Field notes

Insights

Notes from delivery. Patterns we keep reusing and mistakes we'd rather you didn't repeat.

All posts

Token consumption: an aggregate bill cannot be split back apart

Finance did not ask what the agents cost. They asked which business unit owes what, which is a different question and cannot be answered from a total after the fact.

Agent observability: nothing breaks when the answer is wrong

The defect that reaches a user is fluent, plausible and wrong, and no suite raises a flag on it. What production measurement catches, and what late discovery costs.

Agent discoverability: know what shipped and who approved it

Which agents are running, who signed off on each, and what rolling one back involves: usually the answer is a file somebody edits by hand. It is true until it isn't.

Shadow AI: a ban removes the visibility, not the demand

The agent nobody sanctioned has no owner, no off switch, and whatever access its creator happened to hold. The controls that work make the approved path the faster one.

What has to be true before a non-engineer can safely ship an agent

Governance, environment, identity and guardrails have to be proven before the first non-engineer ships an internal app, not layered on after it already shipped.

A runbook only counts if every failure mode in it actually happened

Handover isn't documentation plus a call. It's access narrowed, our own access removed, and a runbook proven against failures that actually happened, not ones we imagined.

A green test suite doesn't mean your agent is ready for pilot

Unit tests, data invariants, business prompts and a faithfulness check each catch a different failure, and a demo can pass every one of them while a real question still breaks it.

Your first enterprise agent should be recommend-only

Write-back is where the risk is, and it is also where most of the value isn't. Start with an agent that tells people what to look at.

The pattern we use to put agents on enterprise data

Source system to BigQuery, curated views, MCP Toolbox on Cloud Run, an agent runtime with no database access of its own. Here is why each piece is there.

Gemini Enterprise rollouts: what actually takes the time

It is not the AI. Across six rollouts this year, the schedule risk lived in identity, connector allowlisting and naming confusion.

Get in touch

Talk to us

Tell us what system the answer lives in and who needs it. We'll reply with a view on whether it's a two-week assessment, a five-week pilot, or something else.

Start a conversation →

or info@insightnext.tech

InsightNext on LinkedIn