Optimisation

What the organisation is actually doing with the AI it pays for, read across every seat and every vendor that reports. Findings are things you could act on this week; the tables underneath are the evidence for them, and the bottom of the page names what we still cannot see.

Seats reporting
8
of 8 registered
Tokens / 7d
553.4M
across the estate
Tool runs
1.0k
21 distinct tools
True cost / 7d
$0.0000
$524.65 at list price, not billed

What we found

  • Decide

    2 different models across 4 seats

    james.o@naturebaby.com → claude-sonnet-5; Claudia.Z@naturebaby.com → claude-sonnet-5; don.p@naturebaby.com → claude-opus-5; jacob@naturebaby.com → claude-sonnet-5. Same product, different tiers. Worth deciding whether that split is intentional or just how each person happened to be set up.

  • Watch

    12 tools are used by more than one person

    Bash (328 runs across 7 seats), mcp_tool (182 runs across 5 seats), Read (118 runs across 5 seats), Edit (109 runs across 4 seats), TaskUpdate (56 runs across 4 seats), Write (39 runs across 5 seats), ToolSearch (34 runs across 5 seats), TaskCreate (34 runs across 3 seats), SendUserFile (32 runs across 6 seats), Skill (23 runs across 4 seats), AskUserQuestion (16 runs across 4 seats), Projects (11 runs across 2 seats). Repeated practice across people is where a shared skill or template pays for itself.

  • Decide

    1 tool averages over 5 seconds

    mcp_tool ~9.1s over 182 runs. Slow tools stall the whole turn, and they are usually the cheapest thing on this page to fix.

Model by seat

SeatModelCallsTokensShare
james.o@naturebaby.comclaude-sonnet-5655.3M1%
Claudia.Z@naturebaby.comclaude-sonnet-5667141.5M26%
don.p@naturebaby.comclaude-opus-525130.2M5%
jacob@naturebaby.comclaude-sonnet-5171.4M0%
hengda.q@naturebaby.comno calls in the window
anna.t@naturebaby.comno calls in the window
amy.w@naturebaby.comno calls in the window
isobel.a@naturebaby.comno calls in the window

Tools across the estate

ToolRunsSeatsAvgRejected
Bash32873.2s
mcp_tool18259.1s
Read118587ms
Edit109412ms
TaskUpdate56417ms
Write39522ms1
ToolSearch3458ms
TaskCreate34322ms
SendUserFile326711ms
Skill23417ms
AskUserQuestion1642ms
Projects1121.1s
Grep4219ms
Glob32166ms
WebFetch313.1s
SearchMcpRegistry21459ms
WebSearch21725ms
PushNotification11
Agent11203.1s
ListConnectors11370ms
SendUserMessage111ms

Rejected counts the times a person declined an edit the agent proposed. They are not failures – they are the only place a harness tells us a human pushed back, which is the control plane working.

Where a human stood in the way

Decisions
122
a person accepted or refused
Refused
1
of those 122
Awaiting approval
0
queued right now

1 of 122 decisions went against the agent.

What this cannot see yet

Discovery

Needs a connector into a system of record

Finding AI nobody registered means reading SSO logs, browser telemetry or a SaaS admin API. Telemetry only ever shows what already reports to us, which is the opposite problem.

Entitlements

Needs a scopes read per harness

Whether an agent can reach more than it needs is a comparison between what it was granted and what it has used. We hold the second half; the first has to be pulled from each vendor.

Benefit

Needs telling us what the work was worth

Hours saved and deals closed are not observable from a token counter. The cost side above is measured; the return has to come from you, per use case.