For agents

Agents that already know your business.

Canopy is what external agents connect to. Canopy Agents are the ones KPI Tree runs for you: platform agents working out of the box on your metrics, with your permissions, plus, on Enterprise, your own agents deployed on the same context.

Canopy
Canopy Agents
Canopy Agentic Workflows

Working from day one. Governed from day one.

Platform agents, out of the box. Ready-made

The personalised action plan, the daily and weekly briefings, RACI assignment, canvas edits and Slack answers, with no setup beyond connecting your data.

Grounded, governed, gated. Trust built in

Your permissions on every run, approval gates where you want them, and escalation up your real org chart. Agents know when to act and when to ask.

Any model. Your keys. Your caps. Cost under control

Claude, OpenAI or Gemini per run, bring your own keys, workspace spend caps and per-agent cost tracking.

The evidence

Anthropic measured it. Context decides accuracy. Not the model.

Anthropic automates 95% of its business analytics queries with Claude. The accuracy comes from the hand-built context layer between the model and the data, not from the model. Canopy Agents run on that kind of layer: grounded in your tested causal model, your definitions and your permissions.

Read our full analysis

Anthropic engineering blog

“Without skills, Claude’s ability to answer analytics questions accurately didn’t exceed 21% on our evals.”

21%

The ceiling on Anthropic’s internal evals without curated context.

~95%

The same model on a curated context layer, in production today.

How Anthropic enables self-service data analytics with Claude

Platform agents, working on day one

Personalised Action Plan identifies declining metrics, their drivers, and the specific actions that fall to you under your RACI. The daily briefing and the weekly review land before the day starts, from the same context. Update RACI Assignments finds unowned metrics and proposes Responsible, Accountable, Consulted and Informed owners with reasoning. Canvas Assistant shapes the metric tree itself. And the Slack Assistant answers and acts on metric questions where your team already talks.

Personalised Action Plan

Agent

Declining metrics, their drivers, and the actions that fall to each owner.

Weekdays 07:00 ~2 min
Run

Update RACI Assignments

Agent

Finds unowned metrics and proposes owners, with reasoning.

Mondays ~1 min
Run
Run history
Succeeded07:02action_planScheduled12k in / 1.8k out£0.04
Succeeded06:41update_raciScheduled8.3k in / 900 out£0.02
SucceededYesterdayaction_planWorkflow14k in / 2.1k out£0.05

Grounded in your context, not the internet

Every agent reads Canopy: the statistically tested causal model, the live org chart and RACI, verified outcome history, budgets and plans, and the objectives and initiatives the business is pursuing. So an agent's recommendation is the same plan a well-briefed person would produce, grounded in which levers have actually moved which numbers in this business.
Liverpool Street
CausalModerate

Revenue

£384k

4.2%

AOV

£7.25

5.8%

Orders

52,998

2.1%

Knows when to act, and when to ask

Actions can sit behind approval gates: the agent proposes, a named person approves, rejections branch to their own path. When nobody responds, escalation follows your real reporting lines. You choose where the line sits between suggest, approve and act.
Trigger
Fired
On target missedMetrics
Completed
Generate Action PlanOpus 5.1Agents
Waiting
Wait for Sarah's approvalUtilities
Escalate to managerWorkdayOwnership

Your permissions, honoured on every run

Agents act on your behalf with your permissions, so what an agent sees is exactly what you see, metric by metric. There is no separate service account with god-mode access, and nothing to audit beyond the access model you already govern.

Revenue

OverviewDataRACITasksContext

Who owns this metric, who acts on it, and who is kept in the loop.

Suggest

Responsible

David Mitchell+1 more

David Mitchell

Area Manager, Operations

Emma Thompson

Growth Lead, Marketing

Laura Fitzgerald

Head of Sales, Revenue

Accountable

Sarah Chen

Consulted

Emma Thompson+2 more

Informed

Nobody

CloseSave

Any model. Your keys. Your caps.

Run agents on Claude, OpenAI or Gemini and switch models per run. Bring your own API keys if you prefer to bill LLM usage to your own account. Set a daily usage cap per workspace, track spend per agent and per user in the AI Usage view, and turn on content moderation when your industry requires it.

Weekly revenue review

scheduled · Mondays
ModelClaudeOpus 5.1
Spend this month£4.20 of £25 cap
Actions require approvalon

A tenth of the tokens. None of your warehouse.

Every agent on this page reads Canopy rather than raw tables. An investigation that costs an agent on the warehouse roughly ten queries, with every result set hauled through the context window, is one precomputed call here. The same answer arrives with an order of magnitude fewer tokens, the aggregation work runs in KPI Tree's engine rather than your warehouse, and the spend that remains is visible per agent and per user in AI Usage.
Token cost comparison loading

Scheduled, triggered, or on demand

Schedule an agent on a cron-like frequency in your timezone, run it on demand, or let the metrics themselves start it through Canopy Agentic Workflows: a missed target, a threshold crossing, a metric gone silent. The morning action plan that lands before you sit down is the same agent, on a schedule.
KPI Tree

KPI Tree app · 08:00

Good morning, Sarah. Tuesday briefing: revenue is tracking £31.5k/day behind target, down 13.1% MTD. Three things need you.

Conversion is the driver, not traffic. The £31.5k/day gap traces to checkout conversion. Sessions are flat.
Liverpool Street down 40.9% MTD. £26.9k against £45.5k in February, your single biggest drag. You are Accountable.
Checkout fix verified: +£32k. Impact confirmed against the causal baseline, 14 days on.
Open briefingView my metrics

Or build your own, on the same context

Create an agent in KPI Tree by describing the job in plain English, choosing the model it runs on and the tools it may call, then give it a schedule in a workflow. Or bring an agent of your own over MCP. Either way it gets everything the platform agents get: the causal model, RACI, verified impact, plan context and the strategic tier of objectives, key results and initiatives, with the same permission model and run history. Shared run history means the whole team can see what has run, what it did, and what it cost. The platform agents come with Growth; building your own is part of the Enterprise plan.

Name

Labour cost watch

Description

Monday check of every site's labour cost against the 28% target

Model

Claude Sonnet 5

System prompt

Every Monday, compare labour cost % for each site against the 28% target. For any site over target, name the Responsible manager and file one task with the gap in pounds and the driver behind it. If every site is under target, say so in one line and stop.

Tools (4)
get_metric_calculationscompare_dimension_metricsfetch_metric_raci_metricscreate_task

Use this agent in a workflow

Run it on a schedule, in response to a metric or task event, or as one step in a longer flow. Send the result to Slack, email, or another step.

Open in Workflows

Common questions

Can an agent act without a human approving?

Only where you allow it. Actions can sit behind wait-for-approval steps with named reviewers; rejections branch explicitly, and escalation follows your org chart when nobody responds. Where you want full autonomy, you can grant it deliberately.

What data does an agent see?

Exactly what the user it runs as can see. Agents inherit your permissions and RACI scope; there is no separate all-access service account.

Can we build our own agents?

Yes, on the Enterprise plan. Describe the job in plain English, choose the model it runs on and the tools it may call, and give it a schedule in a workflow: a Monday labour cost check, a supplier price watch, a Friday pipeline sweep. Your agents get everything the platform agents get, the same context, the same permission model and the same run history with cost. Growth includes the ready-made agents.

Which models can we use?

Anthropic (Claude), OpenAI and Google Gemini out of the box, switchable per run. Bring your own keys for Anthropic, OpenAI or AWS Bedrock if you prefer direct billing.

How do we control spend?

A daily usage cap per workspace, plus the AI Usage view: total spend, cost per agent, cost per user, and the most expensive recent runs. And the baseline is low: agents read Canopy's precomputed context rather than assembling raw queries, so each answer starts an order of magnitude cheaper in tokens.

Can we see what an agent did?

Every run keeps its history: what ran, what it read, what it proposed or did, and what it cost. Agent and workflow management sits under admin control with a shared run history.

Canopy

The business context layer every agent runs on.

Canopy Agentic Workflows

Chain agents, approvals and actions into automated processes.

Metric Ownership

The RACI and escalation model agents respect.

Strategy Execution

The objectives and initiatives agents act in line with.

See it on your own metrics.

Book a demo