Lessons learned spending 4 billion tokens on a new website with AI agents

The goal of the Wallace Corporation 2.0 build was to replace a simple HTML web app with a complete business ecosystem. I wanted to eliminate a number of our paid software subscriptions, provide our customers with useful tools and their own backend portal, publish quality content and then produce something visually interesting, fast, optimised for SEO, secure and future proof. 

We needed a clear visual aesthetic, good quality user experience and I would need a complete backend admin dashboard to maintain content and designs. The scope of the project was big and can be summarised into a few key lessons.

Choose Your Weapon

My preference has generally been for open source or lower-cost AI models that are fast, capable and inexpensive enough to use continuously.

When the project began, DeepSeek V4 Pro Preview was the best balance I had found. Towards the end of the development period, V4 Flash became more useful for a large amount of routine work. The important point was not that these models were the most intelligent models available. It was that they were available at a price and speed that made sustained development practical.

A model with a large context window that can work consistently throughout the day is often more useful than a theoretically superior model that repeatedly hits usage limits, slows down or loses the context of the project.

For the token profile recorded during this build, the difference was significant. Approximately 3.8 billion tokens cost only about A$115 using the DeepSeek usage pattern. The equivalent volume through a top-tier Anthropic model would have been approximately A$4,100 under the pricing assumptions used in the comparison.

That is an enormous difference.

This comparison should be taken with a grain of salt though - a frontier model may have needed fewer tokens, fewer correction cycles or less human review to produce the same result. Even so, lower-cost models changed what was economically possible. I could use them throughout the day, run multiple iterations and ask for repeated critique without treating every experiment as an expensive event.

I also used an OpenAI subscription periodically when I wanted another model family to challenge a design, implementation or assumption.

I use Hermes for agentic work because I do not want Wallace Corporation to have concentrated platform risk around one AI provider. The same principle applies to the software we build for clients. A business should be careful before committing its core operations to one provider, one platform or one model family.

The average business does not need frontier-level intelligence for every task. Most businesses are not developing a novel cure for cancer. They need systems that are smart, reliable, available and affordable. A temperamental and expensive genius is not always better than a capable system that can keep working.

AI is compressing knowledge work

The next lesson is less comfortable.

A large amount of work at the lower end of the knowledge economy is being compressed. Graphic design, web design, basic programming, copywriting and routine analysis are all becoming easier to reproduce with a motivated person and a LLM.

This does not mean that human expertise has become worthless. It means that the market value of generic output is under pressure.
For the average business something better than nothing. Increasingly, they can get something that is good enough with AI and avoid hiring a specialist.

That will put pressure on design and development businesses focused on low-cost, interchangeable work. A small business owner with some ambition, a modest token budget and basic systems thinking can now reproduce parts of many common business products very quickly.

The first 80% of a project can be exciting as someone vibe-codes their way to a prototype. The final 20% usually contains permissions, edge cases, data migration, accessibility, security, deployment, maintenance and the consequences of incorrect assumptions. That final 20% is where a vibe-coding legitimately is road blocked by real software engineering.

This is also why Wallace Corporation chooses not to compete by producing generic output. Our new site contains approximately 48,000 words across services, tools, articles, customer information, policies, FAQs and related content. I read and edited everything. That effort matters.

Generic AI output is often closer to filler than finished work. It is becoming easier for both search systems and human readers to identify content that was generated for convenience rather than written to be useful.

Google’s own current guidance for AI search says that existing SEO fundamentals still apply and emphasises valuable, non-commodity content, original viewpoints and useful information. Google also specifically warns against relying on supposed “GEO hacks” instead of producing strong content.

AI makes mediocre content cheap. That makes genuine editorial effort more valuable.

The SaaSpocalypse is real

Software companies should take this shift seriously.

A motivated person with modest technical skill and access to inexpensive models can reproduce core features of many common business tools surprisingly quickly.

That does not mean every SaaS company will disappear. Distribution, reliability, compliance, support, integrations and accumulated data still have real value. It does mean that businesses are increasingly able to ask whether they should continue paying for a generic platform that forces their processes to fit the software.

During the Wallace Corporation rebuild, we brought several functions into our own platform.

The customer portal now handles functions that might otherwise be spread across separate customer, project, document, referral and account-management tools. The administration platform manages content, customers, projects, newsletters, media and graphics. The site can also produce custom marketing landing pages, which reduces the need for a separate landing-page product.

We have not eliminated every external service - nor do we want to. The difference is that the new Wallace Corporation website now integrates to the key services we value while the products with a small moat have been left behind.

Custom software is becoming faster and cheaper to build. Businesses no longer need to accept every limitation of the software they subscribe to.

That is a major shift.

Database everything

The most important architectural decision was to database almost everything.

The website is not primarily a collection of static pages with text hard-coded into individual HTML files. The templates define the structure and presentation. Database records provide the content.

Services, articles, FAQs, packages, customer information, projects, referrals, contacts, graphics and newsletter issues can all be managed as structured records.

This means that a content change can usually be made in the backend without editing page source files. The same service record can be displayed on a service page, referenced in search, included in an AI knowledge base or used by a recommendation tool.

This is a familiar concept in software development, but it is often poorly understood by people approaching AI-assisted development through vibe-coding.

Generating a pretty page is easy. Designing a coherent data model to fill the page is much harder.

We use Firestore because it suits the flexible document structure of this application and integrates well with Google Cloud. It has allowed the system to evolve quickly while supporting content, customers, workflows and application state.

That flexibility still requires discipline. Firestore does not eliminate data modelling, indexing, access rules or security decisions.

The benefit is that the website becomes useful to software as well as humans.

An agent can work with a service record, customer account, project status, invoice, lead or knowledge article. It does not have to navigate an unstructured collection of pages and imitate a human copying information from one system to another.

Structured data is also important for modern search. It gives search systems and AI tools clearer information about what a business offers, how its services relate to one another and which pages contain authoritative answers.

Platform risk is becoming harder to ignore

I was a strong advocate of no-code platforms such as Bubble and Webflow only one or two years ago. Now I avoid them.

Bubble, Webflow, Base44, Replit and Lovable all provide useful ways to build sites and tools quickly. The question is not whether they are useful today. The question is what happens when a business becomes dependent on one of them.

Can the code be exported? Can the data be transferred? Can another provider run the system?

What happens if pricing changes, a feature disappears or the platform changes direction?

Owning the code does not remove maintenance. It does provide more control over architecture, migration, integrations and future decisions.

The core languages of the web have evolved for decades without being completely replaced every few years. There is a strong case for building custom solutions with simple, well-understood tools and using those tools properly.

Security is another reason to think carefully about platform choice.

Wordfence’s first-quarter 2026 WordPress threat report recorded 2,738 vulnerabilities, including 158 high-threat vulnerabilities, 9.1 billion blocked WAF attacks and 474,000 infected sites. WordPress is not inherently unsafe, but a large plugin and theme ecosystem creates a substantial maintenance and patching responsibility.

The attack surface is also changing because AI agents can act against websites and systems at increasing scale. Australia has recently seen OpenAI agents interacting improperly with government, university and public-sector websites, including attempts to bypass controls. Now it turns out this is happening all around the world. Systems capable of acting autonomously can behave outside their intended boundaries set by humans. 

Know what you do not know

AI creates a serious Dunning-Kruger problem.

The problem with vibe-coders is that they simply do not know anything about what they're actually doing.

A model can produce something that looks authoritative in an area where the user lacks enough knowledge to evaluate it. The result can be convincing, functional-looking and wrong.

The reverse is also true. An expert using AI can immediately see weaknesses that a non-expert misses.

I used adversarial review periodically when I considered it necessary. One model might generate an implementation and another would be asked to challenge it. I would then compare the criticism against the actual code, run tests and decide what to keep.

This is a useful process because different models expose different blind spots. The important part is that the review is directed by someone who understands the system and is willing to perform quality control checks and reject the output if necessary.

A detailed brief is also essential. A vague prompt produces vague work. Clear constraints, context and success criteria dramatically improve the result.

Subagents and review loops are powerful because they turn AI from a chatbot into a working process. The best results came from breaking down a problem, assigning a narrow task, checking the output and iterating towards a defined result.

Is this super intelligence?

A smart person with a smart AI system is extraordinarily powerful.

The combination increases speed, reduces the cost of experimentation and gives one person access to a breadth of knowledge that would previously have required a team of specialists.

That does not mean the AI can do everything.

The Wallace Corporation 2.0 build still represented approximately A$21,000 of billable development value. A large amount of brainpower was still required to choose the direction, identify errors, structure the data, assess the trade-offs, review the content and decide when something was good enough to release.

AI facilitated the work. It did not replace the person responsible for the work.

I am also concerned that habitual AI use could make some people less capable over time. If someone uses a model to avoid learning, checking or thinking, the tool may gradually replace their judgement rather than amplify it.

The rise of generic AI content does little to reduce that concern.

The most useful lesson from the build is therefore not that AI can now build anything. It is that a capable person can now build much more, much faster, with a smaller budget.

The practical lesson

Wallace Corporation 2.0 was built to reduce fragmentation, own more of the system and create a platform that could support customers, content, internal operations and future agents. The project showed that model cost is no longer the main barrier to experimentation. Architecture, data ownership, quality control, security and judgement are more important. For businesses considering their own AI systems, the first question should not be which model to subscribe to. The better questions are:

• Which processes are worth improving?
• Which data needs to be structured?
• Which software dependencies create unnecessary risk?
• What should remain deterministic?
• Where should AI be allowed to act?
• What requires human approval?
• Who will maintain the system when the novelty wears off?

AI can make the build faster. It cannot decide what the business should become.