AI Solutions
Built Around Your Business

What AI development means for a business in 2026

AI development is ordinary software with a model inside it. The model reads text, images or data and returns an answer. The software around it decides what the model gets to see, checks what comes back and puts the result where your people already work. Most business projects we see fall into five types.

The last two get overlooked. Not every task needs a large language model: for fast, structured decisions a System One decision model is often cheaper and easier to check, as we explain in LLM vs System One models. Computer vision is what sits behind our UnimeasureAI work, where a smartphone browser is enough to take body measurements. For a longer introduction, read what is AI development.

When not to use AI

We talk clients out of AI more often than you would expect from an AI development company. Skip it when:

• A rule does the job. "Orders over 5,000 baht need approval" is an if-statement, not a model. Rules cost nothing to run, behave the same every time and are easy to audit.
• A wrong answer is expensive and nobody checks. Models state false things fluently. If there is no validation or human review, and a mistake costs money or trust, fix the design first.
• The data isn't there. An assistant over outdated or contradictory documents gives outdated, contradictory answers. Just faster.
• A product already does it well. Meeting notes, transcription and generic chatbots are bought, not built. Build where the AI touches your own data, workflow or customers.
• Nobody can say what success looks like. "Halve first response time in support" is a goal. "Add AI" is not.

How we build AI applications

The difference between a demo and a product is in the unglamorous steps. This is how an AI application development project runs with us.

1. Discovery. We look at the problem, the people who will use the result, the systems it touches and the data you have. We write down the success measure and what a wrong answer costs.
2. Prototype on your data. A few days with a hosted model on a sample of your real documents, emails or images. You learn whether the idea works before you pay for an interface.
3. Evaluation. We collect 50 to 200 real examples with the correct answer and score every change against them: a new prompt, a new model, a cheaper model. Without that set, everyone is guessing.
4. Guardrails. Answers grounded in your documents, sources shown, outputs validated against a schema or your database, clear limits on what an agent may do, and a handover to a person when the model is unsure.
5. Cost per request. We measure tokens per request in the prototype, pick the smallest model that passes the evaluation, use caching and batch processing where they fit, and set usage limits and alerts.
6. Privacy. We agree with you what may be sent to a model, strip personal data where possible, check the vendor's retention and training terms, and design for GDPR if you serve EU customers and for Thailand's PDPA.
7. Hosting and monitoring. The application runs in your own cloud accounts, usually AWS or Supabase, with logs for quality, cost and response time. Models and data change, so quality drifts: we re-run the evaluation after every change.

We keep the model behind one interface in your code. Switching to a cheaper or better model then takes an afternoon of testing, not a rebuild.

What AI development costs

There are two budgets: building it and running it.

Building. AI work is billed like our other projects: on actual time, with a weekly progress and hours report. A senior developer is $575 a day, a lead developer $795 a day. A prototype on your data is days of work, not months. The full system depends on integrations, the interface and how much evaluation the risk demands. If you want someone in your team for longer, dedicated developers start from $7,900 a month. Everything is on our pricing page.

Running. Hosted models are billed per token, with input and output priced separately. Your monthly bill depends on:

• volume: how many requests per day;
• context: how much text goes along with each request, since long documents and chat history add up;
• the model: the gap between a small and a top model is large;
• output length, which is priced higher than input;
• agents, which make several model calls per task.

As a reference from our own calculation in October 2026: a support assistant handling 1,000 conversations a day on a mid-size model costs around $330 a month in model fees. That is usually the small part. Data preparation, evaluation, integration and monitoring cost more, which is why the cost is not zero. Hosting, monitoring and back-ups for a custom app are $350 a month.

AI projects we have built

• UnimeasureAI: an AI sizing tool that takes body measurements through a smartphone browser. We did the product strategy, branding, UX/UI, the AI model and an approved Shopify app.
• Put-it-on: we rebuilt a business network into a premium international executive platform, with a new brand, a multilingual platform, AI features and new user journeys.
• ThaiRanked: a discovery platform for local businesses in Thailand, built from scratch in 12 languages with an AI-enhanced CMS. It reached 2 million Google impressions and more than 30,000 monthly visitors in its first six months.

Our AI stack and team

We choose the model per task: hosted LLMs such as Claude or ChatGPT for language, decision models for routing and scoring, and custom vision models when images are the input. Around them we build with Python, Node.js/NestJS, SvelteKit, Next.js and React, keep data and embeddings in Supabase/PostgreSQL and host on AWS. Mobile features ship in Flutter or React Native, and we connect to Shopify, Odoo, SAP and Oracle when the AI needs your business data.

We Do Dev Work is a team of 14 developers, designers and project managers in Bangkok, founded in 2025 by Vincent Trienpont, working for clients in Europe and Thailand in English, Thai, Dutch, French and German. You own the code, the prompts and the evaluation set, and they live in your repository and accounts.

Got a process that eats hours every week? Tell us about it and we will tell you honestly whether AI is the right tool for it.