GOAL: qualified demo requests · AUDIENCE: ops leads, IT, finance · CTA: Book a demo (primary, sales-led product), Start free secondary. The objection is “will it do something without asking?”, so the hero answers it.

CAIRN AGENTS

Agents that ask before they act.

Automate the busywork between your tools. Every step is logged, every external action waits for a person, and every run shows what it cost.

Book a demo Start free

WORKFLOW · INVOICE MATCHING

It starts with a trigger.

1 / 5

An email, a form, a schedule or a webhook. You pick what wakes the agent up.

TRIGGER● Running
New invoice in finance@ gmail.watch · PDF attached
TOOL
Extract line items doc.parse · Pebble 1.4 · 14 items
TOOL
Match to purchase order erp.lookup · PO-2291 · Δ $412
HUMAN
Approve the payment? Rule: difference > $250 → Finance lead
RESULT
Payment scheduled erp.pay · Oct 3 · audit #8f2c

Scroll to step through the run, or use the buttons. The agent is paused. Scrolling won’t continue it; only your decision will. Use the buttons to step through the run.

Every run leaves a receipt

Run	Started	Outcome	Decided by	Model · tools	Time	Cost
run_8f2c	Sep 24 · 09:14	Paid	You	Pebble 1.4 · 4 tools	6.8 s	$0.021
run_8f1a	Sep 24 · 08:52	Paid	Auto (< $250 rule)	Pebble 1.4 · 3 tools	3.1 s	$0.009
run_8e97	Sep 23 · 17:30	Rejected	M. Lind, Finance	Ridge 2.0 · 4 tools	5.4 s	$0.038
run_8e40	Sep 23 · 15:02	Needs review	Waiting · 2 h	Pebble 1.4 · 4 tools	—	$0.012
run_8d11	Sep 23 · 11:47	Failed safely	—	erp.lookup timeout	30.0 s	$0.002

Illustrative data · exportable to CSV and your SIEM

Guardrails you set

Human approvalon: pay | send | delete | publishExternal actions always pause for a named person. Approvals expire after 24 h.
Spend cap per runmax_cost: $0.50The run stops before exceeding it and tells you why.
Allowed tools onlytools: [gmail.read, erp.lookup, erp.pay]Agents can’t discover or call anything else.
Data boundariessources: finance/*Agents read only the folders you allow, with the owner’s permissions.
Kill switchadmin: pause allOne click pauses every agent in the workspace.

Evaluated before release

Agent task success on AgentBench-Ops (internal), 600 tasks, Aug 2026. Bars show 95% confidence intervals. Illustrative.

Summit 2.1 · task success88% ± 2.6

Ridge 2.0 · task success81% ± 3.1

Correctly asked for approval99.2% ± 0.7

Stopped at budget cap100% ± 0.4

Axis 0–100%. Known limit: success drops to 71% on tasks with more than 12 steps.

MICRO-CONVERSION: ROI calculator. Conservative defaults, the assumptions are visible, and the result feeds the demo form.

ROI CALCULATOR · HOURS SAVED

People doing this work12

Hours per person per week5 h

Share an agent can take on40%

Loaded hourly cost$65

Estimated hours back per year 1,104

Value of time$71,760

Per person / week2.0 h

Assumes 46 working weeks and that people still review every approval (≈ 1 min each). Pilot results vary; we measure yours.

Book a demo with these numbers