Support Queues Growing Faster Than the Team

Resolving software customers' questions and problems quickly as ticket volume grows, without hiring at the same pace and without handing customers to a chatbot that gives wrong answers or no way to reach a person.

AI is suitable with conditions

Support has unusually strong field evidence for AI. In a study of 5,172 support agents at a Fortune 500 company selling business process software to small businesses, a generative AI assistant, fine-tuned on past support conversations, that suggested replies to agents raised issues resolved per hour by 15% on average, with the largest gains for less experienced agents and small declines in quality for the most experienced. That evidence is for a tool that drafts and an agent who decides. Customer-facing bots are a different matter: Klarna reported in 2024 that its assistant handled two-thirds of customer service chats, then said in 2025 that its AI service had produced lower quality and began hiring human agents again, and in a 2024 field experiment on Alibaba's Taobao platform, published in 2026, chats the AI handled were rated lower even though repeat contacts did not rise. A software company's own AI support bot invented a login policy that led customers to cancel, and a Canadian tribunal held an airline liable for what its chatbot told a customer. AI is suitable as an assistant to agents once tickets are routed well and, for an assistant that retrieves from it, the knowledge base is maintained, and as a customer-facing agent only for narrow, well-documented requests with an easy route to a person.

Problem Statement

A software company’s support team answers customers’ questions and fixes their problems, and the volume of tickets grows with every new customer and every release.

The team can hire, it can help customers help themselves, it can remove the causes of tickets, or it can let AI answer some of them.

Much of the attention goes to the last option.

The evidence behind this playbook points to a narrower role: AI that helps agents works, AI that replaces agents works only for narrow, well-documented requests, and both depend on knowledge that is written down and current.

Business Context

Some of the strongest field evidence on generative AI at work comes from this function.

A study of the staggered rollout of a generative AI assistant to 5,172 support agents, at a Fortune 500 company selling business process software to small and medium-sized businesses in the United States, found that access to the assistant raised the number of issues resolved per hour by 15% on average.

The gains were largest for less experienced and lower-skilled agents, who improved in both speed and quality, while the most experienced became slightly faster and their quality slipped a little.

Customers were more polite to assisted agents and less likely to ask for a manager.

Customer-facing AI has a more mixed record.

Klarna, a payments company, announced in February 2024 that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, with satisfaction it described as on par with human agents and a 25% drop in repeat inquiries.

In May 2025 its chief executive told Bloomberg, as reported by Entrepreneur, that cheaper AI service had led to “lower quality”, that customers should be clear they can always reach a person, and Klarna began hiring human agents again.

Why This Is Hard

The first difficulty is that most of the work is answering questions whose answers are written down somewhere, but often not anywhere current.

An AI that retrieves from an outdated help centre gives outdated answers fluently, and one that fills the gaps itself invents them.

In April 2025, users of Cursor, a coding tool made by Anysphere, were being logged out when they switched devices, and its AI support agent told them this was a new login policy, when no such policy existed, and customers cancelled subscriptions, according to the AI Incident Database.

The second difficulty is that responsibility does not move to the bot.

When Air Canada argued that its chatbot was responsible for its own statements, the British Columbia Civil Resolution Tribunal called this a remarkable submission and held the airline liable, finding it “responsible for all the information on its website” whether that information came from an ordinary page or a chatbot.

The third difficulty is measurement.

A conversation that never reaches a person can be a resolved question or a customer who gave up, and counting both as deflection hides the difference.

In a randomised field experiment on Alibaba’s Taobao platform in 2024, published in 2026, AI shortened chats and repeat contacts did not rise, but customers rated the chats the AI handled lower, which the authors attribute to communication style rather than to whether the problem was solved, and human intervention after an AI failure worked better for technical problems than for emotional ones.

The fourth difficulty is that being honest about the bot has a cost.

A field experiment on sales calls found that disclosing that the caller was a chatbot reduced purchases, because customers judged a bot they knew about to be less knowledgeable and less caring, so a company is tempted to blur the line that the law increasingly requires it to draw.

How The Solutions Fit Together

Routing, macros, and a maintained knowledge base are the floor.

They get each ticket to the right queue with the context it needs, and give agents and customers a current answer to the questions that recur.

Contact-reason analysis runs beside them, sorting tickets by why customers had to ask and sending the largest causes to product and engineering, which is the only solution that reduces volume rather than handling it faster.

An assistant that drafts replies for agents is where AI has the best evidence, and in this playbook’s view a retrieval-based assistant depends on the same knowledge base, although the assistant in the main study was fine-tuned on past conversations instead.

A customer-facing AI agent supports all three for narrow request types with a documented, current answer and a one-step route to a person.

None of these decides anything that costs the customer or the company money.

Implementation Considerations

Fix the knowledge before adding AI, because every AI answer is only as good as the article it was drawn from.

Knowledge-Centered Service, a method maintained by the Consortium for Service Innovation, makes capturing and improving articles part of how each ticket is solved rather than a separate documentation project.

Record each article’s product area and product version, and retire or fix articles when the product changes, with the whole team sharing ownership of the knowledge as KCS recommends rather than assigning each article to one person.

Pilot an agent assistant with a comparison group of agents, as in the study above, and measure resolution time, satisfaction, and quality review results for each experience level, since the most experienced agents may see quality fall.

Define the request types a customer-facing agent may answer, each with its source articles, and keep money, account access, contracts, and policy exceptions out of that list.

ROI Considerations

The return from an agent assistant is resolved tickets per agent hour, net of any fall in quality, and it is largest where agent turnover is high and new agents take long to become productive.

The return from a customer-facing agent is the cost of the conversations it resolves, net of repeat contacts, escalations, lower satisfaction, and the cost of any wrong answer.

Klarna estimated its assistant would improve profit by $40 million in 2024, and a year later its chief executive described renewed investment in human support, which shows why that return should be measured over months rather than at launch.

The return from contact-reason analysis is tickets that never arrive, which is the only one of the three that bends the curve of support cost against customer growth.

Security, Privacy, and Regulation

Tickets contain personal data and often customers’ own data, credentials, or screenshots, so any AI service that reads them needs a data processing agreement, a defined retention period, and redaction of secrets.

Under Article 50 of the EU AI Act, providers must design AI systems that interact directly with people so that those people are told they are dealing with an AI system, unless that is obvious, and that obligation applies from 2 August 2026.

The Air Canada decision means a company should expect to be held liable for what its chatbot tells customers, so the request types a bot may answer and the documents it may draw on are a legal decision as well as a support one.

Give the AI agent no ability to grant refunds, change account ownership, or alter contract terms, so a manipulated conversation cannot make it do so.

Where The Evidence Disagrees

Klarna’s 2024 announcement reported customer satisfaction on par with human agents, while its chief executive said in 2025 that the AI service had produced lower quality, and the Taobao field experiment found lower ratings for the chats the AI handled but no change in overall ratings or repeat contacts.

Neither is a software company, and the sources here give no later satisfaction figures for Klarna, so this playbook does not treat either as settled for software support.

The productivity study found that AI helped least, and slightly lowered quality, for the most experienced agents, so an average gain can hide a loss for the agents a team most relies on.

The study also found the largest gains on moderately rare problems, which suggests the value of an assistant lies less in the common questions a macro already answers than in the long tail, although this is one study at one company.

Symptoms

First response and resolution times slip past the service levels promised in customers' plans, and the backlog grows after every release.
The same how-to questions, billing questions, and known defects arrive every week, and agents retype the same answers.
New agents take months to handle technical tickets alone, and senior agents spend their time answering colleagues' questions.
Leadership asks how much of support an AI agent can take over, while customers complain about bots that loop or will not hand them to a person.
Help centre articles are out of date for the current version of the product, so customers who search still open tickets.

Root causes

Product changes, confusing design, and defects generate questions faster than documentation is updated.
Knowledge of how to fix problems lives in agents' heads and old tickets rather than in maintained articles.
Tickets arrive without the product area, plan, account, or error details needed to route them, so they bounce between queues.
The causes of the most common tickets sit with product and engineering teams, who rarely see support data in a form they can act on.
Support is judged on cost per ticket, which rewards answering faster rather than removing the reasons customers have to ask.

Business impact

Customers who cannot get unstuck use less of the product, which in this playbook's judgement puts their renewal at risk.
Support headcount that grows in step with customers, eroding the margins a subscription business depends on.
Agent burnout and turnover, which lose the knowledge that lives in experienced agents' heads.
Legal and reputational exposure when an automated assistant tells customers something untrue about the product, its pricing, or its policies.

Traditional approaches

Ticket forms, tags, and routing rules that send each ticket to the queue for its product area, plan, and urgency.
Saved replies and macros for known issues, maintained by team leads.
A help centre and community forum for self-service, and in-product guidance.
Knowledge-Centered Service, in which agents capture and improve articles as part of solving each ticket.
Regular reports of top contact reasons to product and engineering, so defects and confusing features are fixed at the source.

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Start with routing, macros, and a knowledge base that agents maintain as they solve tickets, and send the top contact reasons to product and engineering every month so the causes get fixed. Add an assistant that drafts replies for agents from the knowledge base and the ticket history, and measure it on resolution time and quality per agent. Let an AI agent answer customers directly only for narrow request types with a documented, current answer, tell customers they are talking to AI, and give them a one-step route to a person. Agents decide what is sent wherever a draft might be wrong, and refunds, credits, account access, contract terms, and policy exceptions are always decided by a person.

Data requirements

Tickets and conversations high sensitivity: Every ticket and chat with its messages, product area, plan, account, priority, assigned queue and agent, timestamps, and resolution, including the macro or article used. Contains customer names, account details, and sometimes credentials or customers' own data pasted into a message.
Knowledge base low sensitivity: Help centre and internal articles with product area, product version, last review date, and usage, which every AI answer is grounded in and which is often the weakest link.
Account and product context medium sensitivity: The customer's plan, entitlements, configuration, recent errors, and open incidents, so a reply or a route reflects the customer's actual situation.
Satisfaction and quality reviews medium sensitivity: Customer satisfaction ratings, reopened tickets, escalations, and quality reviews of a sample of conversations by team leads, used to judge any change in resolution speed against its effect on quality.
Product defects and releases low sensitivity: Known defects, incidents, and release notes, linked to tickets, so contact reasons can be traced to the product change that caused them.

Evaluation metrics

Tickets per active customer, by contact reason, against the previous quarter, as the measure of whether causes are being fixed.
Time to first response and time to resolution against each plan's service levels.
Customer satisfaction and reopen rate, measured separately for conversations handled by AI, assisted by AI, and handled by agents alone.
Share of AI-handled conversations that were resolved, not only those that never reached a person.
Errors found in quality reviews of AI-drafted and AI-sent replies, especially wrong statements about pricing, policy, or product behaviour.

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.

Common 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

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.

Sources

Practitioners

KCS v6 Practices Guide, Consortium for Service Innovation

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

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.

Last updated 2026-10-03