Pressure
Every team wants agents now
Marketing builds a GPT, finance connects a spreadsheet, support ships a bot. IT finds out when something goes wrong.
Govern · IT & AI enablement teams
Let every team ship agents.
Keep one view of all of them.
The pressure to put agents in production is real and decentralized. Lustena lets marketing, sales and finance build their own agents while IT keeps the inventory, the limits and the cost picture to say yes with control.
Decentralized AI without visibility is a cost center in disguise
Licenses pile up by team, tokens burn inside agents nobody declared, and the people building them were never trained for the risks. Saying no is not an option; neither is not knowing.
Cost
Costs nobody can see
The first consolidated number arrives on the invoice. Which agents are used, by whom and how much, is unknown until then.
Skills
Teams were never trained for this
Most people using AI at work were never taught what a prompt injection is or which data must not be pasted. Guardrails have to work without being understood.
Roll out in an hour, by policy
Publish the extension through Google Admin, Intune or Jamf. Devices enroll on their own with an enrollment token. No change to the agents, no training session before day one.
Productionenrollment token · marketing-2026
Devices
Browsers enrolled with an enrollment token, with the employee behind each one. A revoked device cannot come back by re-enrolling.
48
enrolled
46
seen today
1
revoked
Devices · 5 of 48
| Device · browser | Employee | Last seen |
|---|---|---|
| MBP-anaChrome 141 · macOS | ana@acme.com | 2m ago |
| WIN-joaoEdge 141 · Windows | joao@acme.com | 11m ago |
| MBP-rafaChrome 141 · macOS | rafa@acme.com | 1h ago |
| LT-0412Chrome 140 · Windows | — | 3d ago |
| MBP-luChrome 141 · macOS | lu@acme.com | revoked · 5d ago |
Guardrails set once, applied everywhere
Declare a limit per agent in six dimensions. The people who understand the risk set it; everyone else just uses the agent. Crossing a limit produces one sentence, not a lecture.
ProductionGPT · Proposals · owner ana@acme.com
Limit
What this agent may use and how far it may go. Evaluated on every session.
Six dimensions
- Toolssend_email, read_sheet, create_contactsend_email, read_sheet, create_contact3 allowed
- MCP serversgmail, drive, hubspotgmail, drive, hubspot3 allowed
- Hosts*.acme.com, api.hubspot.com*.acme.com, api.hubspot.com2 allowed
- Data classesnone personalnone personalcrossed · 24h
- Autonomylevel 2 · draft, never sendlevel 2 · draft, never sendwithin
- Schedulebusiness hours · America/Sao_Paulobusiness hours · America/Sao_Paulowithin
Last 24 h
41 sessions · 1 crossed
Previous
v2 · Sep 22 · v1 · Sep 3
On cross
flag · webhook → #sec-alerts
Shadow AI shows where to deploy next
The GPTs, projects and connectors people set up on their own show what each team already brings to AI. The first scan turns that into a list: where an official agent would pay off, and what to govern first.
Productionfirst scan · day 7
What to govern first
Ranked by data exposure, what each agent can reach and how much it is used.
23
agents found
9
without an owner
4
to govern first
Priority
- 1Renewals GPT · personal accountreads CRM export · 3 × card number this weekgovern first
- 2Claude project · ContractsDrive connector · no ownerclaim
- 3Lead enrichment (LiteLLM)2.9M tok · no limit declaredset limit
- 4Support bot (LangGraph)limit v2 · 0 crossedok
Where to deploy next
- Sales · 438 sessions · 3 personal GPTsProposals, follow-ups, renewalsofficial agent
- Finance · 141 sessions · 2 connectorsSpreadsheet analysisofficial agent
Adoption and cost with an owner
Sessions per app, per agent and per team from the browser; model and volume per agent from the SDK and the LiteLLM connector. Rationalize licenses with numbers instead of surveys.
Productionlast 7 days
Adoption and cost
Who uses which agent and how much, with an owner for every line of spend.
1,642
sessions
5
teams
12
idle paid seats
Sessions by team
- Marketing612
- Sales438
- Engineering373
- Finance141
- Support78
Model and volume by agent
| Agent · model | Volume | Spend owner |
|---|---|---|
| Support bot (LangGraph)gpt-4.1-mini · SDKowner · platform team | 18.4M tok | platform team |
| Contract review (CrewAI)claude-sonnet-4.5 · LiteLLMowner · joao@acme.com | 6.2M tok | joao@acme.com |
| Lead enrichmentgemini-2.5-flash · LiteLLMowner · rafa@acme.com | 2.9M tok | rafa@acme.com |
Idle paid seats · no session in 30 days
12 of 60 · ChatGPT Business 9 · Claude Team 3
Questions
Do we need an MDM?
An MDM or Google Admin makes the rollout silent. Without one, an install kit with a double-click wizard enrolls a device in a minute.
Which browsers are covered?
Chrome and Edge, and any Chromium browser that accepts Chrome extensions. Firefox and Safari are available on the Enterprise plan.
What about developers?
They add the Python SDK, the MCP server for Claude Code, Cursor and Codex, or the scanner. Their agents appear in the same inventory under the same limits.