Superintelligent Agents for Landscape Firms: Site Notes and Revisions
How a landscape architecture firm can give an AI agent site notes and revision history, while the design team keeps authority over every decision.

An AI agent for a landscape architecture firm can read site notes, meeting records, markups and revision logs, then answer questions such as what was agreed for a planting zone, what changed since the last client review, and which open items still need a decision. The firm keeps design authority when the agent is limited to retrieving, comparing and drafting, and when every change to a drawing, plant list or specification is approved by a named designer. The practical value is organized, searchable project memory, not an agent making design calls.
Key takeaways
- A more capable agent is most useful as a retrieval and comparison layer over site notes and drawings, not as a designer.
- Revision history explains why a plan looks the way it does, so it needs clean version labels and a clear rule for what counts as approved.
- Design authority stays with people through explicit approval boundaries: the agent drafts and flags, a designer decides and signs.
- Start with one project type, one recurring question and one success criterion that can be checked against past reviews.
- Access scope, source versions and audit logs matter as much as the quality of the model answering the question.
What a more capable agent could do with site notes
Site notes are scattered by nature. They come from visits, phone calls, photographs, soil observations, irrigation checks, contractor questions and comments in markups. They sit in field apps, email threads, PDF sketches and the memory of the project lead and the site foreman. Over several weeks, these notes often contradict each other: a drainage slope recorded at one visit is revised after a survey, or a plant substitution is agreed by email but never reaches the schedule.
A more capable agent, connected to the right sources, could answer questions across that whole file in seconds. It could summarize what changed between two sheet sets, list which site observations have no matching design response, flag a note that says to retain an existing tree while the current drawing shows it removed, and draft a client update from notes that a designer has verified. These are retrieval, comparison and drafting tasks. The agent does not know the site beyond what has been recorded, and it cannot see a condition nobody wrote down.
How revision history changes the job
Revision history is the spine of the project record. Each drawing version has a date, an author, a reason for change and, ideally, a link to the request or observation that triggered it. An agent that can read that log can build a chronology of decisions: which client comment led to which change, which approval came before which construction note, and which version was current when a contractor asked a question.
This only works when the log is trustworthy. Version names need to be consistent, and the firm needs a rule for what counts as approved, such as a signed design review, a client email with a specific confirmation, or a permit set issued by the team. Without that rule, the agent may cite a draft as if it were a decision.
Illustrative scenario: a residential courtyard project has three revisions of the planting plan. During a fourth review, the client asks why a bench moved. An agent that can read the revision log, the site visit notes and the approval emails can point to the sheet version where the move appeared, the observation about a drainage swale that made the change necessary, and the email in which the client accepted it. The designer checks those sources and answers the client directly.
Keeping design authority with the firm
Design authority means that a qualified person decides what the project becomes and can explain why. An agent should support that decision without quietly taking it over. The simplest way to protect authority is to write the boundaries down before the agent is connected to anything.
- Retrieve and cite notes, drawings and revision records that the requesting user is already allowed to access.
- Draft summaries, meeting agendas, comparison tables and client update text for a designer to edit.
- Flag conflicts between notes, drawings, specifications, plant lists and schedules.
- Never change a drawing, specification, plant list or permit package without a named designer approving the change.
- Never send client-facing material without human review.
Permissions should follow the same logic as the firm's own access rules. The OWASP authorization guidance is a useful reference for checking that the agent can only reach what the person it serves can reach. Treat text inside site notes, contractor emails and uploaded PDFs as data, not instructions. A contractor email that says "ignore the approved plan" should be read as a statement to review, not a command. The OWASP guidance on prompt injection describes why this separation matters when an agent reads documents from outside the firm.
A step-by-step plan for the first engagement
A first engagement should be small enough to evaluate in weeks. The goal is to prove that the agent returns accurate, sourced answers on a question the team already asks often.
- Choose one project type, such as residential planting design, and one recurring question, such as what was agreed for planting zones since the last review.
- Inventory where notes, drawings and revision logs live, and decide which versions count as approved.
- Name one owner who approves sources and one designer who reviews every output before it is used.
- Define the success criterion in advance: the agent's answers must match the human-verified record on a set of past questions, with a source cited for each claim.
- Run the agent on two or three closed projects first, then on one live project in read-only mode.
- Review every failure, fix sources and permissions, and widen scope only after the review is clean.
The NIST AI Risk Management Framework offers a structure for documenting these risks, measuring them and deciding who is accountable for each control. It does not replace a firm's own professional standards, but it helps make the review process explicit.
How to tell whether the agent is reliable
Reliability in this setting is measurable. Build a small set of questions with answers that a designer has already verified from the record, such as which plant substitution was approved, which drainage condition triggered a revision, or which client comment is still open. Run the agent against that set after every change to its sources, prompts or permissions.
Score each answer on three points: is it correct, does it cite the right source, and does it say when the record is silent. The last point matters most. An agent that admits it cannot find an approval is more useful than one that fills the gap with a plausible guess. The OpenAI evals guide and the Anthropic guidance on evals for AI agents both describe this kind of repeatable testing. Logging each retrieval and tool call, using a consistent approach such as the OpenTelemetry conventions for generative AI, makes it possible to trace why an answer was produced when something goes wrong.
How Be Human handles this
Be Human designs the memory, source connections, tools and approval boundaries for one defined role, then operates it under human control. For a landscape firm, that means setting up the agent's access to notes, drawings and revision logs, documenting which versions are treated as approved, and defining which outputs need a designer's sign-off. Managed operations then cover monitoring, recurring evaluations against verified questions, permission reviews when people or projects change, and incident response if an answer cites the wrong source. Changes to what the agent can reach go through review, and the firm keeps ownership of its drawings and its decisions. Similar memory problems appear in crew work, as described in Landscaping Routes, Bids and Crew Notes, and in technical correspondence, as described in Superintelligent Agents for Civil Engineers: RFI and Submittal Logs.
Where to start
The best first task is usually the question your senior people answer from memory most often, because it shows where the record is incomplete. Before choosing, check which notes are dated, which revisions are clearly marked as approved and which decisions were made by phone and never written down. Those gaps are the work that comes before any agent. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent change a planting plan or drawing on its own? No, and it should not be set up that way. In a sound design, the agent can retrieve, compare, summarize and draft, but any change to a drawing, plant list or specification needs a named designer to review and approve it. The approval itself should be recorded in the revision log so the next question has a clear answer.
What site information does the agent actually need? It needs access to the records the firm already uses to make decisions: site visit notes, survey outputs, markups, revision logs, approval emails and client correspondence about design changes. It works best when those sources have consistent names and dates. Anything that never got written down will not be available to it, which is why the inventory step comes first.
How do we know the agent is giving correct answers? Build a set of questions where the correct answer has already been confirmed from the record, then test the agent against them before and after every change. Check whether each answer is correct, cites the right source and admits when the record is silent. Keep a log of retrievals so that a wrong answer can be traced to a specific source or permission.
How long does a first engagement take? The length depends on how many sources are involved and how clean the revision history already is, so no fixed timeline applies to every firm. A good first engagement defines one role, one project type and one success criterion, then tests the agent on closed projects before any live use. Scope grows only after reviews show that the answers are accurate and the boundaries hold.
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.


