Lesson objective

Explain why an AI tool can improve one landscape office task and degrade another, and require separate evidence for each task before it is scaled.

Plain-language explanation

Why this matters now

Once a tool works well on one job in the office, the next suggestion always arrives: use it for this too. The renewal letters came out well, so try it on the spring treatment plan. The proposal drafts saved hours, so let it read the weed photos. Each extension feels like a small step, because it is the same tool, the same login and the same person asking.

The evidence says that is exactly where operations go wrong. AI performance is not a property of the tool. It is a property of the tool on a particular task, and the boundary between the tasks it helps and the tasks it harms is not visible from the outside.

Research (not binding): Dell'Acqua and colleagues, "Navigating the Jagged Technological Frontier", published in Organization Science (vol. 37, issue 2, March 2026, earlier circulated as a Harvard Business School working paper in September 2023), ran a preregistered experiment with 758 consultants at Boston Consulting Group, randomly assigned to work without AI, with GPT-4, or with GPT-4 plus a prompting overview. On 18 tasks inside what the authors call the AI capability frontier, AI users completed 12.2% more tasks, finished them 25.1% more quickly, and produced significantly better quality. On a task selected to lie outside the frontier, subjects using AI "were 19% less likely to produce correct solutions compared with those without AI."

Those were management consultants using a 2023 model, so the result applies to landscaping only by analogy. The analogy is the lesson: the same people with the same tool got better on some tasks and worse on another, and they could not see the line.

A question to sit with

Your office has used AI for a year to draft renewal letters and customer notices, and the results have been good. Someone proposes using the same tool to identify weeds from crew photos so the spray route can be planned faster. What evidence do you have about the second task?

What goes wrong

Borrowed confidence. Success on drafting is treated as evidence about identification, estimating or diagnosis. It is not. Each task has its own frontier, and a good record on one says nothing about the next.

The frontier is invisible from inside the output. An answer outside the tool's competence looks the same as one inside it: fluent, confident and well formatted. Professional guidance (voluntary framework): NIST's Generative AI Profile (AI 600-1, July 2024) defines confabulation as "the production of confidently stated but erroneous or false content," says confabulations "are a natural result of the way generative models are designed," and notes that models "sometimes provide logical steps for how they arrived at an answer even when the answer itself is incorrect." A reasoned-looking weed identification is not evidence that the identification is right.

Transferring someone else's accuracy figure. A published accuracy rate describes one tool on one set of material in one place and time. Research (not binding): Wanigasinghe and colleagues (AoB Plants, May 2026) tested two plant identification apps on 5,291 field images of 119 wild vascular plant species in small plots in Alberta, Canada. Flora Incognita identified 79.2% of species correctly and iNaturalist 66.5%, and neither reached the authors' estimated expert accuracy range of 90-95%. That study measured wild Alberta flora. It says nothing about turfgrass, cultivated ornamentals, nursery stock, tree diseases or tree pests, and its percentages must not be carried over to any of them or used to choose an app. What does transfer is the shape of the finding: even in a careful study, the apps fell short of the authors' expert benchmark, and the shortfall had to be measured, not assumed.

Averaging hides the bad task. A single "AI accuracy" figure for the whole office blends the tasks the tool does well with the one it does badly. A 97% overall correction-free rate can contain a task that is wrong a quarter of the time.

The practical answer

Treat every task as needing its own evidence before it scales. The approach below is an AI Coalition Network recommendation, not a legal or professional requirement.

Build a task-by-task evidence table. List each AI-assisted task separately, never "the tool", and record what you actually know about it.

Task Evidence in hand Likely position What must happen before scaling
Drafting renewal letters from the verified contract Own sample: 1 correction in 50 Probably inside Keep the weekly sample; scale in stages
Summarising crew notes into visit summaries Own sample: 4 in 160 Probably inside, with care Source-first review; sample unflagged notes
Estimating mulch and plant quantities from a takeoff No sample yet Unknown Parallel run against the estimator's own takeoff
Identifying a weed from a crew photo Only published studies on other plants Unknown, treat as outside Confirmation by a qualified person every time; Module 8 sampling
Judging a tree's condition from a photo None Outside Not an AI task; the arborist or tree expert decides on site
Reading what a pesticide label allows None; the label is the source Outside as a decision AI may help find a section; the applicator reads the label

Measure each task on your own work. Run the tool beside the person who does the task today, on the same inputs, for long enough to cover the variety your routes produce. Count agreements and disagreements, and have the qualified person decide which was right. Record the rate for that task alone.

Watch the season. A task measured in May may sit differently in October. Spring weeds, summer stress, autumn leaf cover and the plants on a new property all change what the tool sees. Re-measure when the inputs change.

Mark the tasks that are never the tool's decision. Some tasks are outside the frontier by definition, because a named qualified person owns them. The AI may help that person find a page or organise a note. It does not select a pesticide or a rate, decide that a tree is safe, or design an irrigation system.

Say what the evidence is, not what the vendor says. Level 2 taught you to question an accuracy claim: what was measured, on what, by whom. At Level 3 the rule is organisational. No task enters an operation on a vendor's figure or a published study of different material. It enters on your own measured result, or it enters as a task that a qualified person confirms every time.

Remember this

  • AI performance belongs to a task, not a tool; success on one task says nothing about the next.
  • Outside the frontier, the published study found AI users were "19% less likely" to produce correct solutions, and the output did not reveal it.
  • A published accuracy figure describes its own material; the Alberta wild-flora figures do not transfer to turf, ornamentals or tree problems.
  • Measure each task on your own work, and re-measure when the season or the inputs change.
  • Tasks that feed a qualified decision are confirmed by the qualified person every time.

Landscaping example

Kestrel Tree & Shrub LLC is a fictional company that has used an AI tool for a year to draft renewal letters and seasonal service notices. The office manager, Anselm Okoro, has a record: in the last sampled 200 letters, 3 needed correction, 1.5%. In June the spray programme manager, Rosa Lindqvist, a certified applicator, is asked by the owner whether the same tool could identify weeds from crew photos so that the broadleaf route could be planned from the office. The owner points to the letter results as proof the tool is "accurate". Rosa declines to use that figure. She sets up a parallel run instead: for three weeks, crews photograph each weed patch they flag, the tool suggests an identification, and Rosa or a second certified applicator identifies the same patch from the photo and, for a sample, on site. Of 120 patches, Rosa and the tool agree on 87 and disagree on 33, an agreement rate of 87 ÷ 120 = 72.5%. More importantly, 14 of the 33 disagreements involve grassy weeds in turf, and 6 involve a plant Rosa has not seen on these routes before. She records three things: the tool's agreement rate on this task is not its letter-drafting rate; it is weakest on exactly the cases where a wrong call would drive a wrong product onto a route; and the figure comes from Kestrel's own routes in June, not from any published study. The owner agrees the tool may help crews file photos, but every identification that could lead to a treatment decision is confirmed by a certified applicator.

Responsible-use warning

Do not let a good record on one task stand as evidence for another, and do not import an accuracy figure from a vendor sheet or a published study of different plants, places or seasons. The Alberta study measured wild flora in small plots, and its figures say nothing about turf, ornamentals, tree diseases or tree pests. The tasks most likely to fall outside the frontier are the ones that feed qualified decisions, so they are where borrowed confidence costs the most.

Human-review requirement

The owner of each task decides whether its evidence is sufficient: the office or account manager for drafting tasks, the estimator for takeoffs, the certified applicator for anything feeding an application decision, and the arborist or licensed tree expert for tree condition. The operations manager keeps the task-by-task evidence table, refuses to scale any task without its own row and owner, and schedules re-measurement when the season, the tool or the property mix changes. No summary figure for "the tool" may substitute for a task's own evidence.

Practical exercise

Use fictional material only. No real customer, property, product label or photograph.

  1. List every AI-assisted task your company runs or is considering. Write each as a separate row; never write "the AI tool" as one row.
  2. For each task, record the evidence you actually hold: your own sample (with counts and dates), a vendor claim, a published study of other material, or nothing.
  3. Classify each task as probably inside, unknown, or outside the frontier, and give one sentence of reasoning for each classification.
  4. Mark every task that feeds a decision owned by a qualified person: the certified applicator, the licensed irrigator where your state licenses the work, the arborist or tree expert, the estimator, or the contract signer.
  5. For one "unknown" task, design a parallel run: which inputs, how many, over how many weeks, who judges disagreements, and how you will count agreement.
  6. Recompute this fictional result: 87 agreements in 120 identifications. State the rate and write one sentence on why it may not hold in October.
  7. Write the rule your company will use to decide when a task's evidence is enough to scale, and what happens to a task with no evidence.
  8. Name who owns the evidence table and how often each row is re-checked.

Knowledge check

  1. What did the Dell'Acqua experiment find inside and outside the AI capability frontier?

    Answer guidance: Inside the frontier, consultants using AI completed 12.2% more tasks, 25.1% more quickly, with significantly better quality. On a task outside it, AI users were 19% less likely to produce correct solutions than those without AI.

  2. Can the Alberta plant identification accuracy figures be applied to weed identification in turf?

    Answer guidance: No. The study measured wild vascular plants in small plots in Alberta; its figures do not transfer to turf, ornamentals, nursery stock or tree problems, and must not be used to choose an app.

  3. Why is a single accuracy figure for "the tool" misleading?

    Answer guidance: Performance belongs to a task. An overall figure blends tasks the tool does well with tasks it does badly, so a strong average can hide one task that is often wrong.

  4. How should a company decide whether a task is ready to scale?

    Answer guidance: By its own measured result on that task, from a parallel run on the company's own work judged by the qualified person, or by treating the task as one the qualified person confirms every time.

Sources and further reading