Getting Started with AI in Support

This guide is for support leaders evaluating or planning an AI rollout. No technical background required.

What AI can (and cannot) do in support today

AI excels at

• Classifying and routing tickets automatically — Triage reads incoming tickets and assigns category, priority, and team before an agent ever sees them.
• Generating draft responses for agents to review — Reply produces context-aware drafts that agents edit and send, cutting handle time significantly.
• Monitoring customer sentiment in real-time — Sentiment tracks how customers feel throughout a conversation and flags escalation risk before it becomes a complaint.
• Scoring conversation quality consistently — QA evaluates every conversation against your rubric, not just a random sample.
• Identifying knowledge base gaps — KB analyzes support tickets to find topics where self-service articles are missing or outdated.
• Providing real-time operational dashboards — Pulse gives managers live visibility into queue depth, SLA compliance, and team performance.

AI does not

• Replace human agents — it augments them. Every AI output is a suggestion that a human reviews, edits, and approves.
• Make final decisions — agents always have the last word. AI provides recommendations, not mandates.
• Work perfectly on day one — quality improves with feedback, tuning, and domain-specific data. Plan for a calibration period.
• Eliminate the need for training — agents need to learn how to work effectively with AI tools. Budget time for onboarding.

The Simpli approach

• Modular — Start with one capability and add more as you see value. Each capability works independently.
• LLM-agnostic — Swap AI providers (OpenAI, Anthropic, Mistral, or local models) with a single configuration change. No code modifications needed.
• Data sovereign — Run entirely on your own infrastructure with local models for complete privacy. No data needs to leave your network.
• No vendor lock-in — Avoid dependency on any single AI vendor or platform.

Recommended adoption path

Phase 1: Triage (weeks 1-2)

Lowest risk, highest immediate impact.

• Auto-classifies incoming tickets so agents see categories and priorities pre-filled when they open a ticket.
• No change to agent workflow required. Agents work exactly as before, just with better information upfront.
• Measurable from day one: track classification accuracy and time saved on manual sorting.
• Who is involved: Engineering (deployment), one team lead (validation of classification accuracy).

Phase 2: Sentiment (weeks 2-3)

Passive monitoring that runs in the background.

• No agent workflow change required. Sentiment analysis happens automatically on every conversation.
• Alerts team leads when escalation risk is high, enabling proactive intervention before customers escalate.
• Measurable: escalation prediction accuracy, average time to intervention.
• Who is involved: Engineering (deployment), team leads (monitoring alerts and acting on them).

Phase 3: Reply (weeks 3-6)

Biggest productivity gain, but requires a trust-building period.

• Start in "suggestion mode" where agents always review and edit drafts before sending. Never auto-send.
• Track acceptance rate (how often agents send drafts with minimal edits) to measure quality over time.
• Gradually increase trust as agents see consistently good drafts. Some teams reach 70-80% acceptance within weeks.
• Measurable: handle time reduction, draft acceptance rate, customer satisfaction scores.
• Who is involved: Engineering (deployment), all agents (training on the new workflow), team leads (monitoring quality).

Phase 4: QA (weeks 4-8)

Requires thoughtful rubric design before deployment.

• Involve agents and team leads in designing the scoring rubric. Buy-in matters more than perfection.
• Start by scoring all conversations and use the data for coaching conversations, not punishment.
• Compare automated scores with manual reviews to calibrate. Adjust rubric weights until scores feel fair.
• Measurable: QA coverage increase (from sample-based to 100%), coaching effectiveness, quality trends.
• Who is involved: QA analysts (rubric design), team leads (coaching), agents (feedback on fairness).

Phase 5: Pulse + KB (weeks 6-12)

Operational maturity through data-driven management.

• Pulse dashboards give managers real-time visibility into queue health, SLA compliance, and team utilization.
• KB gap analysis identifies topics where customers keep asking questions but no self-service article exists.
• Measurable: SLA compliance trends, self-service deflection rate, knowledge base coverage.
• Who is involved: Managers (dashboards and decision-making), content team (writing KB articles to fill gaps).

Time to value

Deployment times assume your engineering team has the infrastructure ready. "Time to measurable impact" is when you will have enough data to evaluate whether the capability is working well.

Who needs to be involved?

Cost expectations

LLM costs are usage-based, charged per token processed. Here is what to expect:

• A team handling 1,000 tickets/day with Triage and Reply typically costs $50-200/month in LLM fees, depending on the provider and model selected.
• Running local models via Ollama reduces LLM costs to zero, but requires GPU hardware (a single NVIDIA GPU with 24 GB VRAM is sufficient for most models).
• Costs scale linearly with ticket volume. Double the tickets, roughly double the LLM spend.

See Cost Optimization for detailed budgeting guidance and strategies to reduce spend.

Next steps

• Measuring Success — How to define KPIs and track ROI for your AI rollout
• Change Management — Preparing your team for AI-assisted workflows
• Architecture — How AI capabilities work together