AI Agents for Landscape Architecture Firms: Site Notes and Client Reviews
AI agents for landscape architecture firms that keep site notes, planting revisions and client review comments tied to the current drawing set.

An AI agent is useful to a landscape architecture firm when it keeps site visit notes, planting plan revisions and client review rounds tied to the drawing set they describe. Each comment should point to a sheet, a revision and a decision, so the team can answer which planting plan the client approved and why. The safe first step is a narrow agent that files and cross-checks this material, while licensed designers keep control of anything that changes the design.
Key takeaways
- Tie every site note and review comment to a sheet, a revision and the moment it was observed.
- Start with one role, such as a review log keeper, before giving an agent access to drawings or client email.
- Record plant substitutions, approvals and superseded items as explicit entries, never as text buried in a message thread.
- Treat that record as the memory the agent works from, so any later model draws on facts rather than guesses.
- Keep design decisions with a licensed designer, and let the agent draft, file and flag.
Why site notes lose their meaning without a drawing reference
A note such as 'move the bench a few steps west and check drainage near the old oak' is clear to the person who wrote it and ambiguous a few weeks later. Which base survey was on the table? Which grading plan was in force? Which planting plan placed the oak inside its protection zone? Landscape projects usually run several live versions at once: the set issued for permit, a working set marked up on site, and a client version carrying comments from the last review. A note written without a revision reference attaches itself to whichever version the reader happens to open.
The cost of that ambiguity shows up in the review round. A client asks to reverse a plant substitution that was already made, and the team cannot quickly confirm whether the change was discussed on site, approved in writing, or only implied in an email. Traceability lets a team answer that question in minutes instead of rebuilding the project from memory and inbox searches.
What the agent needs to hold as memory
Before an agent does anything useful here, it needs a defined memory. That memory is not a folder of attachments. It is a structured record that the agent reads at the start of each task and writes to at the end of it. For a landscape practice, the record usually includes:
- The drawing set index: sheet numbers, titles, revision letters and the purpose of each issue.
- Site visit notes, each linked to the sheet and revision that was on site at the time.
- Review rounds, with each client comment mapped to a sheet, a location on that sheet and an outcome.
- The approved plant list, substitution rules and any species flagged for availability or site conditions.
- Decisions and approvals, with the person who gave them and the revision they apply to.
The discipline is the same one that helps in other trades: a record should show what was known, against which version, and who decided. The OpenTelemetry observability primer describes the general habit of recording what a system did so its behavior can be inspected later. Applied to a planting plan, the equivalent is a log in which every change names its source. The same memory-first reasoning applies to crews and bids, as covered in landscaping routes, bids and crew notes.
A step-by-step plan for linking notes to the drawing set
Start with the drawing set rather than the notes. The sequence below works for a firm handling several projects at once, and it can be run by a project lead without any engineering work.
- Freeze an index of the current drawing set, with sheet numbers, titles and revision letters, and mark which revision is the issued one.
- Require every site note to name a sheet and a revision. If the observer is unsure, the note is filed as unanchored and flagged for review.
- Log each client review round as separate comments, each mapped to a sheet, a location on that sheet and a proposed change.
- Record the team's response to each comment as accepted, declined with a stated reason, or waiting for a site check.
- When a planting plan is revised, create a new revision entry and mark the replaced items as superseded. Never overwrite the earlier entry.
- Route any change to grading, drainage, structural elements or tree protection to a licensed designer for approval before the record marks it final.
Firms that manage technical submittals face a similar problem, which is why the approach resembles the RFI and submittal logs used by civil engineering teams: every item is tied to a specific document version.
What a traceable revision looks like in practice
The following scenario is illustrative and does not describe a real project. A designer visits a courtyard and notes that a specified shade tree sits close to a utility line shown on the planting sheet at revision C. The agent files the note against revision C, checks the approved palette for two smaller trees with a similar form, and drafts a substitution entry marked as pending. It does not change the planting plan or message the client.
The designer confirms the site condition and approves one of the substitutions. When the next review round arrives referencing the newer revision, the client's comment is mapped to that sheet, and the replaced tree remains in the log as superseded, with the site note and approval attached. Anyone reading the record can see what changed, when it was approved and by whom, without asking the team to reconstruct it.
How should access and approval boundaries be set?
Access should be scoped as tightly as the work allows. An agent that files site notes needs read access to the drawing index and write access to its own log. It does not need permission to edit design files or to send messages to clients. Connections to shared folders and mail can be granted through standard OAuth 2.0 scopes, which limit what a connected application may read or change.
Client emails are also inputs the agent reads, so they can contain text that tries to redirect its behavior. Prompt injection guidance recommends treating retrieved content as data rather than as instructions, and keeping tool permissions narrow whatever a message says. Approval should sit with people for anything that changes design intent, while the agent handles filing, anchoring and drafting.
Be Human designs, implements and operates this kind of setup for companies. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, such as whether each site note is anchored to a revision. Managed operations can add monitoring, recurring evaluations of filing accuracy, maintenance of connected folders and mail, periodic permission review, incident response when something goes wrong, and controlled changes to the agent's instructions.
Where a stronger model fits later
Retrieval is the practical way to give a model access to a firm's record without retraining it. A retrieval-augmented setup fetches the relevant notes and revisions at the time of a question and places them in the model's context, which keeps the source of each answer visible. The AWS comparison of retrieval and fine-tuning makes a related point: when knowledge changes often, as revisions do, a governed store is easier to maintain than knowledge baked into a model.
This is why the record matters more than the model. If the firm's notes are anchored to revisions today, a stronger model added later can draw on the same trail, and its answers can be checked against it. If the notes are not anchored, a better model only produces more fluent guesses. Evaluation should therefore test the record, not the prose: a recurring check of whether the agent cites the correct revision is a more useful test than whether its summaries read well.
If you want to find the first review task in your firm that is worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.
Frequently asked questions
Does the agent need access to our CAD software? Not for the first role. A review log keeper can work from exported PDFs or from a drawing set index that a designer maintains. Expanding access to design files is a later decision, made after the success criterion is met and a permission review is in place.
How is this different from keeping notes in a project management tool? A project tool stores what people wrote. The agent's job is to check each entry against a named revision, flag what has no anchor, and show which items were superseded. The tool can host that record, but the anchoring rules are what make it useful during a review round.
Will a stronger AI model make this unnecessary? No. A stronger model reads the record more accurately, but it still needs to know which revision a note refers to and who approved a change. The record is the durable asset, while the model can be replaced as better options become available.
Who approves changes that the agent drafts? A licensed designer or the responsible project lead approves any change to design intent, grading, drainage or protected trees. Routine filing actions, such as anchoring a note to the correct sheet, can proceed without approval if the firm decides that and the permissions allow it.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


