What it does

Thirty-four things it does, and what stops each one going wrong.

Everything the AI does in Canopy and Trellis: what it is, where you see it, what sets it off, and the specific thing that stops it doing damage when it is wrong. The last column is the one worth reading.

The limits are the product.

Anyone can produce a suggestion. The question you actually have is what happens when it is wrong, and how many times before somebody notices. So each one is listed with the thing that bounds it: a person who has to approve, a practice run first, a dollar ceiling, a requirement to show its evidence, or a rule it is simply not allowed to argue with.

Notice how many of them end with a person. That is deliberate, and most of them stay that way for a full season.

What it does	Kind	Where you see it	What sets it off	What stops it going wrong
Schedule optimization	Math	Dispatch board	On demand, nightly	Publishing requires a person. Dispatcher pins are hard constraints.
Rain-day re-solve	Math	Dispatch, Teams	Forecast threshold	It avoids churning the week, and shows you what moves before you publish.
Snow event call-out	Math	Dispatch, SMS	Accumulation forecast	The trigger depths in the contract are not negotiable.
Seasonal crew loading	Math	Planning	Quarterly	Scenario only. Never publishes to the field.
Equipment service queue	Math	Fleet	Meter threshold	Respects next week's route load.
Mow-cycle skip advisor	Math	Dispatch	Degree-day model	Contract minimum visits are never violated.
Account brief	Writing	Property 360, Teams	Before a scheduled visit	Every claim cites a record id.
Proposal narrative	Writing	Estimate builder	On proposal build	It can only describe work that is actually priced on the estimate.
Renewal package	Writing	Property 360	120 days before expiry	Escalation comes from the contract, not from the model.
Customer visit update	Writing	Portal, email	On ticket completion	The account manager approves the first twenty per account.
Review and scorecard narrative	Writing	Console, Teams	Monthly	Numbers come from your books, never from the model.
Incident and claim write-up	Writing	Mobile, Safety	On incident creation	Stays in draft until a manager signs.
Enhancement suggestion	Memory	Everywhere	Signal arrival	Evidence required, and it will not suggest the same thing twice in sixty days.
Renewal risk scoring	Memory	Console, Property 360	Nightly	Contributing factors always shown.
Why a job is over budget	Memory	Job cost	Three cycles over budget	Ranked causes, never a single verdict.
Property history questions	Memory	Anywhere you can ask	You ask	Answers only from your own records, never another company's.
Ask for any number	Memory	Console	You ask	It shows exactly how the number was worked out.
Whether to bid at all	Memory	Estimating	Request received	Recommends with reasons. A person decides.
Agronomic program design	Memory	Estimating, Operations	Seasonal	Label rates are hard limits, never suggestions.
Aerial measurement	Photos	Estimate builder	On property creation	The estimator confirms every area it measures.
Photo check on close-out	Photos	Mobile	On visit close-out	Advisory. Never blocks a crew from leaving.
Supplier ticket extraction	Photos	Mobile, Purchasing	On photo upload	Anything it is unsure of stops and waits for a person.
Plant and turf health	Photos	Mobile	On photo upload	Suggests, never diagnoses. Routes to an agronomist.
Multilingual voice capture	Language	Mobile	Push to talk	The original recording is kept, and the confirmation shows in the speaker's language.
Correspondence translation	Language	Portal, email, SMS	On send	Your approved wording is used, and longer messages get checked back the other way.
Safety and training localization	Language	Mobile, Safety	On assignment	Acknowledgement recorded against the language version.
Legal text translation	Language	Contracts, notices	Once per clause	The model drafts, a person approves, the version is pinned.
Billing run	Chores	Finance console	Cycle close	Dry run, dollar thresholds, reversible writes.
Collections sequencing	Chores	Finance console	Daily	Drafts only. A person sends.
Three-way match	Chores	Purchasing	Continuous	Small differences pass. Anything outside your tolerance comes to a person.
Payroll pre-close	Chores	Finance console	Per pay period	Never auto-edits time. The approver is the crew's manager.
Thirteen-week cash forecast	Chores	Finance console	Weekly	Advisory. No payment is ever initiated.
Claim defense package	Chores	Finance, Safety	Notice of loss	Assembles only. Never communicates with a claimant.
Regulatory filing	Chores	Compliance	Per state calendar	The applicator of record signs every submission.

How much rope each one gets

Each of the 34 sits at one of five levels, and you set the level per branch and per company rather than once for the whole product. Turning one up is a decision you make once it has earned it. Turning it back down is one click and does not involve us.

Whatever the level, an action always shows what would change, what it is worth in dollars, and how to undo it. Approvals come to you where you already are.

__SPECTRUM__

Most agents live at L1 or L2 for a full season. A few never leave L0, because the number they were built to prove turned out to be small, and that is a useful thing to have learned cheaply.

Evaluation is a build artifact

Golden specifications per capability, run on every prompt or model change: a fixed input set with expected structured output. A capability whose evaluation is failing does not reach a customer, and the evaluations ship as something you can look at rather than something we promise we do.

Everything is traced

Every prompt is versioned and scoped, every run is traceable and explainable, with cost and latency on each. When an agent gets something wrong, the question is not whether you can reconstruct why. It is only how long it takes.