Practical notes from the engineers who build agents, copilots and AI products.
TCO includes inference, retrieval infrastructure, monitoring, evaluation upkeep, model changes and the people who own the system. The build price is the smallest line over three years; the recurring lines are what decide whether a system pays for itself, and most budgets leave half of them out.
Reporting from a sentence, bulk changes, drafting, summaries, next-step suggestions and search top the list; start with three done well. The ten jobs below come from what users type into copilots once they exist, not from what roadmaps imagine, with what each needs from your product and how hard it is to build.
The strangler pattern wraps the legacy system with a façade and replaces modules one at a time behind it, each with rollback. Traffic for each module is switched to the new implementation only after it has run in parallel and matched the old one, so the business never stops and no single cutover can take it down.
A semantic layer defines metrics, joins and time periods once so natural-language questions map to correct, governed queries. Without it the model guesses which column is revenue and which join is right; with it the model picks named metrics and dimensions, and SQL is compiled from definitions finance has signed off.
Custom CRM typically starts around $28,000 and ERP higher, phased over months; the comparison is licence plus customisation of a generic product. Count the five-year total on both sides, including integrations, data migration, training and the cost of bending your process to fit a package you did not design.
The quoted fee is rarely the largest number. The hidden costs of an AI discovery sprint are internal time, data extraction, evaluation set construction, benchmarking API spend, and the cost of acting on the answer. Together they usually exceed the fee, and none of them is a surprise if you budget for them.
A ten-day sprint benchmarks models on your data, sketches the architecture, estimates cost and ends with a go or no-go and a fixed price. It replaces months of vendor conversations with evidence: which model clears your accuracy bar, what the monthly bill looks like, and whether the build is worth funding at all.
The same approach works for MongoDB by generating validated aggregation pipelines over a semantic model of your collections. The model selects governed metrics and dimensions, a compiler emits $match, $group and $lookup stages, a validator rejects writes and unindexed scans, and the pipeline is shown with the answer.
Enterprise buyers ask whether prompts, retrieval and logs are isolated per tenant and whether data trains models; the answer must be yes and no. Tenant isolation for AI in SaaS means every retrieval index, prompt context, cache and trace carries the tenant identity, and no provider is permitted to train on the traffic.