing_ our ideas.

### Solution ⬜️

In service design and high-level technical architecture I think there's still to many case-specific considerations
to take, for AI to come up with the solutions that's worthwhile listening to.

- Who will have process ownership of applications for parental leave? (Service design)
- What's the DB technology to use for the identity service? (Technical architecture)

The most correct generic (AI) answer is HR and Postgres – but it depends on the context. Maybe applications are processed automatically, or we ought to use WorkOS Enterprise SSO for identity management.

Since we're working on a high abstraction level, generating the _content_ isn't what's costly today – it's the thoughts put into the decision.

We don't need AIs ability to generate a big _quantity_ of content, and I think skilled employees will make higher _quality_ decisions.

### Design 🟨

(UX – Figma sketching)

### Specs 🟩

(Graphical design / Figma detailing)

### Implementation 🟩

(coding – Cursor)

### QA 🟥

(testing – )

### Support 🟩

## Personal tech advice

### Don't become an implementer

I know there's a whole lot of developers who will take well specified tickets from JIRA, and implement them.
If that's the only value you bring to work, you're in trouble. AI will be taking your job.

Instead, try to get in on making the tickets.

### Use what's popular

Personally, I want to encourage trying and playing with technology. But with AI you will
gain speed by using frameworks, libraries and patterns simply because they are popular.

Because popularity = more online content = more training data = better AI coder.

[Django](https://www.djangoproject.com/) has 20 years of popularity and online content, strong patterns and amazing documentation. Django developers will therefore get more "correct code" from AI than the average developer.

So be (more) biased towards popular solutions.

### Increase consistency

Understanding code becomes harder, since someone not on your team wrote it (AI).
So anything you can do to make code more understandable on an architectural level
will help you more than previously.

_Note: A nearby trap is early over-optimization. But consistency in 2030 is more important than in 2020._

### Excel at what you do

It's a cliché advice, but it will be more relevant than ever.

Most work is mediocre. If AI is able to produce mediocre+ work, it will make paying a person for this work obsolete. So stop doing mediocre work: either pass it on to AI or start excelling at it.

### Use AI for boring shit

Just a different phrasing for the above. You'll be mediocre at what you find boring. AI is cheaper than you at doing mediocre work. So offload your work to AI before your boss offloads you to AI.

### Give better feedback

Most work-days are filled with _getting stuff done_, and I think we need to start spending more time _evaluating_ what has been done.

Because AI will get stuff done with mediocre quality, so we need to get good at evaluating these implementations, understand what and why something will or won't work – and be great at giving AI the feedback on how it could've been better.

### Implement better QA

It's incredible how large a product becomes after a while. There's so many pages, features and interdependencies.

Now imagine the future where you've first prompted Figma to design a new feature, which with an addon also has implemented it in your code base. (I expect this future to be near)

You've likely QAed the design by looking at it in Figma, or even click an interactive demo there, but **how do you QA the implementation**?

1. You don't. Trust AI to do its job.
2. You read, understand and approve up the code.
3. You click around the app to look for bugs.
4. Automated tests assist with (or replace) 3 ⬆️

While option 1 might work for the first couple of weeks in a new codebase, I have little
faith that this is viable for much longer than that. The product will become a trashfire of bugs. I think many disagree with me here, and I'll call them naive.

Option 2 will be chosen by many dedicated developers for a while, and I am afraid
it will grind their dedication down by the sheer boredom, so they end up at 4.

I bet option 3 will be selected by a bunch of dinosaur companies who will convert
much of their tech team into QA-testers with a sprinkle of vibe bug-fixing. This job will suck, and the amount of work will be costly enough
to push everything towards option 4.

So **option 4 is where we'll land**. And how do we make this reality better? Automated verification that (unrelated) behaviour is unchanged. Easy QA that new behaviour is as expected.

For instance, generating Storybook stories, connecting [Figma to Storybook](https://help.figma.com/hc/en-us/articles/360045003494-Storybook-and-Figma), running browser tests with [Chromatic](https://chromatic.com/storybook), adding [Playwright](https://playwright.dev/) e2e tests.

### Consider e2e testing

(See previous point) I am not advocating unit testing or high test coverage. But I believe vibe coding will introduce much more bugs than previously.
High level tests will therefore provide more value than before – because you can ask the AI to perform changes within a boundary ("tests must still pass afterwards").

At the same time I think small unit tests can be of hinderance to refactorability/change in the code base. So maybe it's time for those e2e-tests?

### Beware of cutting 3rd party software

[Not-invented-here syndrome](https://en.wikipedia.org/wiki/Not_invented_here) will become a more dangerous disease than ever, for two reasons: It's tempting and easy to create something inhouse + it's human to underestimate the complexity of a problem.

> Do what you do best and outsource the rest

AI will introduce a lot of slight differences and more efficient ways of doing _"everything"_ in tech. So workflows will change. This sucks if you have
your own company code nested up in every part of the organization.

For instance, Intercom is using AI to connect its chatbot to serve your customers based on the articles you've created in the support center.
Have you built your support center or chat bot internally? You will have to invest time or loose out on these benefits.

AI Antichurn tool [Revatto](https://revatto.com) integrates into Chargebee to automate anti-churn effort, and Avalara can handle taxes in different areas.
Have you built your own billing system? You will have to invest time or loose out on such benefits.

The [RelevanceAI](https://relevanceai.com/function/marketing) might automate some of your sales team – as long as you're running an internal support ticketing system. You will have to invest time or loose out on such benefits.

Migrating to (or from) 3rd party software takes more time the larger you are. So you might want to start chopping off code today.

## Societal changes

### Elders gets scammed

Many elders has had problems getting scammed with technology, phone sales and fake invoices for quite a while.

The problem will grow, since e.g. bad actors will be able to write or [speak with a convincing tone using AI](https://www.instagram.com/reel/DKPu4MppDnj/)-generated text or voices – maybe even using the voice of their victims children, generated from online content.

**You should** watch out for your parents.

**We need** more secure default messaging channels and better age-adapted technology, e.g. identity verifications of whoever contacts you, or better banking tech.

**Who fixes this?**
Apple. They launch a `Safety` set of features that allows us to increase our parents security settings.
Here, callers must provide biometrics identification and unknown numbers are blocked. Luckily, no AI is needed for this, as Apple won't have successfully embraced it by then.

Banks may also fix this, by selling "Scam-proof" bank services. In order to deliver that, they
might use AI to phone up your granny for a little chat if she tries to transfer abnormal amounts of money.

### Read more critically

The last 15 years has been terrifying in how we relate to information.

We no longer consume the same(ish) media as our neighbors and colleagues, making it
harder to relate to people in our proximity and diving societies. Truthfulness has become
something subjective, with phrases like _MY truth_.
Objective truth does not even phase the owners of "_Truth_ Social".

6+ hours of daily mobile usage, financed by personalized ads on social media,
lets others buy our political opinions. [Cambridge Analytica affected hundreds of elections](https://www.bbc.com/news/world-43476762) this way.

With AI, this problem to become bigger. The price is lowered for generating well-formulated, convincing information, regardless of it's correctness.
I believe we will fix the hallucination problem – but I don't think we'll be able to stop
questionable actors from generating persuasive, but false, content.

**You should** be vary of what information you consume.

**We need** to be able to verify where information originates. And assistance in evaluating truthfulness. And culturually take a step away from the personalized swipe-hole that social media has become.

**Who fixes this?** People _want_ to hear what they _like_ to hear, and won't be paying for tech to tell them that they're about to loose grasp on reality – so it's hard to start small. I think the push comes from big tech/media companies or EU.

> ✅ httrspct:// "This website is respected by other well known brands you already know and trust"

I also hope governments can start putting limitations on phones in schools.

### Customer-relations layoffs

Our current shift in AI is driven by the **L**arge **L**anguage **M**odels, which is made possible
by the internet being a huge dataset of text, image and video. I think it's a safe bet to say that the jobs
that are affected the most by this, are those where these forms of communication is bread and butter:

- Anyone in support / call centers
- Account managers / customer success
- (Low-level) social media representatives
- Generic sales / low-hanging marketing

The good news is that you'll be able to do so much more. You, as 1 individual,
will be able to cater for 10x more customers in support. AI will respond to email
and comments made by the majority of customers, in a patient and polite manner.
Generating posts for your social media? It'll excel at that. So one single marketer
in 2035 will be able to do 10x the amount of work.

The cliché misconception of technological development is that if something makes you 2x more productive, the business will fire 50% of its employees.
And that has usually not been the case: instead the company does 2x the work, either by increasing quality, quantity or offering.
I think that we with AI can generate content that surpasses the quality of the bottom 70% of current work being done. So we need humans to perform better to let them keep their jobs. Below-average performers will be let go.

**You should** excel at your job if you work in one of these sectors, and learn to adopt AI tools. Otherwise prepare to do a different job within a few years.

**We need** to get more creative, in order to stand out from the crowd. As individuals, excel at these tasks or acquire new skills. As companies,
we need to get on this cost-cutting train. Or go the other way, and use our _human_ touch as a
selling point.

**Who will do this?**
Many different actors. I can imagine it's exciting days for Intercom – who may be disrupted by a startup and go die slowly like an Oracle purchase; or start selling a complete Call-center-as-a-service, replacing your current support department.

## Disruptions

### Unemployment waves

Large companies pay a lot of people to produce decent tickets, slides, code, designs and emails, and AI does that work for less.

Earlier I said 2x productivity has meant 2x output, and companies kept the staff. Large companies will break that pattern. They hired more people than the work required, and shareholders will take the savings. McKinsey will put "AI transformation" on the slides that announce the cuts.

Microsoft, Meta and Google already cut tens of thousands while revenue kept growing. With time, I expect other will copy that.

They keep a smaller, more senior staff: people who decide what to build and check that it works. Ticket implementers get cut first.

**You should** get out of jobs where your value is "I completed the ticket." Own a problem, own a system or get so good they can't ignore you.

### More startups

One person can do what a small team used to do, so the cost of starting a company drops.

I expect two-person companies with a Stripe account and Cursor, selling something that used to need a 15-person product team. Most of those products will be mediocre, and a few will eat a product line that a large company used to own.

People leaving those companies will found some of them, since the remaining implementer jobs are shrinking.

You'll spend your time on taste and finding users, because a decent first version is cheap.

**You should** start a startup or a side gig.

### Increased class divide

If decent work is cheap, owners and the people who excel take the money. A 10-person company with the revenue of a 100-person company does not hire the extra 90. The people paid to produce mediocre get fired.

I don't think a course called "how to excel" turns the bottom 70% into the top.

**You should** get equity, or get so good they can't fire you. A replaceable knowledge job is a worse position than it was in 2015.

**We need** more than retraining slogans for the people whose jobs go away. Either governments tax the surplus, or we live with a larger underemployed group.