Founders face a doctrine problem before a tooling problem. They need a short set of operating beliefs that decide where AI may change the business and where human judgment stays in charge — and scarce attention this quarter has to follow those beliefs, not the latest demo queue.
Tool-first work feels practical because it starts with demos. A model drafts a support reply or summarizes a sales call; each demo is plausible on its own. Stacked together, the demos invent a strategy that is only a backlog. The company gains automations and still lacks a theory of advantage.
Information systems and AI strategy research give a more durable frame. Technology matters when it cuts coordination cost or raises the quality of a named decision. Strategy work should therefore open with four questions:
- Which decision becomes materially better if the company can retrieve, synthesize, or generate knowledge faster?
- Which workflow has enough volume and repetition to justify automation, but enough variance to punish naive automation?
- Which proprietary context makes the output better than a generic model response?
- Which human judgment must remain explicit because the cost of a confident error is high?
Precision beats blanket caution. Name the decision and who owns the risk.
Those questions turn AI spend into resource allocation. A founder can rank candidate investments without treating every experiment as equal.
Customer support is a strong candidate when questions repeat and policies already define a cheap escalation path. Sales research is stronger when the bottleneck is account understanding rather than message drafting. Product discovery is weaker when customer context is thin and the team only multiplies speculative requirements. Finance and compliance often warrant assistive workflows and stop short of autonomous execution, because a confident error costs more than a slow human review.
A One-Page Doctrine
A useful doctrine states what the company believes about AI before a vendor pitch or investor question forces the answer. For an early-stage company, the first version can fit on one page:
- Compress search and drafting where context is abundant.
- Keep final calls human where trust, law, or capital is at risk.
- Attach each investment to a named business decision, not a vague productivity goal.
- Expect output quality to improve with proprietary data or operating cadence.
- Design adoption as a learning loop, not only a cheaper task.
Doctrine still needs revision over time. Its main gift is one tool lists cannot give: a clean way to say no.
Advantage goes to companies that know which decisions compound and which workflows teach the system — and that keep human duties explicit when the interface sounds fluent. The longest automation backlog is a weak substitute for that view.