Shadow AI in Customer Support: Turn It Into a Team Workflow
Neal Travis · · 8 minutes read

Ask a support leader how many separate AI setups are connected to their helpdesk right now, and most can’t give you a number. One agent pastes a tricky ticket into ChatGPT to get a first draft. Another has connected Claude to Zendesk on a personal account to summarize long threads. Each of them set it up alone, with their own prompts.
That’s shadow AI, and in support it tends to grow one agent at a time. The people closest to the tickets are the first to spot where AI saves them time, so they go ahead without waiting for a team setup.
Shadow AI is any AI tool people use for work outside a setup the company has approved and can see.
Banning it rarely works, since people will use the tools they prefer regardless of what is available. Leaving it alone means every agent runs a slightly different, invisible version of your support process. In this article, we’ll cover how far shadow AI has already spread in support teams, what goes wrong when everyone sets up AI on their own, why bans backfire, and a step-by-step AI adoption strategy to turn individual use into a shared, visible team workflow.
How far shadow AI has already spread in support teams
Swifteq’s own data on teams connecting AI to Zendesk shows how far this has gone. Nearly half of all accounts exist only because colleagues at the same company each signed up separately. One company had more than 80 separate accounts on a single Zendesk instance. Fewer than 1 in 18 companies ever invited a colleague into a shared setup.
Every one of those accounts is an agent who decided AI would help with their tickets and set it up themselves. That’s a good instinct, but the problem is that their team lead usually has no idea it’s happening.
The wider research points the same way. Microsoft and LinkedIn’s 2024 Work Trend Index found that 75% of knowledge workers use AI at work, and 78% of those users bring their own AI tools. In support specifically, Zendesk’s 2025 CX Trends report found that agents’ use of shadow AI, which it defines as “external tools unapproved for agent use,” jumped by up to 250% year on year in some industries.
There’s a quieter version of the same problem, too. One head of support told Swifteq that their offshore team pastes tickets into ChatGPT with very basic prompts. In that case the leader knows AI is in use, but nobody has shaped how it’s used. The quality of every AI-assisted reply comes down to whoever wrote the prompt that day.
What goes wrong when everyone sets up AI on their own
In my experience at the Academy to Innovate HR (AIHR), we went through this during a company-wide push on AI transformation. We encouraged everyone to use AI for as much of their work as they could, and we rolled out Gemini as the company tool. The encouragement worked, and AI use took off across teams. But not everyone had every tool they wanted, so people started opening their own accounts with the tools they preferred, and a few problems showed up quickly:
- Company and customer data ends up in personal accounts: Once someone connects company data to a personal AI account, you lose control over where that data goes and who can see it. In support, that means customer conversations and account details going into AI tools nobody has reviewed. IBM’s 2025 Cost of a Data Breach report found that one in five organizations reported a breach due to shadow AI, and organizations with high levels of shadow AI paid an average of $670,000 more per breach.
- Skills and answers drift apart: One agent writes careful prompts and checks the output. Another pastes the ticket in and sends whatever comes back. Customers get noticeably different answers depending on who picks up their ticket.
- People pull different data and end up measuring different things: When each person asks AI for numbers from their own sources, reports stop lining up. Two people can ask the same question about ticket volume and come back with two different figures.
- Work gets built in tools other people can’t open: Some colleagues built artifacts in Claude that only people with a Claude account could use. The work was useful, but anyone on a different tool was locked out of it.
The lesson I took from it was that encouraging AI use works best when everyone gets the same tools and the same level of training on them. When the company makes the right tools easy to get, people stop going off to find their own, and most of the risk goes with it.
Why banning shadow AI in customer service backfires
The first instinct when you discover shadow AI is to shut it down. But the agents using it have already seen it save them time. A ban mostly pushes that use somewhere you can see even less of it, and it sidelines the people on your team who’ve done the most experimenting.
Forcing everyone onto one approved chat tool has its own problem, and it’s the one we ran into at AIHR. People settle on the AI tool they like. If the approved one doesn’t fit how they work, they’ll open an account somewhere else.
When I asked Sorin, Swifteq’s founder, where he’d draw the line, his answer was to standardize the connection to your helpdesk and let people keep the AI they chat in. That connection decides which customer data the AI can read and what it’s allowed to change in Zendesk. Get that right, and it matters much less whether an agent types into ChatGPT or Claude.
An AI adoption strategy for support teams in 6 steps
If I suspected my team was already using AI on tickets, this is the process I’d work through:
1. Find out what’s already in use
Ask your team directly, and make it clear nobody is in trouble for using AI. Ask which tools they use and what they use them for. You’ll learn which use cases are worth standardizing, and you’ll find out who your strongest prompt writers are.
2. Write a one-page AI usage policy
Keep it short enough that agents will read it. It should answer four questions: which tools are approved, what customer data can go into them, which AI outputs need a human review before they reach a customer, and who to ask when something falls outside the policy. For the data side, I put together an AI data privacy checklist for support teams that covers what to check before you connect AI to customer data.
3. Give everyone access through one team setup
This is the step that makes AI use visible. With the MCP Server for Zendesk, you create one team account and invite your agents into it, so you can see who is connecting AI to your helpdesk. Each person connects with their own Zendesk login, which means their AI assistant can only see and change what their Zendesk role already allows. It connects to the AI assistants people already use, including Claude and ChatGPT, so nobody has to give up a tool they like. If several of your agents already signed up separately, Swifteq can merge those accounts into one.
4. Decide what AI is allowed to change
Start conservative. The account owner sets what AI can do in Zendesk for the whole team and can restrict individual people further. Several actions, like public replies and publishing help center articles, stay off until an owner turns them on. Each agent can tighten those limits for their own sessions, and only the account owner can loosen them. A sensible starting point is read-only access or internal notes, moving to public replies once you trust the output and your review process. The MCP Server setup guides walk through each permission.
5. Share prompts and drafting standards
Once access is shared, prompt quality is the next gap to close. A team that pastes tickets into ChatGPT with basic prompts gets basic replies, and a slightly different one from each agent. Collect the prompts that work into a shared library, or into skills your AI assistant can reuse, so every agent starts from prompts that have already been tested. For reply drafting inside Zendesk, Agent Co-writer gives the whole team one AI drafting setup. Drafts are grounded in your help center articles and macros, and team leads set the writing guidelines and shared prompts every agent works from.
6. Train everyone to the same level, then review monthly
Run a short session so every agent knows the approved setup and where the shared prompts live. Make sure anything built with AI, like a report or a reusable prompt, lives somewhere the whole team can open. Then review it with the team each month, updating the prompt library and the policy as the work changes.
Bringing shadow AI into the open
Agents adopting AI on their own is a sign they want to do their jobs better. A support leader’s job is to give that a shared home: one visible connection to your helpdesk with limits you set, plus a prompt library the whole team draws from. That turns dozens of private experiments into one workflow you can see and improve.
Start with step one this week. Ask your team what they’re already using. Their answers will show you how far along your team is and where to focus first.
Running the monthly review on real tickets
The monthly review in step six is what keeps the policy and the prompt library useful, but it usually runs on what people remember. Agents mention the prompt that worked well and the reply that went wrong last week, and a team lead might reread ten or twenty tickets by hand. So the policy changes based on a handful of examples, while most of the replies your team drafted with AI never get looked at.
Swifteq’s MCP Server for Zendesk connects Claude or ChatGPT to your Zendesk, so the same team setup from step three can also feed that review. A team lead can give Claude the drafting standards from step five and ask it to score last month’s solved tickets against them. For every point it marks down, Claude gives the reason, for example a reply that promised a refund outside your policy, so the review starts from a short list of real tickets instead of impressions. If you save those standards as a Claude skill, a set of saved instructions Claude follows each time, every review checks the same things, whoever runs it.
Because each person connects with their own Zendesk login, anything Claude changes on a ticket for them, like an internal note or a tag, is recorded under their name in Zendesk, the same as a change they made by hand. That keeps the work people do with AI in the ticket history your team lead already reads, instead of in someone’s personal chat account where nobody else can open it.
The MCP Server is free for 500 calls a month per Zendesk account, shared by everyone at your company, with no credit card required. Each ticket Claude reads is one call, so reviewing 100 tickets takes about 101 calls. Once the whole team uses it every day, €49 a month covers your account and its first user, and each extra user is €9 a month. Connect Claude or ChatGPT to Zendesk and invite your team into one account.


