Field guide

AI workflows, AI applications and the Claude vs ChatGPT question

Straight answers from a studio that ships AI software for a living — what to build first, which model to use for what, what it costs, and how to keep it safe. No hype, no theory.

What is an AI workflow, and how do I build one?

An AI workflow is a repeatable business process where a model does one clearly-scoped step — draft, classify, extract, summarise, decide — and software does the rest. Build one by picking a process that runs at least weekly, writing down the current manual steps, replacing only the judgement-heavy step with a model call, and keeping a human approval gate for anything that writes to a customer, a ledger or a contract. Start with one workflow, instrument it, and only then chain steps into an agent.

Claude vs ChatGPT — which one should I use?

Both are excellent; they are good at different things. Anthropic's Claude is strong on long documents, careful instruction-following, coding and tool use, and tends to be more conservative and steerable — which matters when the output touches money, law or customers. OpenAI's ChatGPT has a broad consumer ecosystem, strong image generation and a wide plugin/app surface. For production systems the honest answer is that the model is a swappable component: build behind a gateway so you can route each task to whichever model wins on quality, latency and cost, and re-test quarterly. Debouillet builds primarily on Claude and keeps the routing layer open.

How do I use AI at work without it going wrong?

Three rules. First, never let a model be the final authority on an irreversible action — put a person or a deterministic check in front of writes, payments and sends. Second, give it the context it needs (your docs, your data, your tone) instead of expecting general knowledge to cover your business. Third, evaluate: keep a set of 20–50 real examples with known-good answers and re-run them whenever you change a prompt or a model. Teams that skip evaluation ship regressions they cannot see.

What is an AI agent, and when do I actually need one?

An agent is a model given tools (search, database, API calls) and a loop, so it can take several steps toward a goal instead of answering once. You need one when the task genuinely branches — research, triage, reconciliation, multi-system operations. You do not need one when a single prompt plus a form would do; agents cost more, fail in more interesting ways, and demand tighter permissions. Scope tools narrowly, log every call, and cap the loop.

What is RAG and do I need a vector database?

Retrieval-augmented generation means fetching your own content and putting it in the prompt so the model answers from your facts instead of its memory. You need retrieval; you do not always need a vector database. For a few hundred documents, plain keyword search or full-text search in Postgres is often better and far cheaper. Reach for embeddings when meaning matters more than wording, and always show sources so answers are checkable.

How much does an AI application cost to build and run?

Build cost at Debouillet starts at $2,500 for a focused AI workflow or internal tool shipped in about two weeks, and $7,500 per month for a continuous product team. Running cost is usually far smaller than teams expect: most internal workflows land between $20 and $300 a month in model usage. The expensive mistakes are unbounded agent loops, stuffing entire document sets into every prompt, and re-processing the same content instead of caching it.

Which AI use cases pay back fastest?

In order of how often we see them pay for themselves within a quarter: inbox and ticket triage, document extraction (invoices, contracts, forms), internal knowledge search over your own docs, first-draft generation for repetitive writing, and QA/reconciliation checks that catch human error. Customer-facing chat is the most requested and usually the slowest to pay back — do the internal ones first.

How do I keep company data safe when using AI?

Call models server-side so keys never reach the browser, use enterprise/API tiers where inputs are not used for training, redact or tokenise identifiers before they leave your systems, scope each tool to the minimum data it needs, and log prompts and outputs so you can audit any answer after the fact. Row-level security on your database matters more than anything you do at the prompt layer.

Want one of these built instead of read about?

We scope, build and ship AI workflows and applications on Anthropic's Claude — fixed price, first release in about two weeks.

Request a build