Most teams do not get stuck on calling a model. They get stuck before that, on what to build first. "Add AI" is too vague to start, and "rebuild search" is too big to finish, so the work sits in a planning doc and nothing ships. The fix is a small first use case that is real, owned, and measurable.
Here is a short way to choose one.
The four tests
Score each candidate use case against these. A candidate should pass all four before you build it.
1. It is a real workflow, not a feature idea. Name the workflow and where it runs. "Summarise each inbound support ticket before it reaches an agent" is a workflow. "Use AI to improve support" is not. A workflow has a trigger, an input, and an output you can point at.
2. A named person owns it. A name, not a team. If the owner is "whoever has time," the first slice will not ship. The person who owns the workflow should own the first use case, because they can say what "good" means.
3. The result is measurable. Decide up front what you will compare: time saved per ticket, summaries a reviewer accepts without edits, or transcription accuracy on your own audio. If you cannot name the measurement, you cannot tell whether the first slice worked.
4. A first slice ships in days, not quarters. The first slice should touch one workflow, one endpoint, and real data. If it needs a new database, a new queue, and a new UI, it is not the first use case; it is the second or third.
Write three to five candidates
Keep each to one line, then score them. For example:
- Summarise each inbound support ticket before it reaches an agent.
- Transcribe recorded sales calls and extract the next steps.
- Draft first-response replies for a human to edit and send.
- Tag incoming documents by type before routing.
Now mark each against the four tests. The one that passes all four and has the smallest first slice wins.
A worked example
A ten-person B2B support team writes three candidates:
- Summarise tickets before they reach an agent. Real workflow, owned by the support lead, measurable as minutes saved per ticket, first slice is one call on one ticket.
- Draft replies automatically. Real workflow, owned by the same lead, but the measurable result is harder (quality is subjective), and the first slice needs review tooling.
- Rebuild the help centre with AI search. Too big for a first slice.
They pick candidate 1. The first slice is a single call: send one
ticket to POST /v1/chat/completions and print the summary.
If that is useful on real tickets, they wire it into the queue. The
workflow maps to chat completions; a transcription-first workflow would
map to POST /v1/audio/transcriptions instead.
Start the first slice
You do not need anyone's permission to test. Read https://webwright.ai/v1, point an existing
OpenAI-compatible client at the base URL, and run one request on your
own data. If you already call the OpenAI API, an OpenAI-compatible base
URL is a base-URL and model-name change, not a rewrite. The evaluation
guide and the cost calculator on the site are there for the next step.
When a working session is worth it
A working session earns its place when the choice is genuinely open: several candidates pass the four tests, the workflow spans more than one endpoint, or you want a second opinion before committing engineering time. Bring the scored list above and the session starts where a cold call would take half an hour to reach.
The WebWright Team