Start with the queue, not the platform
The easiest way to waste time with AI is to begin with a tool. The better starting point is the work that keeps getting postponed: a list that never gets cleaned, a set of sources nobody has checked, follow-up that disappears between meetings, or a document that needs structure before anyone can act on it.
That work already has a customer, an owner, and a consequence. It is a better candidate than a vague ambition to “add AI.”
“Start with one problem, not one platform.”
Make “done” visible
A task is not ready for automation until a person can recognize the finished state. “Research the market” is not a finished state. “Return a table of named companies, source URLs, and a note for every unresolved record” is closer.
Good scope answers four questions:
- What goes in? The sources, documents, systems, or facts the work may use.
- What comes out? The artifact a person can inspect and use.
- What counts as correct? The acceptance checks, source requirements, and exception rules.
- What is not allowed? The systems, data, decisions, and external actions outside the boundary.
Use AI where repetition is expensive
AI is valuable when the work contains a lot of reading, comparison, extraction, classification, drafting, or routing. It is less valuable when the real problem is an unresolved decision or missing authority.
Let the machine handle the repeatable surface. Keep the judgment surface visible. A person should be able to see what was assumed, what was found, what failed, and what needs a decision.
Verification is part of the job
Verification cannot be bolted on as a celebratory step after the work is already called complete. It belongs in the scope from the beginning.
- Require a source or trace for each important claim.
- Check the output against the requested format and acceptance criteria.
- Separate verified records from unresolved or ambiguous records.
- Never convert “the model produced text” into “the business problem is solved.”
This approach may return fewer rows, fewer automations, or fewer confident answers. That is not a defect. It is the difference between a useful result and an expensive hallucination.
Design for the exception queue
Every real workflow has cases that do not fit. A supplier has two names. A document is missing a page. A permission expires. The source disagrees with itself. Systems designed only for the happy path hide these cases until they become somebody else’s emergency.
Make exceptions visible and actionable. The best AI workflow often does not remove the human from the process; it gives the human a smaller, better-defined queue.
Only then ask whether it should become a product
A successful experiment is not automatically an offer. Before productizing, ask whether the outcome is repeatable, whether the boundary can be explained, whether verification is affordable, and whether the result matters enough for someone to pay attention—or money.
GNI’s public rule is simple: experiments support the ideas; they are not automatically products. A capability becomes an offer only when the outcome earns that status.