10 Real-World Case Studies

Situation. A global carrier handled 12,000 rebookings on a typical disruption day. Each one took an average of 14 minutes of agent time. Customer satisfaction during disruptions was their weakest scorecard metric.

Approach. They built an agent that read the disrupted booking, checked customer tier, fetched alternative inventory, applied policy, and proposed three options. Action policies required the customer service agent to confirm before issuing tickets.

Result. Average handling time dropped to 4 minutes. NPS during disruptions improved by 18 points. The agent operated on rung 3 of the human review ladder (one-click approval).

Situation. A home goods retailer launched an AI assistant for store managers. Three teams used three different definitions of available inventory. The assistant gave three different answers.

Approach. They paused the rollout, built a small semantic layer with twelve agreed definitions, and made the assistant consume them on every answer.

Result. Adoption tripled within a month. Inventory-related complaints fell sharply. No model change required.

Situation. A health insurer wanted an agent to triage claims faster without losing the human touch on sensitive cases.

Approach. They used the Human Review Ladder. Simple claims auto-approved. Mid-tier required one-click approval. Sensitive claims always reached a human regardless of rules.

Result. Throughput up 4x on simple claims. Complaint volume on sensitive claims dropped because the agent never auto-processed them.

Situation. A bank's customer support agent used the most expensive model for every question. 78% of questions were simple status lookups.

Approach. They added a cheap router. Simple questions went to a small model. Complex ones escalated to a larger model. They added a semantic cache for the top 200 questions.

Result. Cost per resolved chat dropped 74%. Latency improved. The big model handled fewer, harder questions and looked smarter than before.

Situation. A regional healthcare provider built an assistant that suggested differential diagnoses and summarized patient history.

Approach. Context: a clinical ontology. Control: prescribing required physician confirmation. Cost: a router between general and clinical models. Choice: open formats for evaluation and audit.

Result. The assistant launched on schedule and passed an external regulatory audit on first attempt.

Situation. An HR team wanted to draft performance review summaries from manager notes, ratings, and goal data.

Approach. They built an agent that drafted summaries but could never publish without manager approval. Action control blocked any direct write to the HR system from the agent.

Result. Time per review dropped by 60%. Compliance was satisfied because no agent action reached an employee record without a human in the loop.

Situation. A consumer goods company's forecasting agent burned tokens by pulling pages of policy and notes into every prompt.

Approach. They moved to structured-first retrieval. The agent queried a structured planning database first and only fell back to documents when needed.

Result. Tokens per answer dropped 60%. Accuracy improved slightly because the agent stopped grabbing irrelevant text.

Situation. A manufacturer ran analytics on a proprietary stack. When a faster query engine appeared, the migration looked like a year of work.

Approach. Years earlier, they had standardized on an open table format. They pointed the new engine at the existing data.

Result. Migration took weeks, not a year. The dividend of openness paid back the small ongoing tax of years prior.

Situation. A SaaS company shipped a support assistant. Quality was good but adoption was flat. Users did not trust it.

Approach. They added three signals to every answer: the source link, the definition used, and the freshness of the data.

Result. Adoption doubled in six weeks. Accuracy was unchanged. Trust, not intelligence, had been the bottleneck.

Situation. An e-commerce company's AI workflows were tightly coupled to a specific model vendor. A better model launched from a competitor.

Approach. They had built every AI workflow against a thin internal SDK. The vendor was a plug-in behind the SDK.

Result. The swap took a long weekend. Their nearest competitor needed six months to do the same.