Practical notes from the engineers who build agents, copilots and AI products.
For enterprise work the difference is rarely raw intelligence. Claude and GPT diverge on deployment channel, data handling, structured output guarantees, refusal behaviour and deprecation cadence. Claude suits long-document reasoning; GPT suits schema-strict tooling. Most production systems use both.
Atlas Vector Search is vector indexing inside MongoDB Atlas, so embeddings sit beside the documents they describe. A dedicated vector database is a separate service tuned only for vectors. Atlas wins when your data already lives in MongoDB; a dedicated store wins at large scale or for index features Atlas lacks.
Deployment independence is the main thing microservices buy that a modular monolith cannot. A modular monolith is one deployable unit with enforced internal boundaries; microservices split those modules into separately released services. Pay for the split when separate teams must ship on separate schedules.
Managed AI services mean somebody else runs your AI systems day to day: evals, cost monitoring, prompt regression, model swaps and incidents. In-house LLMOps means your own engineers own that rota. The difference is not capability but who is on call when a model changes behaviour overnight.
A rewrite builds a replacement system in parallel and cuts over in one event. Incremental modernisation leaves the old system running and replaces it a route at a time behind a stable facade. Choose a rewrite when the platform cannot be extended at all; choose incremental when the business cannot stop.
The cheapest AI quote is rarely cheaper work; it is usually less work. Evaluation, data preparation, integration, security review and handover are the lines that disappear first, and each one returns as unplanned spend after go-live, when it is more expensive to buy.
A brief produces accurate quotes when it answers four questions: what decision or task is being automated, what data exists and in what shape, which systems must be read and written, and how success will be measured. Everything else in a brief is context; these four set the price.
Choose fixed price when the scope can be written down and the outcome matters more than the method. Choose a retainer when a live system needs continuous care. Choose a dedicated pod when the work is real but the roadmap will change every fortnight and you want to steer it weekly.
Tokens are one line of seven. A production LLM run-rate also carries retrieval infrastructure, evaluation runs, observability, guardrail and safety calls, re-indexing, and the engineering time to keep all of it correct as models change underneath you.