Boards and competitors make AI feel mandatory. Shipping a chat box next to every screen is easy. Shipping AI that changes a customer outcome — and that you can operate without lighting money on fire — is not. We already wrote about production RAG versus wrappers, and about agentic coding needing verification. This piece is earlier in the funnel: should this product get AI at all right now, and what must be true before you fund it?
Add AI when you can name the workflow
Good AI starts with a sentence like: “Every week our users spend N hours doing X; a correct automated assist would remove the repetitive half.” Bad AI starts with: “We need ChatGPT in the app.” If you cannot point to a repeated decision, document pile, classification task, or drafting burden your users already hate, you are decorating — not productizing. Decorations still burn tokens, support time, and trust when they hallucinate.
Five gates before you build
1. Data you are allowed to use. Rights, PII, retention, and tenancy. No clear story means no retrieval feature. 2. A success definition. Precision/recall, time saved, conversion lift, or ticket deflection — pick one you can measure on a golden set. 3. A failure mode users accept. Wrong answer with citation to open? Force human approve before send? Refuse when retrieval is weak? Design the miss, not only the hit. 4. Cost model. Tokens per successful job, caching, model routing, per-tenant caps. Demo margins lie. 5. Ownership. Who maintains prompts, evals, and incidents after launch — the same seniors who ship, not a rotating bench.
When to wait (even if competitors shipped a bot)
Wait if core product reliability is still shaky — AI will amplify chaos. Wait if you have no evaluation harness and no one accountable for regressions. Wait if the only “requirement” is a pitch deck checkbox. A thin wrapper can win a demo and lose the next six months of roadmap to cleanup. It is rational to ship a narrower deterministic automation first (rules, search, templates) and reserve models for the ambiguous slice. Many “AI” wins are fifty percent good product design.
A practical sequencing for SaaS teams
Start with assistive features inside an existing screen — draft, summarize, classify — with human confirmation. Add retrieval only when document sets and tenancy are real. Add agents only when tools are permissioned and audited. Instrument cost and quality from week one. Expand autonomy after evals stay green. That sequence is slower than a viral launch video. It is how you avoid becoming the case study about the chatbot that emailed the wrong customer.
Questions for your next planning meeting
“Which user job gets faster in a way we can measure in 30 days?” “What data is in-bounds?” “What does a wrong answer cost us?” “Who owns evals after launch?” “What is the monthly token budget at 10× usage?” If those answers are vague, the AI initiative is not ready — the market trend does not override your stage.
Signals you are ready vs signals you are performing
Ready: users already paste content into ChatGPT outside your product; support tickets cluster around summarization or classification; you have a labeled dataset or can build a golden set in two weeks; legal has cleared the data path. Performing: the roadmap item is titled “Add AI”; success is defined as “launch before the conference”; no owner after the sprint; the only eval is “the demo looked smart.” Performing AI creates a permanent support surface and a credibility tax when answers are wrong.
Budget the boring half of the work
Half of a serious AI feature is not the model call. It is retrieval hygiene, tenancy, prompt versioning, eval fixtures, observability, cost alerts, and UX for uncertainty. If your estimate only covers “wire the API,” multiply it — or cut scope until the boring half fits. Also budget for rollout: feature flags, cohort pilots, and a kill switch. Shipping to 100% of tenants on day one with no quality bar is how theatre becomes an outage narrative.
How CodeFlamme scopes these decisions
We start with the workflow and the failure mode, not the model card. If the job is better solved with search, templates, or rules, we say so. When AI is justified, we wire tenancy, evals, and cost controls into the first vertical slice — the same discipline we apply to production RAG and agent verification elsewhere on this Insights channel.