The AI workshop is over. There are ten ideas on the whiteboard and yesterday’s enquiries are still waiting in the inbox. Which idea should take on some actual work? Choose one recurring task whose output your team can check, and turn it into a clearly bounded pilot brief.
A useful first step could be a draft reply: AI adds approved information to an enquiry, then a colleague checks and sends it. That gives you something concrete to test. Taking on the entire customer service process leaves far more questions open. Leave your AI workshop with one task whose inputs, output and owner are clear.
1. Choose one recurring task
Describe where the task starts and ends. “Handle enquiries faster” might become: an incoming product question is matched with approved information, and a colleague receives a draft reply with sources and unresolved points. They check the content and approve it for sending.
Start with one common type of enquiry. Write down what makes the result usable and which cases go straight to a person. A price negotiation or a complaint may need individual judgment. A bounded workflow is easier to walk through with the people involved. Keep the other ideas for a later decision.
2. Gather real cases and permitted sources
Collect examples from everyday work: typical requests, useful past replies and awkward cases. Include incomplete messages, conflicting documents and questions your sources cannot answer. These cases reveal whether the system makes uncertainty visible and hands the work over with enough context.
Specify which documents and systems the pilot may use. Who maintains the product information? Which version is current? Who may see each piece of information? Check this with the people doing the work today. Where two files disagree, settle which source applies before testing. The assistant can then use that decision.
3. Choose the smallest suitable implementation
Work back from the required output. An existing tool may already cover the task once its sources and settings are configured. You may need a connection to the inbox and a handover to your business system. Custom development makes sense where existing tools cannot reliably support the required workflow.
An internal AI assistant can help with questions about company documents. Our guide to knowledge systems with evidence explains why answers should show their sources. If the solution needs to carry out several steps, AI process automation covers the handovers and approvals. Choose the technology around that workflow; a dedicated agent needs a specific job.
4. Test with the people who will use it
Ask the responsible colleagues to handle the same types of case using the current and proposed workflows. Record total handling time, including review and corrections. Note missed cases and which outputs were actually usable. A quick draft that needs lengthy editing belongs in the full comparison.
Discuss the differences together. Was a source missing, was the task unclear, or did the system misread available information? Each calls for a different correction. Keep some cases aside for a later test so changes are also checked against fresh examples. Agree in advance which errors would prevent launch and who makes the release decision.
5. Put an owner behind the operation
Before launch, name the person responsible for the pilot. Agree who uses it, checks exceptions, maintains sources and requests changes. The team should know where to report a wrong result and how to continue with the previous workflow during an outage.
Record access arrangements, connected systems and the agreed support. Rerun the relevant test cases after changes to sources or integrations. AI Værk’s work includes introduction, ongoing operations and development. The scope is agreed for the particular system, so responsibility remains clear after the first successful test.
A pilot brief in five fields
| Field | Pilot agreement |
|---|---|
| Task and output | A product question arrives; the result is a draft reply with sources and unresolved points. |
| Users | The responsible service team reviews the drafts and tests everyday and difficult enquiries. |
| Sources and systems | Approved product information and permitted correspondence; access is agreed before testing. |
| Human decision | A colleague checks and sends. Price negotiations, complaints and missing evidence go directly to a person. |
| Success check | Compare usable replies and total handling time with the current workflow, including review and corrections. |
How a first step grows into ongoing work
At Kirsch & Drechsler, an initial assessment and workshops grew into a long-term collaboration. It includes internal knowledge management, a website chatbot, ongoing support and the joint development of CherryGate. For your own first step, consider what foundation the pilot could create for further applications.
Bring your chosen task, the systems involved and a few typical cases to the next conversation. Together, we can assess which implementation fits and what needs preparing before the first use.
