FocusNow AI Club, meet-up #2 – 28 May 2026.
The second AI Club session was built around a live demonstration: a working application, built in two days, that takes a requirement written in plain business language and ends with configuration deployed to a ServiceNow instance.
The pipeline
A requirement goes in as a business analyst would write it – in the example shown, a business-unit split where stakeholders need additional notifications on related incidents. The tool generates user stories from it, then checks them against the target instance to surface overlaps with configuration that already exists.
Each story arrives with acceptance criteria and the artefacts it would create: an email notification, a table, a business rule. Stories can be edited, approved, rejected or deleted, then grouped into a release. Pushing the release generates the code and deploys it – to a scoped application or global scope – using ServiceNow’s own SDK. Automated test setup through ATF sits at the same step.
For SPM and delivery work this is the direction that matters. The value is not in generating text about a requirement; it is in compressing the distance between an agreed requirement and working configuration, with a human approving each story before anything reaches the instance.
What it took, and what went wrong
Two days of design and build, running on a €12/month VPS with 16GB of RAM behind a private network.
The obstacle was ServiceNow-specific. Early attempts to work with the ServiceNow SDK produced hallucinated methods and broken code, and it took several iterations before output became dependable. That matches the pattern across every session so far: AI accelerates ServiceNow work substantially, but only for someone who can recognise a wrong answer when it appears.
Paying for it sensibly
A practical cost point emerged. Most teams route everything to the most capable model by default, which is the expensive way to work. The better pattern is to let a strong model plan and orchestrate while cheaper, faster models execute the well-defined steps beneath it. On long-running tasks the difference is substantial, and multi-model gateways make the switching straightforward.
Key takeaways
- Requirements-to-configuration pipelines are buildable now – two days of work produced a functioning one.
- Keep human approval between story generation and deployment. Review is where the quality comes from.
- The ServiceNow SDK is the deployment path, but expect hallucinated code until prompts are tuned.
- Match the model to the task. Reserve the expensive model for planning, not execution.
Join the next one
The FocusNow AI Club meets monthly, with working demonstrations rather than slideware – and our consultants are currently certifying on Claude. If you are weighing where AI belongs in your ServiceNow roadmap, get in touch.
Source: ServiceNow SDK documentation