2025 was the year we put AI agents into real operations for our clients. Lead qualification, ticket routing, follow-up emails: five projects, eight agents, a year of data. What worked.
Three patterns that paid off
1. Automatic lead qualification
An agent reads contact forms, classifies by service, sector, urgency. Result: sales only sees pre-qualified leads with a 1-10 score. Conversion rate +28% vs pre-AI, because sales only talks to those worth talking to.
2. First-line ticket routing
The agent reads incoming email tickets, classifies them (bug / feature / billing / other) and routes to the right team. First response time: -65%. Routing error: 6% (vs 12% of manual routing by the prior operator).
3. Contextual follow-up reminders
The agent reads a client's history and suggests "what to say today" to sales. Nothing magical — a digest based on meeting notes, deadlines, open topics. Sales actually uses it.
Three patterns that failed
1. Automatic customer replies
"AI replies directly to the customer" became "AI replied badly and now they're angry". Moved to "AI drafts, human validates and sends". Acceptable error rate: 5%, but high cost when it happens.
2. Automatic sales call summarization
Nice in theory, disappointing in practice. Summaries lose the needed details and introduce fluff. Replaced with an assistant that, on demand, extracts specific data (next steps, pricing discussed, objections).
3. Predictive churn without quality data
For most clients, the data needed for predictive churn isn't there or is poor. The agent "did" predictions but with random outcomes. Suspended.
2026 lesson
The agent doesn't replace people — it replaces work people didn't want to do anyway. Routing, qualification, dispatching: yes. Final customer communication, strategic decisions, empathy: no.