The Reality of AI in Customer Service: Balancing Risks and Opportunities
Updated September 29, 2026. Ian Landsman rewrote this guide to reflect how support teams use AI today and what HelpSpot offers now. The original version was published April 10, 2024, by Dave Hicking.
AI can handle a real share of customer service work. It can answer well-documented questions, collect missing details, sort incoming requests, and draft replies. Its failures are just as real, and they are specific: confident answers that are wrong, actions taken on incomplete information, and customers who can't reach a person when they need one.
So the useful question isn't "AI or people." It's which work goes to each, and what controls make each assignment safe. This guide covers the main disadvantages of AI in customer service, why implementation is harder than a demo suggests, and a practical way to divide the work.
The real disadvantages of AI in customer service
Confident, wrong answers
Language models write fluently whether or not they're right. A wrong answer in a helpful tone can be worse than no answer, because the customer has no reason to doubt it.
This isn't only a reputation problem. In Moffatt v. Air Canada (February 2024), a British Columbia tribunal held the airline responsible for incorrect bereavement-fare information from a chatbot on its website. The chatbot's answer contradicted the airline's own policy page, and the tribunal rejected the argument that the chatbot was responsible for its own statements. Whatever your AI tells customers, your business said it.
Stale or conflicting sources
AI answers are only as good as what it can read. Old help articles, outdated saved replies, and past tickets that reflect a retired policy can all become answers. Many "AI mistakes" are documentation problems that AI repeats faster and more widely.
Acting on incomplete context
Answering a question is one risk. Taking an action is another. An AI that can issue a refund, change an account, or close a ticket can do real damage when it misreads the request or lacks the history a person would check first. The risk grows with the permissions you give it.
Tone that misses the moment
A correct answer can still be the wrong reply. A customer who is angry, grieving, or facing a deadline needs acknowledgment and judgment, not a polished template. AI can write well and still miss what a short, frustrated message is really saying.
Dead ends
A loop with no exit is the fastest way to turn a small problem into a lost customer. If customers can't reach a person, or the handoff drops the conversation history, AI has made the problem worse.
Privacy exposure
AI features send text to a model. Know what data goes where, which provider processes it, what it retains, and whether sensitive values such as card or identity numbers are redacted first.
Metrics that hide failure
"Resolved" can mean the customer got an answer, or it can mean they gave up. Deflection and containment numbers look good either way. If you only count how many conversations AI closed, you won't see the ones it failed.
Why implementation is harder than the demo
Demos use clean questions and a tidy knowledge base. Your queue doesn't. Most of the work that decides whether AI helps or hurts happens before it answers anyone:
- Content quality. Audit the articles and saved replies AI will draw on. Retire what's outdated. Fill the gaps behind your most common questions.
- Scope. Decide which request types AI may handle and which it should leave alone. Start narrow.
- Permissions. Give AI the least access that does the job. Reading data and changing it are separate decisions.
- Handoff design. Decide when AI stops, who picks up the request, and what context comes with it.
- Ownership. Someone on the team needs to own AI configuration, review results, and fix sources when answers go wrong.
- Measurement. Track reopens, escalations, customer satisfaction, and a regular sample of AI-handled conversations, not just volume.
Match the work to the risk
The most practical way to divide work is by what a mistake would cost. Here's a starting framework, from lowest to highest stakes. Adjust it for your business, your customers, and any regulatory obligations.
Well-documented questions
Hours, setup steps, where to find a setting.
- AI: Answer when approved sources clearly support it.
- People: Keep sources current and review samples.
Gathering information
Order numbers, error messages, account details.
- AI: Ask for what's missing before a person picks up.
- People: Define what each request type requires.
Sorting and routing
Categories, tags, priority signals.
- AI: Classify requests based on your rules.
- People: Set categories and rules, and correct misroutes.
Drafting
Replies to routine and mid-complexity requests.
- AI: Draft from ticket context and sources.
- People: Decide which drafts need review before sending.
Actions with consequences
Refunds, account changes, cancellations.
- AI: Act only with narrow, explicit permissions and a record of what ran.
- People: Set limits and handle exceptions.
Complex or sensitive cases
Multi-step problems, billing disputes, outages, upset customers.
- AI: Summarize history, surface related requests, and translate.
- People: Own the case from start to finish.
Two things stand out. First, AI has a useful role in every category, even where a person owns the outcome. Second, the lines move. As you see how AI performs on a type of request, you can give it more room there and less where it struggles.
Should a person review every AI reply?
Not necessarily. Reviewing every reply is a sensible way to start, and it's the right permanent choice for some high-stakes request types. But blanket review everywhere gives up most of the time savings, and it tends to decay into rubber-stamping.
A better approach is to set review by request type:
- Start with review on, and track how often reviewers change what AI produced.
- Where output is consistently accurate and mistakes are cheap to fix, allow more autonomy and keep sampling.
- Where output is often corrected, or a mistake would be costly, keep review on or keep AI out.
- Revisit these choices when your products, policies, or documentation change.
Whatever you allow, customers should always have a clear way to reach a person, and that person should see the full conversation.
The cases left for people get harder
When AI takes on routine work, the queue your team sees changes. What remains is the multi-step problem, the customer who has written in four times, the issue that needs another department, the exception to policy. These cases take longer, and they need better tools, not fewer.
That means:
- Clear ownership. One person responsible for each case, so nothing falls between a handoff and a busy team. Single assignment exists for exactly this.
- Full history. Every message, internal note, and earlier request from the same customer in one place.
- Private collaboration. Internal notes and staff mentions, so specialists can weigh in without the customer seeing the back-and-forth.
- Follow-through. Reminders, SLAs, and statuses, so long-running cases don't stall.
A tool built only to automate answers leaves your team without these. A help desk with no AI leaves routine work on their plate. You need both.
How HelpSpot approaches AI
HelpSpot is a full help desk where AI handles the work it can and people efficiently manage complex cases. As of September 2026, HelpSpot includes:
- AI Composer, which turns an agent's notes into a full draft using ticket history and knowledge base content. Agents review and edit Composer drafts before sending.
- AI categorization, summaries, and translation to sort incoming requests, condense long emails, and work across languages.
- Knowledge base article generation, which turns a strong reply into an unpublished draft for your team to refine and publish.
- The HelpSpot MCP server, which lets assistants like Claude and ChatGPT work with your help desk. An administrator has to enable it. Every operation runs with the connected user's HelpSpot permissions, and admins can disable individual tools or all write tools. Whether the assistant asks before each change depends on the AI client and its settings.
- An admin switch that turns AI features off entirely.
These sit alongside the help desk your team uses for complex work: single-assignment ownership, private notes, collision detection, automation, reporting, and a choice of US Cloud, EU Cloud, or self-hosting at the same price. See how HelpSpot supports people handling complex cases.
HelpSpot first introduced AI features in 2023. That post describes the original release, not the current feature set.
If you'd like to see how this would work for your team, contact us for a demo.
Confident but wrong answers, answers drawn from outdated sources, actions taken on incomplete context, tone that misses an upset customer, and dead ends where customers can't reach a person. Privacy exposure and metrics that count abandoned conversations as resolved are also common problems.
Not necessarily. Review everything when you start and for high-stakes request types. Where AI output is consistently accurate and mistakes are cheap to fix, you can allow more autonomy while continuing to sample results. Customers should always have a clear path to a person.
AI can resolve some well-documented, low-risk requests without a person when it works from current, approved sources and has a clear handoff. Complex, ambiguous, or sensitive cases still need a person who owns the outcome, supported by a help desk that gives them full context.
HelpSpot is a full help desk where AI handles the work it can and people efficiently manage complex cases. Current AI features include AI Composer drafts that agents review before sending, categorization, summaries, translation, knowledge base draft generation, and an MCP server that works within each user's HelpSpot permissions.
