For agents
Canopy is the business context layer for AI agents.
Every other context layer helps AI answer. Canopy is the only one that closes the loop: each metric carries what drives it, who owns it, whether the last action actually worked and the strategic bet riding on it. Every action your team takes tests that graph, a real experiment the model keeps score of. So your agents move the business instead of just describing it.
What agents get on Canopy. None of it lives in your warehouse or semantic layer.
Correlation, causality, then proof. Noticed, interrogated, proven
Daily correlations across all your metrics, our own proprietary ML models and an array of statistical tests to separate cause from coincidence, then proof from real experiments: the actions and initiatives your team completes. Most tools stop at the first rung.
Actions that respect your org chart. Live RACI on every metric and objective
Canopy knows who is Responsible, Accountable, Consulted and Informed, and who reports to whom, so an agent notifies, assigns and escalates to the right person, not a mailing list.
Tell it what the data cannot show. Plain-language context, curated
A launch, an absence, a blocker or a decision, added in plain language or ingested from Slack and Notion. Every item records who said it, its weight follows your org chart, and anything unreliable is held for review rather than served.
What one agent works out, every agent keeps. Cross-session memory
A chat session's reasoning dies when it closes. Canopy consolidates what outcomes prove into the graph itself, with who and what formed it, so every agent starts from everything already proven rather than from zero.
Ten queries become one, and the bill stays flat. Our compute, not yours
Comparison periods, rolling totals and outlier checks are precomputed around every metric, and every aggregation runs in KPI Tree's proprietary encrypted in-memory engine, not your warehouse. An order of magnitude fewer tokens, and question volume never reaches your bill.
Plans and reforecasts in context. Not just actuals
Budgets, forecasts, targets and ramp profiles flow through the same pipeline as your actuals, so whether you are ahead of plan is a question an agent can actually answer.
The strategic plan is context too. Objectives down to tasks
Canopy carries the long-term plan alongside the day to day: objectives, key results, the funded initiatives targeting each metric and the tasks beneath them, so an agent weighing today's action knows which strategic bet it serves.
Agents know when to trust a number. Outliers and staleness tracked
Every metric is continuously checked for outliers, gaps and stale syncs, and that status travels with every answer, so an agent qualifies its answer instead of confidently quoting a broken number.
Curated context in, noise kept out
Tell it what the numbers cannot show: a launch shipped on Tuesday, a key person off this week, a metric blocked on a supplier. Anyone can add context in plain language, or connect a Slack channel or Notion page, and agents factor it in wherever the metric appears. Every item records who said it, its weight follows your org chart, and sensitive items stay visible to owners only. And context is curated on the way in: anything unreliable or contradictory is held for review rather than served, because an agent grounded in bad context is worse than an agent with none.
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 a hand-built layer of canonical definitions, curated references and daily evals between the model and the data, not from the model. Canopy is that layer as a product, for any agent you connect.
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.
Five kinds of context. One governed layer.
A warehouse carries the numbers, and none of the organisational knowledge an agent needs to act on them. Canopy holds all five kinds: how the business is structured, how it operates, what its people actually did and what worked, what was true at any point in time, and where it is strategically heading. Miss one and agents hallucinate rules, contradict each other across teams, or drift as the business changes. Everything below is one of these five, in the product.
Active Customers
2,847
NPS
54
Paid Clicks
18,402
Structural context
The causal tree and the org chart. An agent knows what drives a metric, and who sits where.
Revenue
Finance · daily
R
A
COperational context
RACI, workflows and approvals. Who owns each metric, who acts when it moves, and where a human signs off.
Fix checkout flow
Verified · +£32kLinked to Revenue · measured by the actuals pipeline
Pause underperforming ad sets
Measuring…Behavioural context
The decision traces. What your team actually did, and what verifiably worked, measured on outcomes.
Revenue
19 Mar 2025Trailing 30 days vs previous · +2.9%
Retail week (4-5-4) · +0.7%
Same day last year
…20+ frames precomputed for every date
Temporal context
What was true when. Plan against actual and every comparison frame, at any date in your history.
Strategic · the quarter
Churn below 2% · objective
Marketing efficiency · objective
…every bet wired to the metric it moves
Strategic context
The objectives, key results and funded initiatives the business is pursuing, and the metric each bet targets.
Launches, absences, blockers and decisions, on the metric they explain
Context comes in every shape the business does: a launch or price change lands as an event on the metric's timeline, an absence explains a quiet week, a blocker records why nothing is moving, and a decision keeps its rationale attached. Initiative check-ins flow in automatically as the work progresses. All of it surfaces on the metric's Context tab and in every agent's answer, and the Knowledge area is the one place to browse, curate and retire it as things change.
Active Customers
ContextNew pricing went live on Tuesday
eventon the timeline
Head of Sales out this week
absenceowners only
CRM migration blocked on security review
blocker
Chose annual billing push over discounting
decisionrationale kept
On the metric’s Context tab, in every agent answer.
What one agent works out, every agent keeps
A chat session is working memory: whatever an agent reasons out disappears when it closes, and no other agent benefits. Canopy is the long-term memory underneath. The causal tree, the ownership, the outcome history and the context people added persist for every agent and every person, compounding as the business runs. And every piece of memory records who or what formed it, down to the model that extracted it, so trust is inspectable. A typical company brain stops at indexing what was written; this one holds a tested model of what moves the business.
Agent · chat session
“NPS looks like the real driver of Active Customers.”
Session closed. Nothing kept.
Canopy remembers
Active Customers
driver: NPS
“Key account manager off this week”
said by the metric owner · visible to owners
Kept for every agent, with who and what formed it
It learns over time
Most context layers are built once and decay from that day on. Canopy's causal model moves in the opposite direction, because every driver relationship climbs a ladder of evidence: noticed as a correlation, interrogated by our proprietary ML models and an array of statistical tests, then proven by the work your team was doing anyway. Edges climb as evidence accumulates, fall back when reality disagrees, and carry their rung with them, so an agent always knows whether it is quoting a hunch or a proven lever.
Metric
NPS
54
Metric
Active Customers
2,847
Email CTR ↔ Trial Signups · r 0.41
Paid Clicks ↔ Pricing Views · r 0.37
…22,791 pairs tested tonight
01Noticed, nightly
Every pair of metrics is swept for correlation after each sync, seasonality stripped, in and beyond your trees. A candidate at this stage is noticed, not believed.
NPS → Active Customers survives. Most candidates do not.
02Interrogated
Proprietary ML models and a battery of statistical tests, corrected across thousands of pairs, separate cause from coincidence. What survives is strong evidence, still fitted to observation.
Follow-up sprint on detractors
Measuring…MOMarta Okafor · owner
Declared before the result:
Active Customers ↑
03The team acts
An owner’s action declares the metric it intends to move, and the direction, before the result is in. Completed work doubles as a real-world experiment.
Follow-up sprint on detractors
VerifiedActive Customers 2,847 → 2,921 · measured by the actuals pipeline
“New pricing went live” · kept with the metric, for every agent
04Proven and remembered
The outcome, measured by the same pipeline as your actuals, verifies the edge for every agent. Conclusions that stop holding are discounted and rolled back.
Your org chart, in context
Canopy knows who reports to who and keeps an always-current RACI on every metric and objective, so an agent's answer can become an action routed to the right named person rather than a paragraph describing a problem. When someone leaves or moves team, the accountability moves with them. There is no ownership primitive in a warehouse or a semantic view; this is context only the layer above can carry.
James Harrington
CEO
Sarah Chen
VP Revenue
A · RevenueDavid Mitchell
VP Product
Sofia Martinez
VP Marketing
Emma T.
Growth
Laura F.
Sales
Reporting lines and RACI, current on every answer.
It reads the long-term plan, not just today's numbers
A context layer that only knows the day to day produces tactically sensible, strategically blind answers. Canopy holds the strategic tier alongside the tactical one: every objective, its key results, the funded initiatives targeting each metric and the tasks in flight beneath them, all wired to the same tree. Ask an agent whether a dip is worth acting on and it knows the quarter's bet on that driver, the budget behind it, who owns it and what is already being done. Day-to-day recommendations stay aligned to the long-term plan because both live in one model, either side of the same strategic and tactical line.
Strategic · reviewed quarterly
The plan: an objective, the number that proves it, the bet that moves it
Reduce labour cost percentage to below 28%
83%
outcome · Department objective · 1/3 tasks done
Key result: below 28% by quarter end
29.4
reads itself from the pipeline · nobody types a status
Labour cost %, live
moved by
Rota optimisation rollout
40%
execution · £12k of £30k spent · 3 tasks
Tactical · runs daily
The loop: detect, route, act, verify
Detect
Up 4.2% this weekLabour cost %
Outlier-checked against seasonality, so it is signal, not noise.
Route
ASarah is told, with the driver attached
RACI from your directory decides who, the tree decides why.
Act
Task: pilot the new rota at Canary Wharf
Filed inside the rota rollout, so the work counts towards the bet.
Verify
Impact measured on the metric
What worked is remembered; next quarter plans against proof.
Canopy, the business context layer, holds both tiers
People and agents read the same model, over the app or MCP.
Roughly ten queries become one, and the warehouse bill stays flat
An agent investigating a metric straight on the warehouse runs roughly ten queries: the current value, each comparison period, rolling totals, outlier checks, every one a live warehouse hit and a context-window round trip. Canopy precomputes that metadata so a single query returns it, and every aggregation, comparison and correlation runs in KPI Tree's proprietary encrypted in-memory engine, not your warehouse. There are no pre-aggregations to define and no cache warm-ups to schedule; the precompute is automatic for every metric and refreshes as new data lands. Agents burn an order of magnitude fewer tokens, answer at in-memory speed, and your warehouse bill stays flat while question volume grows.
Any date, any grain, every comparison. No query.
Every metric re-aggregates automatically to any granularity, daily, weekly, monthly, quarterly or yearly, while respecting each metric's additivity: sums sum, rates average, balances carry their last value. And for every date in your history, more than twenty comparison frames are already computed: rolling 7 and 30 days, week on week through year on year, every to-date frame against last period and last year, same day last year, retail 4-5-4 calendars, even Black Friday alignment. An agent can time-travel to any date and read all of it in one in-memory call, at answer speed. Budgets and reforecasts aggregate through the same pipeline, side by side with actuals.
Confidence built in
An agent on raw tables cannot tell a stale number from a fresh one, or an outlier from a trend, so it answers confidently either way. Canopy tracks outliers, gaps and staleness automatically and surfaces them as context, so the agent knows when to trust a number and says so when it should not. Budgets, reforecasts and ramp profiles are part of the context too, flowing through the same pipeline as actuals, so an agent knows what the plan actually was, not just what happened.
Same agents. Same MCP. Different context.
Connecting an agent to your warehouse or catalogue is real and good: governed definitions kill hallucinated SQL and permissions are inherited. Pick what your agent sits on today; the comparison is about everything after the number.
MCP| KPI Tree Canopy | ||
|---|---|---|
| Metric Context | ||
| Cross-platform | One warehouse only. Metrics living in other warehouses, databases or tools are invisible to the agent.Governs the Snowflake estate, including external sources it catalogues, but the context stops at the platform boundary.Governs the Databricks estate. Context stops at the platform boundary. | One context layer across Snowflake, BigQuery, Databricks, Redshift, Azure SQL, Postgres and Google Sheets, plus anything with an MCP server. Metrics from different platforms live in the same tree, with the same causality, ownership and actions. |
| Metric definitions | Raw tables and SQL. The agent re-derives every metric per question, so ask twice and you can get two answers.Strong. Semantic views give governed, composable definitions with lineage. This is the layer Canopy syncs from, not a rival to it.Strong. Unity Catalog metrics give governed definitions and lineage. This is the layer Canopy syncs from, not a rival to it. | Governed definitions synced from your semantic layer, identical on every surface, with the causal model on top. |
| Data lineage | Object dependencies and access history exist, but the agent reconstructs lineage query by query.Strong. End-to-end column-level lineage across the Snowflake estate.Strong. Table and column lineage across the Databricks estate. | Full dbt lineage travels over the same MCP: which models and columns feed every metric, with the causal model above it. Where the number came from and what moves it, in one context. |
| Queries per question | Roughly ten warehouse round trips per metric: the value, each comparison period, rolling totals, outlier checks.Definitions are governed but values are still computed per question; every comparison period is another query and another context-window round trip.Definitions are governed but values are still computed per question; every comparison period is another query and another context-window round trip. | Comparison periods, rolling totals and outlier checks are precomputed, so one call returns the full picture. Roughly ten queries become one. |
| Warehouse bill | Every agent question is warehouse compute, and agents ask a lot of questions.Every agent question is still warehouse compute; governance does not change where the queries run.Every agent question is still cluster or SQL warehouse compute; governance does not change where the queries run. | Aggregations, comparisons and correlations run in KPI Tree's proprietary encrypted in-memory engine. The bill stays flat as question volume grows. |
| Data confidence | None. The agent confidently quotes whatever the table returns, including broken numbers.Lineage and popularity signals point at authoritative assets, but there is no per-metric outlier or staleness state attached to an answer.Lineage and tags help discovery, but there is no per-metric outlier or staleness state attached to an answer. | Outliers, gaps and stale syncs are tracked continuously per metric, and that status travels with every answer. |
| Business Context | ||
| Causality | None. The LLM narrates whatever pattern it spots in the moment, and correlation gets presented as causation.Not covered. Semantic views define how metrics are calculated, not what drives them.Not covered. The catalogue defines metrics, not the causal relationships between them. | Driver edges statistically proven daily by our own proprietary ML models and an array of statistical tests, and pruned by the people who know the business. |
| Ownership | None. Ownership lives in people's heads.Role-based access says who may read a table, not who is accountable for the metric moving.ACLs and tags say who may access the data, not who is accountable for the number. | RACI on every metric and objective plus the live org chart, so an agent knows who is Accountable, who to notify and how to escalate. |
| Financial plan context | Plans usually live in planning tools or spreadsheets outside the warehouse, with no shared aggregation logic.Not covered. Governed context describes the data estate, not your budgets and reforecasts.Not covered. The catalogue describes the data estate, not your budgets and reforecasts. | Budgets, reforecasts, targets and ramp profiles flow through the same pipeline as actuals, so an agent knows what the plan was. |
| Strategy context | None. The strategy lives in decks and OKR tools the agent cannot see.Not covered. Governed context describes the data estate, not the objectives the business is pursuing.Not covered. The catalogue describes the data estate, not the objectives the business is pursuing. | Objectives, key results and funded initiatives are wired to the same tree, so an agent knows which strategic bet a metric serves and what work is targeting it. |
| Action Context | ||
| Acting on a change | None. The answer ends in the chat window.Out of scope. Agents can read governed context; acting on a metric move is not what the catalogue does.Out of scope. Agents can read governed context; acting on a metric move is not what the catalogue does. | A metric move pushes to the named owner with the driver attached, and workflows escalate up the reporting chain when nobody acts. |
| Agent write-back | An agent can run SQL, but there is nowhere to record an action, an owner or an outcome.Read-oriented. There is no task, ownership or outcome object for an agent to write to.Read-oriented. There is no task, ownership or outcome object for an agent to write to. | The context is writable through the same MCP: agents file tasks against metrics, update ownership and propose tree edits, with every change traced. |
| Learning over time | Stateless. Every session starts from zero and the org never learns from itself.No action or outcome history; nothing accumulates between sessions.No action or outcome history; nothing accumulates between sessions. | Every action declares its expected effect up front, so completed work doubles as a real-world experiment on a driver edge. What passes re-weights the model, every change is evidenced and reversible, and it is all remembered between sessions for every agent. |
Run your own workflows on this context
Because Canopy holds the org chart, the RACI and the metric triggers, you can automate business processes on top of it. Someone is out of office? The action escalates to their manager automatically. A metric goes silent, a key result drifts or a period closes off-target? A workflow fires, with approvals as the human gate. This is context that does things, not context that gets quoted.
Prefer us to run the agents? Meet Canopy Agents
Everything on this page is what external agents consume over MCP. When you would rather not bring your own, Canopy Agents are the agents KPI Tree runs for you: the Personalised Action Plan, RACI assignment, canvas and Slack agents, plus custom agents you deploy, all on this same context with your permissions, model choice, spend caps and full run history.
Weekly revenue review
scheduled · MondaysOpus 4.8Ask questions that were previously impossible
What is driving churn, who owns the fix, did last month's action actually work, and which objectives are at risk? No warehouse query answers those. Canopy's MCP carries the driver edges, RACI, verified impact, budgets and reforecasts, objectives and initiatives, and full dbt lineage behind every metric, and it is writable: agents file tasks, update ownership, propose tree edits and register driver hypotheses for outcomes to test, all traced. Every answer is scoped to the asking user's permissions. And because the context is defined once, connecting another agent costs nothing: one-click setup for Claude, ChatGPT, Gemini, Copilot and the other MCP clients, with no separate integration per agent.