Software / SaaS

AI Customer Support Assistant for a Growing B2B SaaS Platform

Round-the-clock answers grounded in the help centre, with clean hand-offs to human agents when it matters.

• Conversations resolved by AI

  Share of support conversations resolved without human involvement after three months.

• First response time

  Median first response time, down from over five hours.

• Agent handling time

  Reduction in average handling time on escalated tickets thanks to AI-generated summaries.

• Customer satisfaction

  Share of customers rating AI-handled chats as helpful, in line with human-handled tickets.

Summary

HelpLoop, a B2B SaaS platform for field-service scheduling, launched an AI customer support assistant in its web app and help centre. Within three months it was resolving around 42% of conversations without an agent and cutting first-response times from hours to seconds.

Lessons learned

• Start with the knowledge base, not the model. A two-week clean-up of outdated help articles did more for answer quality than any prompt change.
• Make hand-off a feature. Customers were happy to talk to a bot once they could see a clear route to a person.
• Measure with real tickets. An evaluation set built from 500 historical conversations let us test every change before release.
• Keep account actions narrow. The assistant can look up account status but cannot change billing or permissions.

For teams exploring similar work, our AI chatbot development service and AI web application development guide explain how we approach grounding, evaluation and rollout.

• Engagement

  • Client

  • Industry

  • Duration

  • Team

  • Services

• Technologies

  • Anthropic Claude

    We build AI assistants, agents and document workflows on Anthropic's Claude models, known for long-context reasoning, careful instruction following and strong writing.

  • OpenAI

    We integrate OpenAI models into business software for assistants, document processing, search and content workflows, with grounding, evaluation and cost controls.

  • AI & Machine Learning

    Applied AI and machine learning for forecasting, classification, recommendations and document understanding, alongside large language models where they add value.

  • Microsoft Azure

    We design, build and run applications on Microsoft Azure: App Service, containers, Azure SQL, Functions and Azure OpenAI, with security and cost control built in.

The challenge

HelpLoop's customer base had tripled in two years, but its support team had grown by only a third. Customers in several time zones were writing in around the clock, and the team of twelve agents was struggling to keep up.

An analysis of six months of tickets showed a clear pattern:

• Most questions were repetitive. Around half of tickets were "how do I" questions already answered in the help centre, such as configuring recurring jobs or exporting timesheets.
• Response times were slipping. Median first response had grown to over five hours, and overnight tickets often waited until the next morning.
• Agents lacked context. Agents regularly had to ask customers for their plan, account ID or the page they were on before they could help.
• An earlier chatbot had failed. A rule-based bot introduced two years earlier had frustrated customers with rigid menus and was switched off after a few months.

HelpLoop wanted an assistant that could genuinely answer product questions in natural language, never make up features or policies, and hand over to a human cleanly when it could not help, without adding a night shift.

Our solution

We built a retrieval-grounded AI assistant embedded in HelpLoop's web app and help centre, connected to its existing helpdesk.

Grounded answers

The assistant answers only from HelpLoop's help centre, release notes and a curated set of internal support macros. Content is chunked, embedded and stored in a vector index that updates whenever an article is published. Each answer links to the source article so customers can read more, and if no relevant source is found the assistant says so rather than guessing.

Context-aware conversations

Because the assistant runs inside the logged-in app, it knows the customer's plan, enabled features and the page they are on. A small set of read-only tools lets it check account status, recent job sync errors and subscription details.

Clean hand-off to agents

When a customer asks for a person, raises a billing or data issue, or the assistant's confidence is low, it creates a ticket in the helpdesk with a summary, the conversation transcript and the account context attached. Agents pick it up without asking the customer to repeat themselves.

Models and safeguards

We used Anthropic Claude for conversation and summarisation, behind a provider abstraction so models can be swapped. Safeguards include strict grounding instructions, PII redaction in logs, rate limits and an evaluation suite run on every prompt or content change. This work combined our AI agent development and chatbot expertise.

What we built

Key features

• Retrieval-grounded answers

  Answers come only from the help centre and approved support content, with source links for every reply.

• In-app account context

  The assistant knows the customer's plan, features and current page, so answers fit their setup.

• Read-only account tools

  Looks up account status, sync errors and subscription details without being able to change them.

• Helpdesk hand-off with summaries

  Escalations create tickets with a summary, transcript and account context attached.

• Continuous evaluation

  An evaluation suite of 500 real conversations is run before every change goes live.

Delivery

How we delivered it

• Ticket analysis and scoping

  Analysed six months of tickets to identify high-volume question types and agree what the assistant should and should not handle.

• Knowledge base clean-up

  Worked with the support team to update, merge or retire outdated help articles before indexing.

• Build and evaluation

  Built retrieval, tools and hand-off flows, testing against a set of historical conversations.

• Staged rollout

  Launched to 10% of customers, then 50%, then everyone, reviewing transcripts at each stage.

• Ongoing tuning

  Monthly reviews of unresolved conversations feed new help articles and prompt improvements.

Gallery

Inside the solution

• In-app support assistant

• Support deflection dashboard

In the client's words

• The assistant resolves the repetitive how-do-I questions and hands everything else to us with a proper summary. Our agents spend their time on the tricky tickets now, and customers get answers at 2am without us adding a night shift.

Project

Related projects

• AI Sales Assistant for B2B Teams

  An LLM-powered sales assistant that qualifies inbound leads, drafts tailored follow-ups and summarises accounts from CRM data for a B2B industrial supplier.

• Smart Education ERP

  A unified education management platform for schools, colleges, universities, and multi-campus institutions, covering admissions, attendance, timetabling, fee management, examinations, academic records, and parent and student portals in one integrated system.

• Clinic Appointment Mobile App

  A patient booking app and clinic scheduling back office for a group of private outpatient clinics, with online booking, reminders and secure pre-visit forms.

Solutions

Solutions used

• AI Customer Support

  AI customer support assistant that answers questions from your own knowledge base, takes actions through your systems and hands off to human agents with full context.

• Custom SaaS Platform

  A production-ready SaaS foundation — multi-tenancy, authentication, billing, admin and APIs — that we customise into your own software product, so you launch sooner.

• AI Business Automation

  AI business automation platform that reads documents, classifies requests and runs multi-step workflows across your systems, with human approval where it matters.

Services

Services involved

• AI Chatbot Development

  AI chatbot development for customer support and internal help desks: assistants trained on your knowledge base, integrated with your helpdesk, website and Teams, with smooth human handover.

• AI Agent Development

  AI agent development for multi-step business tasks: agents that use your APIs and tools to research, update records and resolve requests, with approvals, audit logs and clear limits.

• SaaS Development

  SaaS development for start-ups and established companies: multi-tenant architecture, subscription billing, onboarding and admin tooling, built to scale securely on Azure or AWS.

More results

Other case studies

• Retail Chain POS Modernisation

  How a 60-store retail chain replaced ageing on-premise tills with an offline-capable cloud POS, cutting checkout times by 35% and ending overnight price syncs.

Support volume outgrowing your team?

Talk to our AI team about a grounded assistant that works alongside your agents.

Talk to our AI team about a grounded assistant that works alongside your agents.