Engagement #1: The "AI customer service" spec
Mid-market e-commerce client, ~$40M/yr revenue, wanted "an AI agent that handles all customer service." We dug in. Their actual ticket volume was 800/month. 60% of those were three repeating questions: "where's my order?", "how do I return this?", "what's your promo?". The rest were edge cases that needed a human.
The right answer wasn't an AI agent. The right answer was a better-designed help centre with a search bar that worked, three clear answers to those three questions, and a Shopify-native "where's my order" widget. We pointed them at a $3K Shopify-app bundle that solved 80% of their volume in a week. They kept the remaining 20% with a human who now had time to do it well.
Engagement #2: The "AI scoring" platform
B2B SaaS client wanted to ship "AI lead scoring" as a marketed feature on their platform. They had no data — they were a year-old startup with 200 customers, no labelled outcomes, no consistent schema across customers' use of the product.
The right answer wasn't AI. It was a rules-based scoring engine built in two weeks, deployed to all customers, and explicitly framed in the UI as "your custom scoring rules." Once they had a year of usage data — actual conversion outcomes labelled against those scores — we could revisit. They shipped the rules engine. It still ships their scoring product today, fifteen months later.
Engagement #3: The "AI internal search"
Enterprise client wanted "AI-powered semantic search across all our Confluence + Slack + Google Drive." We asked: how does your team currently search? They said: "we don't, we ask each other on Slack."
The problem wasn't search quality. The problem was that nobody curated knowledge — most useful answers lived in Slack threads that were 6 months old and unindexed. AI search wouldn't fix that; it would surface stale outdated nonsense at higher precision.
The right answer was a documentation initiative — pick the 50 most asked questions, write actual answers, put them in one place. Three months later they could revisit AI on top of a corpus that was actually worth searching. They picked someone else for the AI build; we sent them a list of doc-ops contractors instead.
The pattern
All three engagements share the same shape: the team pattern-matched on AI as the answer before diagnosing the problem. AI fits a specific class of problem — fuzzy input, contextual reasoning, output that needs to flex per request. It's a poor fit when:
- The problem is small enough for rules.
- The problem is structural (no data, no curation, no clear job).
- The problem is human (process, communication, ownership).
Asking "should we use AI for this?" before "what is this?" gets you into trouble fast. The AI build checklist is partly designed to surface this earlier — if the first three questions don't have clear answers, you don't have an AI problem yet.
The bonus engagement we did take
A client wanted "AI for sales call analysis." We said: "build sales call analysis without AI first — get the recording pipeline working, the storage schema settled, the dashboards your team actually uses." They shipped that in three weeks. Then we layered Claude-driven summarisation and theme extraction on top, which took another two weeks. The non-AI foundation was the load-bearing part. Without it, the AI features would have been demo-ware.