ChatGPT launched in late November 2022. By the time GPT-4 hit in March 2023, the studio's workflow had already changed. Three months into this new era, here is the honest write-up of what we have learned, what we are doing differently, and what we are still figuring out.
What changed almost immediately
Three things shifted in the first sixty days, all at once.
Client conversations shifted. Half the discovery calls we took in January and February started with a variation of "we want AI in our product." Some of them had thought it through. Many had not. The studio's job has expanded to include a serious diligence step on what "AI in our product" should actually mean for a given business, before we scope a single sprint.
Our internal workflow shifted. Engineers started using ChatGPT-style tools daily for code generation, refactoring assistance, and writing tests. We did not mandate it; the team adopted it organically. Productivity on routine code went up. Productivity on novel architectural decisions did not, which is the right outcome.
Our pricing conversations shifted. Some clients heard about "AI" and assumed software was about to become cheap. That assumption is wrong for the kind of work we do, and we have been pushing back firmly. The architectural, scoping, and integration work where studios earn their keep has not been commoditized. The boilerplate has.
What we are actually building with AI right now
Most "AI in our product" requests, when we work through them, land on a small number of useful patterns.
The first is structured data extraction. Pulling fields out of PDFs, scanned invoices, contracts, and unstructured emails into a database. This was always possible with OCR and a lot of post-processing. With large language models the accuracy and speed improved enough that the business case for clinics, accounting firms, and legal teams started to make sense.
The second is conversational triage. Front-line patient intake for clinics, FAQ deflection for service businesses, basic customer support that knows when to escalate to a human. This is not chatbot theater. Done correctly it deflects real volume from human staff and routes the rest with better context.
The third is content drafting inside existing products. Sales reps drafting follow-up emails, marketers drafting first-pass copy, HR teams drafting role descriptions, all inside an interface that knows the business context. The model writes the first draft, the human edits, the output ends up where it needs to go.
None of this requires owning a foundation model or fine-tuning anything. The OpenAI and Anthropic APIs are good enough for ninety percent of business use cases. The work is integration, prompt design, evaluation, and guardrails.
What we are not building
A lot of "AI" requests resolve to "I want my SaaS to be the next ChatGPT." We decline these. The right answer is almost always a focused feature inside an existing product, not a standalone wrapper.
We have also declined a handful of generative-content product ideas where the output quality would have been mediocre at best and the legal exposure unclear at worst. The studio's reputation is downstream of every shipped product. We will not put our name on auto-generated content that we would not be comfortable defending.
What we are betting on next
A few directions look durable from here.
We are betting that retrieval-augmented generation will become the default architecture for AI features that touch a business's private data. Fine-tuning will remain useful for narrow domains; retrieval will be the workhorse for everything else.
We are betting that "AI as a developer" tools become standard inside engineering workflows over the next eighteen months. We are not betting that they replace engineers. We are betting they shift where the time gets spent.
We are betting that a meaningful number of products built in the next two years will need AI evaluation and observability infrastructure as a first-class concern. The shipping of an LLM feature is the easy part. Knowing whether it is still working correctly six months later is the hard part.
What this means for our clients
If you have an existing product, the right question is rarely "should we add AI?" The right question is "which two or three workflows inside our product cost the most human time, and would AI legitimately reduce that?" Most of the time there is a real answer. Sometimes there is not, and the honest recommendation is to wait.
If you are starting a new product, AI capability is now a normal part of the stack, not a special category. We will scope it the same way we scope payment processing or authentication: deliberately, with a clear purpose, sized to the actual need.
We will keep writing about what we learn as the field moves. This year is going to reward studios that are honest about what these tools can and cannot do, more than studios that promise the maximum. We intend to be the first kind.