All use cases
Phones & Booking

A landscaper’s AI call was marked successful even though it failed. Reading the transcript caught it

Audit real calls and add a human fallback when the caller and the bot cannot understand each other.

Real storyDone with Roth Digital AI voice agent

What they did

Roth Digital AI published a transcript-based review of SouthCal Landscape’s phone agent covering June 22 through September 21, 2026. It separated spam, owner tests and genuine customers rather than counting every answered call as a win.

How it went

Of 121 total calls, the agency identified 23 genuine customer calls: 16 handled by AI, six correctly escalated and one mishandled. On the failed call, poor audio caused repeated clarification for over two minutes. The platform still labeled it successful. The fix was a fallback after about two attempts without useful new information: offer a callback from a person.

Worth knowing

It is a small sample, and the agency that built the agent performed the audit. Answered is not the same as booked, and a positive sentiment score is not proof the customer got help. This is a supplier-published case study, not an independent audit. The prompt below is our adaptation, not the company’s original prompt. A chat draft does not reproduce an integrated software system. Wisconsin angle: Truck cabs, wind and noisy job sites make this especially relevant to field-service calls. Give customers an easy way to reach a person and verify callback numbers.

Try it yourself

Act as the office helper for my Wisconsin landscaping business. Help me test one small workflow before buying or connecting software. TASK: Audit real calls and add a human fallback when the caller and the bot cannot understand each other. MY INPUTS (replace brackets; remove customer and employee identifiers): - [ANONYMIZED CALL TRANSCRIPTS]: paste 10 real examples - [PLATFORM SUCCESS LABELS]: if available - [APPROVED HUMAN HANDOFF]: person, hours, callback wording - [REQUIRED INTAKE FIELDS]: job type, town, callback number DO THIS: 1. Classify calls as customer, spam, test or other; show the evidence. 2. For customer calls mark resolved, correct handoff, failed or unclear. Do not equate friendly wording with success. 3. Flag repeated questions, unconfirmed phone numbers and promises without evidence. 4. Write a two-attempt clarification fallback and a human-request path. 5. Draft a weekly audit sheet separating genuine customers from test/spam calls. RULES: Use only my supplied information. Write NEEDED for missing information and cite the input row, document page, or note behind each finding. Separate observed facts from assumptions. Do not invent prices, measurements, certifications, warranties, availability, or customer approvals. Do not send messages, change records, book work, charge cards, or contact anyone. Give me drafts and a review checklist only. Stop and flag safety concerns for a qualified person; never diagnose equipment or tell a customer to attempt hazardous work. Keep personal details out of shared outputs. OUTPUT: 1. A short plain-English summary. 2. The requested working table and message drafts. 3. Missing facts and exceptions requiring an owner's decision. 4. A five-item check before using this on a real job. 5. A simple two-week test log: date, task, manual minutes, AI-assisted minutes including checking, errors caught, and outcome. Do not claim savings until I fill it in. Then ask me for the single most important missing input. Keep the language natural, like a helpful local shop owner, not a sales brochure.

Want help getting this working? Rise Systems in Oconomowoc sets these up.

Want this running in your business without the DIY? Rise Systems in Oconomowoc sets these up.

Free 15-minute AI check-up
Read the original

Source: Roth Digital AI case study · Publication date not stated; checked November 7, 2026

Related