Scaling self-service support

Context

Admirals supports thousands of customer requests across multiple regulated markets. A large share of that volume came from recurring questions — users needed help finding information, understanding account actions, solving payment issues, and navigating support topics.

Help Center content existed, but it wasn’t strong enough as a first line of support. Articles were inconsistent, hard to scan, and often written from an internal perspective rather than around what the user was trying to do. Many users still moved from articles into live chat. At the same time, Admirals didn’t yet have a mature AI self-service layer that could resolve repetitive questions before they reached agents.

In this role, I scaled self-service through two connected workstreams: improving Help Center content, and creating the AI chatbot experience from scratch.

Problem

Support agents were spending too much time on repetitive questions that better content and better routing could solve.

Help Center articles weren’t self-sufficient. Many had the right information but users still struggled to resolve their issue — unclear structure, weak hierarchy, inconsistent terminology, duplicated topics, and answers that explained the company process rather than the user task. Conversion from Help Center views to chat was higher than it should have been.

No AI support layer existed. Users who didn’t find an answer in the Help Center had to go directly to a live agent. There was no chatbot that could understand common questions, guide users through next steps, or resolve repetitive cases before they hit a human.

Approach

I rebuilt support self-service as one system, not two parallel tools.

1. Rebuilt Help Center content around user intent

I restructured the Help Center based on real support questions and user tasks:

• rewrote articles in plain language and improved scanability,
• simplified troubleshooting steps,
• removed duplicated or conflicting content,
• standardised terminology across the corpus,
• analysed live-agent transcripts to surface gaps in self-service coverage,
• created new materials for unresolved and frequently escalated questions,
• and retired outdated articles based on view metrics, engagement, and CSAT.

The focus shifted from documenting information to helping users solve problems independently.

2. Built the AI chatbot from scratch

A major part of the initiative was building the chatbot from the ground up with a developer. I owned the conversational experience, prompting, knowledge quality, and content behavior:

• designed conversation flows and support scenarios,
• trained the bot on Help Center articles,
• defined tone of voice and conversational principles,
• created fallback and clarification responses,
• improved the variety and naturalness of answers,
• refined how the bot handled ambiguity, follow-up questions, and user frustration.

I treated the chatbot as part of the content ecosystem, not a side tool. The bot had to speak the same language as the Help Center, use the same terminology, and guide users toward the same next steps. I improved it continuously based on real conversations, unresolved cases, and escalation patterns.

3. Connected Help Center and chatbot into one system

The biggest impact came from treating articles and chatbot flows as one self-service loop. Help Center articles became the source of truth; the chatbot was trained on that content and used it to answer conversationally. The result was a tiered system: simple questions resolved by articles, more specific ones by the bot, and only complex cases reaching live agents.

Artefact 01 · Deflection cascade

Three layers of support — most cases never reach a human.

Help Center articles handle the first wave of questions. The AI chatbot resolves what slips through, trained on the same content. Only complex or ambiguous cases reach a live agent — 71–85% of users never need one.

Layer 01

Help Center

~40%

close on articles

• 8,525 article views (peak month)
• Best chat-conversion: 35% (Nov 2025)
• Strong months: May 52% · Aug 51% · Dec 45%

→

Layer 02

AI chatbot

40–45%

of escalations resolved here

• 52–75% monthly AI-closure rate
• Average 68–70% across 2025
• Trained on Help Center as source of truth

→

Layer 03

Live agent

15–29%

of total users reach a human

• Reserved for complex / ambiguous cases
• 71–85% deflected upstream
• Down sharply from pre-system baseline

Outcomes 71–85% resolved without a human agent·Avg 68% AI-chat closure·Lowest chat escalation of the year at highest article traffic

Results

The self-service improvements showed a measurable impact throughout 2025.

Help Center performance

A key KPI was the conversion rate from Help Center article views to chat — lower meant more users solved their issue without escalating.

The strongest single month was November 2025: 8,525 article views with a 35% conversion to chat — the lowest chat-escalation rate of the year despite the highest article traffic. The content was successfully resolving more questions, even at higher volume.

Other strong months: May 2025 (52% conversion), August 2025 (51%), December 2025 (45%).

AI chatbot performance

The chatbot became a strong second layer. Monthly AI-closed chat rate ranged from 52% to 75%, averaging around 68–70%. Most users who entered chat still didn’t need a human.

Combined impact

Out of every 100 users entering support:

• ~40 resolved their issue through Help Center articles,
• ~60 moved to chat,
• the chatbot closed 52–75% of those chats,
• only 15–29% needed a live agent.

That means roughly 71–85% of support cases were resolved without human involvement — the combined effect of Help Center content and the AI chatbot.

My contribution

• Help Center strategy and restructuring
• Rewriting support articles around user intent
• Improving information architecture and scanability
• Creating the AI chatbot experience from scratch with a developer
• Training the chatbot on Help Center content
• Designing conversation flows and support scenarios
• Defining chatbot tone of voice
• Improving answer accuracy, variety, and relevance
• Optimising fallback and clarification logic
• Analysing unresolved conversations and escalation patterns
• Reducing repetitive workload for support agents through scalable self-service content

---

This case describes work I led at Admirals in 2025 across the Help Center and AI chatbot. Numbers reflect monthly performance reported by the platform; copy and screen examples are paraphrased to respect employer confidentiality.