Support — Charter

Support is where the company meets reality. It is also the most heavily automated customer-facing function in most organisations — and the one where automation most often makes things worse.

The distinction that decides which way it goes: automation that gets a customer an answer faster is good; automation that makes them work harder to reach a person is not. Most bad support automation is the second kind wearing the language of the first.

What this role owns#

Service levels. Response and resolution times, by severity, actually met.

Escalation. A working path to someone who can fix it, not a queue that loops.

Product feedback. Support sees every defect and every point of confusion first. If that information does not reach Engineering and Product, the same tickets recur forever.

Known issues. Documented at release, so support is not discovering them alongside the customer.

The knowledge base, which is also what any support automation reads from.

KPIs#

Deflection without satisfaction is the metric that quietly destroys support functions. A chatbot that prevents contact and a chatbot that resolves problems look identical on that number and are opposites in reality.

Where automated support genuinely works#

• Answering documented questions instantly, at any hour, from a knowledge base that is current
• Gathering context before a human sees it — account, history, what was already tried
• Triage and routing to the right team with a summary
• Status and progress on an existing ticket
• Draft responses for an agent to review and send

Where it destroys trust#

• No route to a person, or one deliberately buried. This single design choice does more brand damage than any outage.
• Confidently wrong answers about your specific product, from a model that has not been given your documentation. See RAG — the fix is grounding, not a better model.
• Not disclosing that it is automated. Beyond trust: EU AI Act transparency obligations applied from 2 August 2026 — customers must be told when they are interacting with an AI system.
• Automating the apology. A generated response to a serious complaint reads as exactly what it is.
• Answering from stale documentation. The automation is only as current as what it reads.

AI agents in this function#

First-line assistant — answers documented questions from the knowledge base, grounded, with sources, and hands to a person the moment it cannot answer or the customer asks.

Triage agent — classifies, prioritises, routes with a summary.

Context assembler — gathers account and history before a human opens the ticket.

Response drafting agent — proposes a reply for an agent to review, edit and send.

Trend agent — clusters tickets by root cause and surfaces what is actually driving volume.

What stays human: complaints, anything involving money owed or refunded, security or privacy concerns, and any customer who asks for a person. That last rule should be absolute and immediate.

SOPs#

• Severity definitions and the response committed to each.
• Escalation — when, to whom, and what the customer is told while it is in progress.
• Handover to a person — immediate on request, with full context, no repeating themselves. Making a customer restate their problem is the most common failure of automated support.
• Known issues at release — Engineering supplies these before shipping, not after tickets arrive.
• Knowledge base maintenance — an article is updated when a ticket proves it wrong. This is what keeps the automation accurate.
• Root cause reporting — monthly to Product and Engineering, ranked by volume and by cost.

Templates#

SOP for recurring procedures · Incident Report for customer-affecting incidents · known issues register · escalation record.

Workflows#

In: customer contacts · incident notifications · release notes and known issues · product changes.

Out: resolutions · escalations · root cause analysis to Product and Engineering · customer sentiment · knowledge base updates.

Handoffs: Engineering for defects · Security for privacy and security reports · Sales for renewals at risk · COO for anything systemic.

FAQ#

Should we use an AI chatbot for support?#

Yes, if it is grounded in your actual documentation, discloses that it is automated, and hands to a person immediately on request. No, if it exists to reduce contact volume — customers can tell the difference within one exchange, and the second kind costs more in churn than it saves in headcount.

How do we measure whether support automation is working?#

Deflection and satisfaction together, plus repeat contact rate. If deflection rises while satisfaction falls, you are not resolving issues — you are preventing contact, which is a different and much more expensive thing.

What should never be automated in support?#

Complaints, refunds, anything with a security or privacy dimension, and any customer who asks to speak to a person. Automating those saves very little and costs the relationship.

How do we stop the same tickets recurring?#

Report ticket volume by root cause to whoever can fix the cause, and make sure it is ranked by cost rather than by count. Support that only resolves tickets is a cost centre; support that eliminates their causes is not.

What else is coming for Support