# Agent context: Support Queues Growing Faster Than the Team

Source: https://aiplaybooks.app/industries/software/customer-support/support-backlog
AI suitability: conditional

## Objective

Resolve software customers' tickets faster and reduce how many arrive, by routing tickets well, maintaining the knowledge base, fixing the top contact reasons, and using AI to draft replies for agents and to answer narrow, well-documented questions, without letting AI make statements or decisions that bind the company or cost the customer.

## Constraints

- Ground every AI answer and draft in current knowledge base articles or reviewed past resolutions, and cite the source. If no source covers the question, say so and hand over to a person.
- Never let an AI agent grant refunds or credits, change account access or ownership, delete data, or state contract terms, prices, or policies that are not in its sources.
- Tell customers when they are talking to an AI system, as Article 50 of the EU AI Act requires of AI systems that interact with people in the EU from 2 August 2026, and give them a one-step route to a person.
- Measure resolution, satisfaction, and reopen rate, not deflection, and compare against agents for the same request types.
- Redact credentials and secrets from tickets before they reach any model, and send customer data to an outside model only under a data processing agreement.
- Pilot agent assist with a comparison group and report quality by agent experience level, not only average speed.

## Inputs

- Tickets and conversations with product area, plan, account, version, priority, and resolution.
- Help centre and internal articles with product area, product version, and review date.
- Account context such as plan, entitlements, configuration, recent errors, and open incidents.
- Satisfaction ratings, reopen events, escalations, and quality review scores.
- Known defects, incidents, and release notes.

## Outputs

- Routing decisions with the rule that fired, and service level timers per plan.
- A monthly contact-reason report with ticket counts per reason, customers and plans affected, and a product owner for each top reason.
- Reply drafts for agents, each linked to the articles or tickets it drew on.
- Where justified, direct answers for approved request types, with a hand-over record when the AI could not answer.

## Success criteria

- Tickets per active customer fall for the contact reasons that were fixed.
- Time to resolution falls against each plan's service levels without a fall in satisfaction or a rise in reopens.
- Agent assist raises issues resolved per hour against the comparison group without lowering quality for any experience level.
- AI-answered request types keep satisfaction and wrong-statement rates within agreed limits, measured weekly.

## Known pitfalls

- Launching a customer-facing AI agent on top of an outdated knowledge base, so it confidently repeats what the help centre gets wrong or fills the gaps itself.
- Measuring deflection, meaning conversations that never reached a person, instead of resolution, so customers who gave up count as successes.
- Letting an AI agent state policies, prices, or terms that are not in its source documents, as a software company's support bot did when it invented a login policy.
- Making it hard to reach a person, which turns a failed bot conversation into a lost customer.
- Cutting support headcount on the strength of a launch announcement, before quality and repeat contacts are measured over months.
- Treating support volume as fixed demand, so AI answers the same avoidable question faster instead of anyone removing its cause.

## When not to use AI here

- The knowledge base is incomplete or out of date for the current product, so there is nothing reliable to ground answers in, which rules out a retrieval-based assistant or bot but not an assistant trained on reviewed past conversations.
- Most tickets are new or caused by a current incident or defect, where the fix is engineering work and a clear status update.
- The request involves money, account access, data deletion, or a contractual commitment, where a wrong answer is costly and the decision belongs to a person.
- Ticket volume is low enough that routing, macros, and a maintained help centre already meet service levels.
- The customer is distressed or the issue is a complaint, where field evidence found the chats an AI handled rated lower and human intervention after an AI failure less effective for emotional escalations than for technical ones.

## Candidate approaches

- Ticket routing, macros, and a maintained knowledge base (primary): Structured ticket forms, routing rules by product area, plan, and urgency, saved replies for known issues, and help centre articles that agents capture and update as part of solving tickets, so each ticket reaches the right person with a current answer.
- Contact-reason analysis for product fixes (supporting): Classifying every ticket by the reason the customer had to contact support, such as a defect, a confusing setting, a missing feature, or a billing question, and sending the largest and fastest-growing reasons to the product and engineering teams that can remove them.
- Agent assist reply drafting (supporting): An assistant that reads the customer's conversation and account context, retrieves relevant help centre articles and past resolved tickets, and drafts a reply or next step that the support agent edits and sends.
- Customer-facing AI agent for narrow request types (supporting): An AI agent in the help centre and chat that answers customers directly, limited to a defined list of request types with current documented answers, that tells customers it is AI and hands them to a person in one step.

## Human oversight

Support agents review and send every AI-drafted reply in the assisted workflow. A support lead owns the list of request types an AI agent may answer alone, reviews a sample of its conversations every week, and removes a request type when its errors rise. Refunds, credits, account access and ownership changes, contract and pricing terms, and exceptions to policy are always decided by a person. Product managers own fixes for the top contact reasons.

## Sources

- [KCS v6 Practices Guide](https://library.serviceinnovation.org/KCS/KCS_v6/KCS_v6_Practices_Guide), Consortium for Service Innovation, practitioner
- [Generative AI at Work](https://doi.org/10.1093/qje/qjae044), The Quarterly Journal of Economics, 2025, research
- [Agentic AI and Human-in-the-Loop Interventions: Field Experimental Evidence from Alibaba's Customer Service Operations](https://arxiv.org/abs/2605.14830), arXiv preprint, 2026, research
- [Frontiers: Machines vs. Humans: The Impact of Artificial Intelligence Chatbot Disclosure on Customer Purchases](https://doi.org/10.1287/mksc.2019.1192), Marketing Science, 2019, research
- [Klarna AI assistant handles two-thirds of customer service chats in its first month](https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/), Klarna, 2024, case-study
- [Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396), Entrepreneur, 2025, case-study
- [Incident 1039: Anysphere AI Support Bot for Cursor Reportedly Invents Login Policy, Leading to Subscription Cancellations](https://incidentdatabase.ai/cite/1039/), AI Incident Database, 2025, case-study
- [Moffatt v. Air Canada, 2024 BCCRT 149](https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do), Civil Resolution Tribunal of British Columbia, 2024, case-study
- [Regulation (EU) 2024/1689, the Artificial Intelligence Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), Official Journal of the European Union, regulation