“Just add AI” is not a brief. It is a mood.
I hear it after a demo video, a board meeting, or a competitor’s chatbot screenshot. The ask sounds small — sprinkle intelligence on the store — and the implied budget sounds like a plugin fee. If I nod along, I inherit an unbounded science project and a client who thinks we agreed to magic.
I still want the work. So I say no to “just add AI” by reframing it into a scoped job with a cost ceiling. This is the playbook.
What “just add AI” usually means underneath
Decode the mood into jobs-to-be-done. Common translations:
- “Our support volume hurts” → deflection + handoff assistant with policy retrieval
- “We cannot write descriptions fast enough” → catalog enrichment pipeline with metafield writes and QA
- “Search feels dumb” → embeddings + hybrid retrieval on product data
- “We want to look innovative” → a thin, measurable feature — or a polite no
If the only success metric is “we have AI now,” I will not find a responsible build. Vanity features become support liabilities.
The reframe script I actually use
I do not lead with a lecture on tokens. I lead with choices.
I can help — and I need to turn this into a product decision. Which outcome matters in the next 60 days: fewer support tickets, faster catalog publishing, better on-site answers, or something else? Pick one primary. We will price that pipeline, including model spend, and set a monthly ceiling. If we cannot name the outcome and the ceiling, we should not start.
That paragraph has saved more relationships than any technical deep dive. It invites the client to be a buyer of outcomes, not a collector of demos.
Five numbers before any prototype
When the conversation is ready, I ask for (or help estimate) five numbers — the same discipline I use in AI API proposals:
- Tokens in/out per successful job
- Monthly volume including retries and failures
- Default model vs upgrade path
- Extras: embeddings, vision, logging, evals
- 3× demo case — what the bill looks like if usage spikes
Then we put those into a shared estimate (often via the AI API cost calculator) and write a ceiling into the SOW. No ceiling, no build.
Scope shapes that are easy to say yes to
I keep a menu of “small enough to ship” AI jobs:
- Tag and route support email intents to the right mailbox
- Generate first-draft descriptions into metafields for human edit
- Answer FAQs from a retrieved policy with hard refusal when empty
- Extract fields from vendor PDFs into a review queue (not straight to production)
- Summarize long internal threads for a store ops channel
Each has a clear input, output, owner, and failure mode. “ChatGPT on the site” does not.
How to decline without burning the lead
Saying no to the unbounded version is not saying no to the client.
Patterns that work:
- Offer a discovery sprint: paid, time-boxed, produces the five numbers + recommendation, not production code
- Offer a single vertical slice: one template, one intent, one catalog collection — not the whole brand brain
- Name the risk in writing: hallucinated policy, unbounded token spend, half-integrated widget tax on page speed
- Propose the non-AI alternative when it wins: better help center, better metafields, better search filters
Clients who only wanted a checkbox will leave. Clients who wanted leverage will lean in. Both outcomes are acceptable.
Protecting the relationship when stakeholders disagree
Often marketing wants a public chatbot tomorrow and ops wants zero wrong answers. I put both requirements on the same page:
- Public bot is allowed only with retrieval + refusal + handoff
- Internal draft tools can be looser with human publish gates
- Budgets are separate line items so one team cannot silently burn the other’s runway
Disagreement becomes a prioritization meeting instead of a vague “AI project” that fails everyone.
Contract language that prevents mood-scope
I include lines like:
- Out of scope: unbounded conversational agent across all policies without retrieval
- Model spend billed to client account / passed through with a monthly soft cap and alert
- Changes to prompt length or feature surface require a written estimate update
- Success metrics defined up front (e.g., % of tickets auto-tagged, not “feels smart”)
This is not bureaucracy for its own sake. It is how freelancers and small studios survive contact with hype cycles.
A short story without fake clients
I have taken calls where the ask started as “just add AI like [big brand].” After the reframe, we shipped a narrow FAQ assistant with a spend alert and a human escape hatch — or we stopped at a discovery note that said “fix your help center first.” Both were professional outcomes. Only one looks exciting on social media. I optimize for the invoice and the support queue, not the screenshot.
Red flags that I pause the sales process
I slow down or walk when I hear:
- “We will figure out the use case after we integrate the API key”
- “It should work like ChatGPT, but on our brand”
- “Unlimited questions for customers, we will see about cost later”
- “Can you also make it train itself on all our emails this week?”
- “Our intern prompted something cool — just productionize that”
Those are not evil people. Those are underspecified projects. Pausing is cheaper than a refund conversation.
Internal champions need ammunition
Sometimes the person on the call already agrees with scoping, but they need language for their boss. I give them a one-pager: outcome, non-goals, monthly ceiling, risks, and a 2-week vertical slice. That document travels better than my accent on a noisy Zoom.
Make the champion look prudent. Prudent champions approve more good work over the long run than hype champions who greenlight everything once.
After you say yes to a scoped job
The playbook does not end at the signature. Week one should prove the ceiling is real: logging on, alerts on, eval set started, and a demo that uses production-like prompts — not a hand-picked happy path.
If the first week already threatens the monthly cap, we redesign before we “temporarily” raise the limit. Temporary limits have a habit of becoming the product.
The principle
Say no to the unbounded mood. Say yes to a scoped job with a cost ceiling.
If you are the stakeholder writing the brief, do the reframe before you hire. If you are the builder, do not be afraid to lead it. Clarity is part of the craft — especially from Karachi, where I cannot afford to spend a week on a demo that was never a project.
When you are ready to turn a mood into numbers, start with volume and tokens in the AI API cost calculator, then write the ceiling into the proposal before a single prompt hits production.