unrealistsFinding opportunities vs fixing one task
These are two jobs. One is a short, messy look across the business: where time goes, which piles of files exist, what would even be legal to try. The other is a named task with an owner (“extract requirement IDs from this audit pack”). We get asked to do both in the same 4-8 weeks. That is how you end with a chatbot nobody wanted and no number anyone trusts.
How we tell them apart on the first call
If they can point at a process, a person who hates that process, and a stack of examples, it is a task. If they say “we should do something with AI” and the room disagrees on whether that means invoices, the helpdesk, or the next board deck, it is exploration. “Chat with all our SharePoint” is exploration wearing a task costume. There is no owner and no definition of done.
Other tells. No process owner in the meeting: exploration. They already time how long a reviewer spends per file: task. Twelve stakeholders and a strategy slide: exploration. One team lead with a folder of ugly PDFs: task. If we guess wrong, we say so in the first week and switch. We do not quietly turn a workshop into a half-built assistant.
What exploration is for
A two- to four-hour workshop with people who actually do the work, then a short written list: a handful of candidates, where the data lives, who would own a pilot, and which items are a script or an integration, not a model. Typical output is “do this invoice extraction next, park the chatbot, forget the voice clone.” Timebox is days, maybe two weeks if we have to sit with three departments. It is not an eight-week build.
Exploration does not produce accuracy numbers. We might open three files or listen to one call to see if the data is even usable. That is a sniff test. Scoring belongs in a PoC after one item is chosen. If the workshop ends with forty use cases and no owner, we failed. Cut to one or two, or stop.
What a task looks like
One sentence the business already knows how to argue about, samples they can hand over, and a metric they already use. Then we run the PoC path: holdout, cost, go/no-go. We refuse to add “and also a chatbot on the same corpus” in week three. That is a second project. It can go on the exploration backlog for later.
If they arrived with a task but the samples are a political mess (legal will not release files, IT will not connect the ERP, two departments own the same process), we back up into a smaller exploration: unblock access, then restart the PoC clock. Related: How we score a 4-8 week PoC, Local inference vs an API.