When AI agents are discussed, it often stops at demos and pilot projects. Platform provider Vercel offers an example from live operations: Julian Eckerle, Head of Central Europe at Vercel, described at the t3n format Arbeit in Progress how the company has integrated AI agents into almost every business area.
Two use cases stand out. In inbound sales – that is, enquiries that come in from interested parties on their own initiative – an agent handles processing. In customer support, another agent answers 91 percent of all tickets according to the company. Jobs have not been lost as a result.
Why these two processes in particular
Inbound sales and first-level support are not a random choice. Both areas have characteristics that make them accessible to automation:
- High volume with recurring patterns. A large proportion of enquiries are very similar in content.
- A reliable knowledge base. Documentation, previous tickets and product information provide the material from which answers are generated.
- Clear success criteria. Whether a ticket has been resolved or an enquiry could be qualified and passed on can be measured.
- A regulated exit. Anything the agent cannot conclusively clarify goes to humans.
Precisely the last point also explains the figure of 91 percent. It does not describe fully automated support, but a division of labour: the agent covers the well-structured part, and the remaining cases land with people who have more time for them.
An agent is more than a chatbot
The term AI agent refers to a system that not only generates text but, within defined limits, executes steps: retrieving information from systems, categorising enquiries, formulating answers, forwarding or completing cases. The difference from a classic chatbot lies less in the language model than in the integration – with the ticket system, CRM, documentation and product data.
This also shifts the effort. The model itself is interchangeable and available from providers. The actual work lies in integration, in clean data sources and in the question of which actions an agent may perform independently and which it may not.
No layoffs – what that means and what it does not
The fact that, according to the company's account, no jobs were lost at Vercel is a remarkable statement, but one that requires explanation. It does not necessarily mean that nothing has changed. More plausible is that activities shift: away from the initial processing of uniform enquiries, towards more complex cases, maintenance of the knowledge base and monitoring of the agents.
For growing companies, this is a realistic pattern. When enquiry volume rises, automation can absorb additional load without the team having to grow to the same extent. However, no conclusion can be drawn from this about companies with stable or declining business – no information on that is available.
What can be transferred to your own operations
The example comes from a provider with pronounced technical expertise and digitally shaped processes. What is transferable is less the specific figure than the approach. For small and medium-sized companies with a website, online shop or customer portal, comparable starting points present themselves:
- Pre-qualify enquiries via contact forms, enrich them with context and route them to the responsible person instead of leaving them in a shared mailbox.
- Answer recurring support questions about orders, access, invoices or features based on your own documentation.
- Editorial support in the content management system, such as suggestions for metadata, summaries or translations that are approved editorially.
- Internal search across manuals, policies and project documentation so that employees can find answers without interrupting colleagues.
The prerequisites are underestimated
The biggest mistake in such projects is to start with the agent. The reverse order makes more sense. First, a process is needed that can be described at all: which enquiries come in, how are they handled today, where do waiting times arise?
Then comes the data basis. A support agent is only as good as the documentation it draws from. Outdated content, contradictory information on the website and in internal files, or missing product information lead to answers that sound formally convincing and are factually wrong.
And finally, measurement points are needed. Without metrics on resolution rate, escalations and satisfaction, it is impossible to assess whether the agent is helping or when it needs to be refined. A rate of 91 percent is only a reliable statement if there is a definition behind it of what counts as resolved.
Name limits openly
Not every process is suitable. Binding commitments, legally delicate information, price negotiations or complaint cases with an emotional component remain with humans. Transparency towards customers is also an issue: anyone speaking with an automated system should be able to recognise that and have a simple way to reach a human contact person.
There is also the question of dependency. Anyone who ties sales and support to agents needs a plan for outages, model changes and changed terms from the provider.
A clean separation between your own logic and a replaceable model pays off here.
Context for web and software projects
For website and portal projects, this shifts the focus: the quality of content, data structures and interfaces will in future determine not only findability, but also how reliably agents can work on them. Anyone who organises documentation, product data and ticket processes today creates the foundation for automation without committing to a specific provider. A realistic approach is to start with a clearly defined use case with a measurable result – not with the simultaneous migration of all customer channels.
Sources
- No layoffs? How Vercel replaces employees with AI agents — t3n.de - News