An AI assistant on your website, portal or WhatsApp can answer routine questions itself and pass everything else to your people. Whether that works depends on one moment: the handover, when the conversation becomes a ticket in Zendesk, Freshdesk, Salesforce Service Cloud, Microsoft Dynamics 365 Customer Service or TOPdesk. Done well, the ticket arrives complete, lands in the right queue, and the ticket system stays the single record of what was asked and done.
This post is for the IT manager, informatiemanager or head of customer service at a 250 to 5,000 FTE organisation that already runs a ticket system and works with an ICT partner. Crux Digits builds the AI side and the connection to your ticket system; your ICT partner keeps running what it runs today.
What does a service team gain when the AI assistant is connected to the ticket system?
Four things: faster first answers, evenings and weekends covered, staff freed for cases that need a person, and consistent answers.
Each gain has a mechanism you can check. The assistant answers from the same knowledge articles your team already maintains, so a question about delivery terms gets the approved wording at 14:00 and at 23:00. Conversations that need a person become tickets overnight, already categorised, so the queue at 08:00 is sorted rather than raw. And because the transcript travels with the ticket, the colleague who picks it up starts at the customer's actual problem instead of asking for the order number again.
Without the connection, staff copy chats into tickets by hand, and contacts the assistant handled never reach your reports.
What must the handover carry into the ticket?
Nine fields: transcript, identity and verification status, category, language, attachments, summary, handover reason, channel and conversation ID.
- Transcript. The full conversation, not a paraphrase, so the agent can see exactly what the assistant said.
- Identity plus verification status. A logged-in customer is verified; an email address typed into a chat is a claim. The ticket should say which of the two it is.
- Category and subcategory, taken from your own category list, never invented by the AI.
- Language, so routing can send a French-speaking customer to someone who answers in French.
- Attachments, such as a photo of the damaged delivery or a PDF invoice, attached to the ticket itself rather than linked to the chat tool.
- A summary of three to five lines, written as an internal note. In Zendesk a comment with public set to false is an internal note; in TOPdesk the initial action can be marked invisible for the caller.
- The handover reason: the customer asked for a person, the question fell outside the knowledge base, it is a complaint, or it needs an action the assistant may not take.
- The channel (website chat, WhatsApp, portal), so reporting per channel keeps working.
- A conversation ID from the AI side. Zendesk has an external_id field "to link Zendesk Support tickets to local records"; TOPdesk has an externalNumber of up to 60 characters. This ID is also what stops a retry from creating a duplicate ticket.
How do Zendesk, Freshdesk, Salesforce, Dynamics 365 and TOPdesk accept a handover?
Zendesk through its messaging switchboard, Freshdesk and TOPdesk through their ticket APIs, Salesforce and Dynamics through their messaging and routing layers.
Zendesk. A third-party bot escalates by calling passControl on the switchboard, which makes Zendesk's Agent Workspace the active integration. The call can carry metadata: standard fields such as group_id, brand_id, tags and priority via dataCapture.systemField, custom ticket fields via dataCapture.ticketField, and a first_message_id that sets where the injected conversation history starts. One detail matters for design: Zendesk respects custom-field metadata only on the first escalation of a ticket, so the category must be right the first time. Files go through the Uploads API, whose token is valid for 60 minutes.
Freshdesk. The Create Ticket API needs one requester attribute (email, phone, requester_id or another listed identifier) and accepts group_id, type, tags and custom_fields. Source 7 marks the ticket as chat. If no contact exists for the email address, Freshdesk creates one. Attachments go as multipart form data with a total limit of 20 MB.
Salesforce Service Cloud. With Messaging for In-App and Web, the assistant can pass context through hidden pre-chat fields, which the customer does not see. A custom parameter maps each field to a flow variable in an Omni-Channel flow, which can route the conversation and write the value onto the Messaging Session record.
Microsoft Dynamics 365 Customer Service. A Copilot Studio assistant hands over with "the full history of the conversation and all variables collected", and Dynamics routes the escalation to a queue. It needs a Copilot Studio licence and the Chat Add-in for Dynamics 365 Customer Service, in the same environment.
TOPdesk. The incident API creates first-line incidents. A few rules from the specification shape the mapping: the brief description is limited to 80 characters, a subcategory that does not belong to the given category returns an error, and a caller can be looked up by email only if that person has the Caller checkbox set. Files are uploaded per incident, with a flag that hides them from the caller.
How does an escalated conversation reach the right queue?
The AI fills the fields your routing rules already read, and the ticket system's own routing picks the person. The assistant never assigns a ticket to a named employee.
All five systems have routing built in. Zendesk omnichannel routing assigns on agent status and capacity on every plan, with custom queues, skills and priority on Professional and above. Freshdesk's Omniroute offers round-robin, load-based and skill-based assignment within groups. Salesforce Omni-Channel uses queue-based or skill-based routing, and a work item uses one model or the other, never both. Dynamics 365 unified routing classifies a work item with rules or machine learning, then assigns on skills, availability and workload.
So find the fields that drive routing today, and have the assistant fill exactly those. Three choices usually cover it.
- Language and product line become skills where your plan supports skills, or groups where it does not.
- Every handover reason maps to one queue. A complaint goes to the complaints group, a billing question to finance support.
- One fallback queue, for example "AI handover: unclassified", catches anything the assistant could not place, and one named team lead owns it.
A worked example, with assumptions stated: a Dutch technical wholesaler on Freshdesk, with groups for orders, returns and invoices, and two product lines, pumps and valves. The assistant sets group_id from the category and adds a product-line tag, and Omniroute's skill-based assignment sends pump questions to the colleagues with the pump skill. Nothing about routing moves into the AI.
How do you know who the customer is before a ticket exists?
Only a logged-in session gives a verified identity. Anything a customer types into a chat is a claim, and the ticket should record which of the two it received.
Zendesk supports authenticated messaging with a JWT signed by your own server, using external_id as the primary identifier; authenticated users get a check mark in the agent view. It works for the Web Widget and the mobile SDKs. Copilot Studio supports single sign-on with Microsoft Entra ID for users who are logged in to the page where the assistant runs.
For unverified customers, two design choices prevent damage. First, match before you create: Freshdesk creates a new contact for an unknown email address, so a typo produces a duplicate contact unless the connector searches first. Second, gate account-specific answers. General questions need no identity; order status, contract details and anything personal need a verified session.
Why does the ticket system stay the system of record?
Because SLAs, reports, audits and GDPR requests already run on it; a second record of each contact doubles that work.
Under the GDPR, a customer can ask for a copy of the personal data you process (Article 15(3)), and you answer within one month (Article 12(3)). If transcripts live in both the ticket system and the AI supplier's database, both are in scope. The storage-limitation principle in Article 5(1)(e) says data is kept no longer than the purpose needs, and Article 28(3) requires a contract with your processor that sets out the duration of processing.
The practical rule: if a fact matters after the conversation, it goes into the ticket. The AI side keeps its own log only as long as handover and quality review need it, and that period is written into the processor agreement. Reports on volume, handling time and topics come from the ticket system, so they include AI-handled and human-handled contacts in one view.
Who owns which part: service team, IT, the ICT partner or the AI supplier?
Four owners: the service team for content, IT for the ticket system, the ICT partner for identity and network, the AI supplier for assistant and mapping.
The service team decides which questions the assistant may answer, maintains the knowledge articles, defines the category list and the handover reasons, and designs the queues. IT, usually the application manager of the ticket system, creates the API user with the least rights that work, owns the field schema, and puts every change to fields or routing through the normal change process. In TOPdesk, for example, an operator account needs create permission on first-line incidents, and edit permission to upload attachments. Nothing broader is needed for the handover.
The ICT partner keeps doing what it does today: the identity provider, the network and firewall rules, and the accounts. The AI supplier owns the assistant, the mapping table from conversation to ticket fields, the evaluation set, and its own transcript retention. Write down who watches the fallback queue and who is told when handovers fail. Our post on AI incident ownership covers that split in more detail.
Should you buy an assistant off the shelf or build one around your ticket system?
Buy when your channels are standard and a ready connector exists; build when routing, identity or the ticket system itself needs custom mapping.
The ticket vendors sell their own AI options too. TOPdesk lists TOPdesk Robin for end users in its Excellent package, and Microsoft's route for Dynamics is Copilot Studio. Compare those first if you already run that platform.
SimplyBoost is our own product, so weigh this paragraph with that in mind. SimplyBoost is an AI support agent that answers from your website, PDFs, FAQ and catalogue on website chat, WhatsApp, Instagram and Facebook Messenger. An account includes three AI agents. It hands over to a person, or to Zendesk, Freshdesk, Gorgias or Slack, with the conversation attached. An agent can run on logged-in portal pages and recognise users by verified email, and it replies in Dutch, French, German, Spanish and English. For a team on Zendesk or Freshdesk with standard queues, that is often enough.
A build makes sense when the handover must land in Salesforce, Dynamics 365 or TOPdesk with custom fields, when routing depends on data from your ERP or CRM, or when identity comes from your own customer portal.
Where does Crux Digits fit?
We design and build the AI assistant and its connection to your ticket system, alongside your ICT partner. We do not run your ticket system, your helpdesk or your infrastructure.
The work starts with the mapping: your categories, queues, identity sources and the nine handover fields, tested against real sample conversations before anything goes live. The AI audit costs €2,500, a Production-ready MVP €20,000, and production from €50,000, all excluding VAT, and the code is handed over to you. For a broader rollout across departments, see AI implementation in the Netherlands.
More on the service: AI customer support agent.
Last updated: 28 October 2026.
SimplyBoost, our own product, answers from your documents on website chat and WhatsApp. When the answer has to come from your own systems, we build that part.
AI customer service agent →
