Loading
Model access now costs dollars a month.
Hassan Sarwar ·
Access to capable language models now costs dollars a month through hosted APIs, so the model is rarely the constraint. The expensive, slow work is integration: connecting a model to the CRM, ERP, ticket queue and spreadsheets a company already runs, then keeping its answers correct as that data changes.
That shift matters for how you plan. A pilot that proves a model can summarise a document is not the project. The project is the plumbing, the permissions, the evaluation set and the runbook that let the summary land in the system where somebody acts on it.
Three things got cheap at once: hosted model APIs you pay for by usage, open-source orchestration tools you can self-host, and cloud compute you rent by the minute. None of them require a research team. All of them stop at the edge of your own systems.
The cost that did not fall is the cost of the last mile: authenticating against your CRM, mapping your field names, handling the records that do not match the schema, and deciding what happens when the model is unsure.
The processes worth automating first share a shape: they run often, they follow a rule a person could write down, and somebody currently does them by copying between two screens. Rare, judgement-heavy work is the worst place to start, because you cannot build an evaluation set for a process that happens twice a year.
Start by counting. How many times a week does this happen, how long does each one take, and what does a wrong answer cost? If you cannot answer those three questions, the first piece of work is measurement, not model selection.
A short assessment answers more than a long pilot. Ours typically runs one to two weeks and it is allowed to conclude that AI is the wrong tool for the process you brought. Reaching that conclusion in week two beats discovering it in month four.
The questions that decide a project are boring ones: which system holds the record of truth, who is allowed to see it, what the current error rate is, and how you will know next quarter whether the thing still works.
Pick the single process that costs your team the most hours, measure it for a week, and scope an integration against that number. One working integration in production teaches you more about your own data than six months of evaluating vendors, and it gives the next project a baseline to argue from.
If you want a second opinion on which process to pick, bring it to a working session. Thirty minutes, no deck, and a written go or no-go at the end.
No discovery marathon. Bring one process that costs your team hours and we will map what an integration would look like in a free 30-minute working session.