Most companies do not need more AI ideas. They need three that work

Usually three things. The use case is chosen for measurable volume rather than demo value. There is a graded test set built from your own cases, so quality is a number rather than an opinion, and regressions are caught before users find them. And the workflow is integrated where the work already happens instead of behind a separate login. Pilots stall most often because nobody could prove the output was good enough to trust, not because the model was not capable.

In narrow, high-volume workflows with a clear correctness signal: document and invoice extraction, classification and routing, search over your own knowledge base, drafting where a human still approves, summarising long case histories, and code assistance. General-purpose assistants deployed company-wide are the most commonly funded and least commonly measured initiative we see. Value concentrates where the task is specific, repeated hundreds of times a week, and someone can tell quickly whether the answer was right.

Less than people expect for generative work, more than they expect for prediction. Retrieval systems need your documents to be findable and correctly permissioned rather than voluminous, and a useful evaluation set can be a few hundred well-chosen examples. Training a predictive model is different: there you need history, labels and enough volume for the pattern to be real. Part of the assessment is telling you which of the two you are actually attempting.

Agentic workflows hold up well when the task has a bounded set of tools, a verifiable result and a safe way to fail — retrieving and cross-checking information, triaging tickets, running a defined sequence of API calls with checkpoints. They hold up poorly on long open-ended chains where each step compounds the last one's errors and nothing checks the work in between. We scope agents to steps that can be verified, and we keep a human decision point wherever an error would be expensive to reverse.

By treating it as an engineering constraint from the design stage. Each step gets the smallest model that passes evaluation, prompts and retrieved context are kept tight, repeated work is cached, and spend is tracked per workflow so an expensive feature is visible rather than buried in a monthly total. We also re-test against newer and cheaper models on a regular cadence — that has cut running costs on live systems more than prompt tuning ever has.

We start by classifying what each system does and who it affects, since obligations differ sharply between a drafting aid and something influencing credit, employment or safety. From there it is concrete engineering: documented data sources and retention, provider terms that exclude training on your data, access controls carried through to retrieval, prompt and output logging for audit, disclosure where users are interacting with a machine, and a defined human review step where a probabilistic answer is not acceptable. We are not a law firm, and we work to your counsel's interpretation.

A first pilot on a well-chosen use case typically runs six to ten weeks from kick-off to real users, with the assessment ahead of it taking two to three. Reliable numbers arrive a few weeks after that, once enough volume has passed through to compare against the baseline. Anything promising transformation across a business within a quarter should be treated with suspicion — the work that takes time is data access, integration and review, not the model.