A team built an AI agent for ServiceNow. The wins came from narrow jobs, not the broad roadmap

The most useful AI agent write-up this week is not a launch. It is a team admitting their roadmap was too broad.

What happened

In a VentureBeat piece, Brian King, a Senior AI Product Engineer at Bytemethod.ai, and Richard Mendis described building a "Digital Worker" for ServiceNow IT service management. It began on browser automation and later moved to Model Context Protocol (MCP) so it could call APIs directly and finish tasks faster. They built their own harness that supports several LLMs and pairs the model's reasoning with deterministic code, which they say is what turns a system that can answer ServiceNow questions into one that performs tasks consistently.

The guardrails are plain: the agent connects through OAuth using the user's own ServiceNow identity, so role based access and permissions still apply, and a human reviews and approves any write operation before it runs.

The results came from three narrow jobs:

• Catalog items. From uploaded requirements in their standard format, it creates a catalog item, workflow and scripts included, in about 20 seconds. They expect this to cut development costs by 80% and reclaim about 25% of team capacity.
• Trend analysis. Work that could take several hours becomes a 10 to 15 minute conversation.
• License reviews. Role audits and premium license checks that used to happen yearly can now run on an ongoing basis.

Their closing advice: start with one workflow, not a broad ITSM agent. Pick something repetitive, rules based, high volume, easy to measure and easy to reverse.

My take

This matches what I see in client work. The agent that gets approved is never "an AI for our CRM". It is "an agent that builds the same kind of record our admin builds 40 times a month".

Three things worth copying:

1. Let the agent borrow the user's permissions. It cannot do anything the person could not already do, so security review gets much shorter.
2. Gate writes, not reads. Reading and summarising is safe to automate fully. Changing records goes through an approval step until the error rate earns trust.
3. Mix the model with plain code. The model interprets the request. Validation and the API call stay deterministic.

Notice the numbers too. The 20 seconds is measured; the 80% is a projection. Report both kinds honestly to your clients.

If you run HubSpot, GoHighLevel or Zoho instead of ServiceNow, the pattern carries over unchanged. I build these narrow, permission scoped agents for teams, and you can see the kind of work I do at romielwillautomate.dev.