Practical LLM work by non-engineers

Research checked on 2026-10-15, Asia/Tokyo.

The strongest cases use real work material: event listings, consultation speech, farm records, testimony, repertoire lists, and the operator's own drafts. The assistant performs a defined task. Several accounts explicitly describe checking the result before acting.

This note retains 25 practices across 21 named operators. Several practices belong to the same person. These are documented accounts of work performed, not independently audited business outcomes. Dates describe the evidence. They do not establish that every operator still uses the same setup today.

How to read the evidence

• Practitioner account: The person describes their own work. Examples, screenshots, or process details strengthen the account.
• Direct interview: A named practitioner describes actual work to a journalist or professional association.
• Vendor account: The software supplier publishes the customer story. The commercial relationship matters even when the workflow is specific.
• Repeat value: My explanation of why the task can keep helping. Documented frequency is stated when available. This is not a measured return on investment.

The focus is the person's non-engineering work. Headcount is not established for every business. A family operation, a small law firm, and an individual working in a larger business are not interchangeable categories.

Operations and records

1. Check event listings before customers see the wrong time

Operator: Larissa Guetter, ATV Big Air Tour. Two founders manage the business, with a larger touring crew.

• Pain: Promoters and community sites publish conflicting dates and showtimes across a 26-event tour.
• Actual use: A scheduled morning ChatGPT briefing flags inconsistencies and supplies links. It found an 8 p.m. listing for a 7 p.m. show. She checks the sources and requests corrections.
• Repeat value: New events and changed listings create fresh checking work every week.

Evidence: Detailed owner account and a public demonstration. Her company also participated in an OpenAI promotional film. This is not an independent endorsement. Owner's workflow, 2026-08-11, updated 2026-08-16. Vendor-hosted demonstration, 2026-08-12, especially 08:05 onward. Commercial relationship disclosed, 2026-09-13.

2. Turn farm data into mandatory water reports

Operator: Rachael Sharp, Sharp family farm, South Carolina.

• Pain: Reporting water use for wells and surface-water pools took substantial administrative work.
• Actual use: Uploads the blank form and farm data to ChatGPT. Reviews the populated form and corrects information placed in the wrong fields.
• Repeat value: Reporting requirements recur. The workflow converts existing records into required paperwork.

Evidence: Direct farmer interview. Reported time savings are not independently measured. This supports document preparation, not autonomous regulatory compliance. Farm Progress interview, 2026-06-26.

3. Build tools for shifts and daily bakery reporting

Operator: Masahiko Miyashita, Bakery Heidi, Japan.

• Pain: Scheduling and reporting add administration to the owner's baking work.
• Actual use: Built apps with Claude through instructions, screenshots, testing, and corrections. Staff enter sales on an iPad. LINE distributes reports. Google Sheets collects attendance and weather data.
• Repeat value: Staff reuse the tools for daily entries and monthly reporting. The owner does not need a new LLM conversation for each transaction.

Evidence: Named owner interview with app screenshots. He describes building without writing code. The English translation corrupts subscription prices. The Japanese original gives approximately ¥10,000 per service and ¥20,000 total. Those prices are not needed to establish the workflow. Japanese original, 2026-08-20.

4. Read incoming farm documents and ask questions about them

Operator: Travis Senter II, Arkansas farmer.

• Pain: Broker emails, grain-merchant messages, and farm-program documents require repeated reading.
• Actual use: Uploads them to ChatGPT for summaries and questions. Also supplies historical field information.
• Repeat value: He reports almost daily use. Each new message or programme creates another reading task.

Evidence: Direct interview. His separate yield-prediction experiments had mixed results. They are not treated here as proven forecasting. DTN interview, 2024-07-03.

5. Create a consistent inventory naming system

Operator: Richard Lane, manager at Reno Salvage.

• Pain: About 1,000 metal items needed recognizable, consistent part numbers.
• Actual use: Gave ChatGPT rules based on item names, thickness, and width. It produced an Excel list of part numbers.
• Repeat value: A reusable naming convention helps subsequent stock handling. This is continuing value from an initial setup, not proof of daily LLM use.

Evidence: Vendor customer story with an example prompt and output. Headcount and inventory-error reductions are unverified. Exaggerated secondary claims about eliminating all downtime were not retained. OpenAI customer story, 2026-02-04.

Client service and professional preparation

6. Check deposition objectives while the witness is still available

Operator: Brooks Schuelke, Austin personal-injury lawyer.

• Pain: He sometimes believed he had secured an admission, then found ambiguity in the transcript later.
• Actual use: Used Skribe in three depositions. During breaks, asked its OpenAI-based agent to compare live testimony with his objectives, identify evasions, and suggest follow-up questions.
• Repeat value: He can address a gap before questioning ends. He reports that this helped more than once.

Evidence: First-person repeated-use account. The product's official account also comments on the post. Case outcomes were not independently measured. Practitioner's post, accessed 2026-10-15. Texas Bar directory lists 2–5 lawyers, accessed 2026-10-15.

7. Find potentially missing documents in a discovery production

Operator: Brooks Schuelke.

• Pain: Reviewing whether delivered documents satisfy a request requires cross-referencing large collections.
• Actual use: Supplied requests and received documents. Asked a document-chat AI to compare supplied files with documents referenced inside them. It listed possible omissions and drafted a follow-up letter.
• Repeat value: Every production creates another completeness review.

Evidence: First-person account of an actual review. The product and underlying model are not disclosed. LLM use is inferred from the described prompting and outputs. Possible omissions are leads for lawyer review. Practitioner's discovery review, accessed 2026-10-15.

8. Convert spoken appointments into physiotherapy notes

Operator: Gil Jelin, solo physiotherapist at Wildwood Holistic Physio, Queensland.

• Pain: Hands-on treatment left notes unfinished. Writing in her second language added effort after appointments.
• Actual use: Heidi captures consultation information and drafts notes in her SOAP template. She reviews and files each note.
• Repeat value: Every appointment requires documentation. The draft reduces the writing required after each appointment.

Evidence: Vendor interview with named practitioner and demonstration clips. Her own clinic confirms the small home-based business. After-hours savings remain a testimonial. Claims of perfect notes were rejected. Heidi interview, 2025-09-16. Her clinic, accessed 2026-10-15. Heidi's explanation of its LLM workflow, 2026-08-17.

9. Turn job requirements into candidate searches

Operator: Jasmine Cheng, founder of boutique recruiter TopKnack.

• Pain: Company lists and Boolean searches took time to assemble for each vacancy.
• Actual use: Gives ChatGPT industry, location, and company-size criteria. Converts job descriptions into search strings, then checks results and adjusts criteria manually.
• Repeat value: She reports daily use. Each vacancy needs another sourcing plan.

Evidence: First-person edited interview. Her time estimates are unmeasured. She did not claim an isolated revenue increase. Business Insider interview, accessible Spanish edition, 2023-04-21.

10. Rehearse specialist interpreting assignments

Operator: Nora Díaz, Spanish–English interpreter and translator.

• Pain: Unfamiliar topics require vocabulary preparation before an assignment.
• Actual use: Before a spinal-surgery consultation, requested bilingual terminology, then checked and extended it through her own research. Uses simulated dialogues for rehearsal.
• Repeat value: A different subject creates another preparation task. Her translation work also uses alternative phrasings as options.

Evidence: First-person professional-association article with a concrete assignment. She also teaches AI workshops, which is a commercial interest. No productivity measurement is supplied. AUSIT account, accessed 2026-10-15.

11. Organize options for disrupted travel bookings

Operator: Marie Rowe, independent travel advisor.

• Pain: Disruption decisions require comparing booking details with multiple terms and rules.
• Actual use: Put regulations, company terms, and ATOL material into a custom GPT. During Middle East disruption, entered booking scenarios for a structured review of options. Also checks itineraries for local charges.
• Repeat value: New bookings repeat the comparison work.

Evidence: Direct trade interview and a verifiable independent travel business. This establishes use, not legal accuracy or current-policy coverage. TTG interview, 2026-06-24. Her business, accessed 2026-10-15.

Research and creative preparation

12. Prepare a weekly newsletter from a large reading list

Operator: Shakeel Hashim, independent journalist at the source date.

• Pain: A weekly newsletter required processing more than 150 articles.
• Actual use: Claude and o1 helped build an article-export tool and summarizer despite his limited coding skills. He checks facts and rewrites summaries.
• Repeat value: Each issue brings another batch.

13. Get another editorial review when working alone

Operator: Shakeel Hashim.

• Pain: No colleague was available for an editing pass.
• Actual use: Supplies drafts to a Claude Project containing prior articles and style instructions. Selects useful edits and criticism.
• Repeat value: Repeats for new articles.

Evidence for 12–13: First-person account with examples and a screenshot. Current Transformer is a larger newsroom, so this is a historical solo workflow. Workflow, 2024-10-07. Current organisation, accessed 2026-10-15.

14. Prepare podcast material without another long transcription session

Operator: Kelsi Bailey, working family and newborn photographer.

• Pain: Podcast transcription and preparation consumed hours.
• Actual use: Uploads episodes to ChatGPT for transcription, show notes, and publishing material.
• Repeat value: New episodes repeat the task.

15. Describe the gap between reference photographs and her own work

Operator: Kelsi Bailey.

• Pain: She struggled to articulate qualities she admired in reference images.
• Actual use: Supplies references and her photographs. Asks what differs and what she appears drawn to.
• Repeat value: Supports further creative development. Client or revenue improvement is unmeasured.

Evidence for 14–15: First-person account. Her client-photography business predates the AI article. Workflow, 2026-04-13. Professional identity, accessed 2026-10-15.

16. Translate a tattoo client's visual brief into image prompts

Operator: Ruby White, tattoo artist, Gatineau.

• Pain: Complex designs require substantial unpaid preparation.
• Actual use: Gives ChatGPT reference images, pose, colours, and shape. Uses its prompt with Midjourney, then selects and assembles elements in Photoshop and Procreate.
• Repeat value: She says she also uses this process for flash designs. New commissions need new preparation.

Evidence: Owner account with two process examples. Captions were checked, not the embedded videos. This combines an LLM, image generation, and manual design. Total savings cannot be attributed to the LLM alone. The page's unsupported copyright percentage claim was excluded. Artist's workflow, 2024-06-29.

17. Build a concert programme from repertoire he already performs

Operator: Ghenadie Rotari, concert accordionist and teacher.

• Pain: A replacement festival appearance needed a programme matching a youth theme.
• Actual use: Uploaded his repertoire to an AI assistant and requested thematic connections. Chose suitable material himself. The product is unspecified for this task.
• Repeat value: Reusing the repertoire for future briefs is a reasonable inference. The source documents this concert task.

18. Review the musician's own social results

Operator: Ghenadie Rotari.

• Pain: High-effort posts sometimes performed poorly.
• Actual use: Uploads weekly or monthly statistics to a ChatGPT Project and adjusts content by platform.
• Repeat value: New results inform subsequent posts.

Evidence for 17–18: First-person podcast. No measured improvement is established. Episode, 2026-07-08. Repertoire account, accessed 2026-10-15. Performer biography, accessed 2026-10-15. The episode's full indexed transcript was available, but direct opening was unreliable.

19. Get candidate readings of artists' signatures

Operator: Walter O'Neill, Antiques Arena.

• Pain: Dyslexia makes stylised handwriting slow to read.
• Actual use: Photographs signatures for possible readings. Checks proposed artists and comparable objects himself.
• Repeat value: He reports weekly use when handling new stock. He also uses structured listing templates.

Evidence: Owner-written account. The surrounding article strongly promotes his academy and includes affiliate links. No attribution accuracy or sales lift is established. Dealer's account, 2026-04-13.

Marketing and written work

20. Reuse the farmer's own writing for customer communications

Operator: Leslie Svacina, Cylon Rolling Acres, Wisconsin.

• Pain: Direct meat sales require regular communication alongside farm work.
• Actual use: Edits her drafts with ChatGPT and repurposes website material into captions and newsletters.
• Repeat value: Existing articles support later messages. She retains the farm knowledge and voice.

Evidence: Owner account with a demonstration. She also sells farm education and uses affiliate links. Generic copy and factual errors are discussed explicitly. Farm marketing workflow, 2025-05-08. Farm identity, accessed 2026-10-15.

21. Write flower descriptions and preserve shop procedures

Operator: Cameron Pappas, Norton's Florist, Birmingham.

• Pain: Descriptions were tedious. A delivery accident also exposed missing reporting instructions.
• Actual use: Supplies arrangement details and a short length limit to ChatGPT, then edits. Used it to draft accident-reporting and warehouse procedures.
• Repeat value: New arrangements need copy. Completed procedures remain available for future staff and incidents.

Evidence: Trade-association coverage of named participants' work. Headcount and procedural safety are unverified. Society of American Florists, 2023-08-16.

22. Turn product-review notes into sentences and soften difficult emails

Operator: Lily Canter, freelance journalist.

• Pain: Rough review notes need prose. Blunt emails can need diplomacy.
• Actual use: Gives ChatGPT her own notes and audience, then edits the draft. Requests email tone changes and removes excessive enthusiasm.
• Repeat value: She reports use several times weekly across writing and correspondence.

Evidence: First-person examples of actual text. She distrusts LLM research, so this is not evidence of AI reporting for her. Journalist's workflow, 2024-11-06.

23. Polish English reports for mushroom-growing clients

Operator: Jeroen van Lier, Total Mushroom Service, Netherlands.

• Pain: Writing clear technical English takes extra effort for a Dutch-speaking consultant.
• Actual use: Uses ChatGPT regularly to improve reports and articles. Treats technical answers as leads for further investigation.
• Repeat value: Each client report repeats the language-editing task.

Evidence: Signed practitioner article. No before-and-after report or measured savings. Do not infer improved crop yields. Practitioner's account, 2026-01-22.

24. Draft scene descriptions for serial fiction

Operator: Jennifer Lepp, who publishes as Leanne Leeds.

• Pain: Routine description under book deadlines.
• Actual use: Selects and edits Sudowrite descriptions from supplied context.
• Repeat value: Recurs across series books.

Evidence: Historical observation. Writing workflow, 2022-07-20. A later interview confirms paid Sudowrite writing. Follow-up, 2022-12-24.

25. Prepare product posts in a shoemaker's workshop

Operator: Olamilekan Oke, Feeted, Lagos. Psychology graduate and shoemaker.

• Pain: A small workshop could not afford the promotional resources of larger brands.
• Actual use: A reporter watched him photograph new shoes, use ChatGPT to generate a model image, and request captions for the intended post. He also uses it for product names.
• Repeat value: Each new design needs promotion. This combines LLM writing with image generation.

Evidence: On-site reporting with process images. No demonstrated sales increase or image fidelity test. Ripples Nigeria report, 2025-10-05.

What the validation changed

The research used three background investigations covering physical businesses, professional services, and creative work. A further pass checked evidence strength, work frequency, occupation, company identity, and the difference between a real action and an intended experiment. The coordinating pass looked for contradictory detail and commercial interests.

Concrete corrections and exclusions follow.

Additional search angles and stopping point

The search went beyond small-business success stories. It included occupational associations, trade publications, owner blogs, professional mailing lists, direct interviews, recorded demonstrations, and podcast transcripts.

Additional occupation searches covered plumbers, carpenters, repair shops, cleaners, dog groomers, funeral directors, antique dealers, auctioneers, jewelers, bakeries, farms, florists, mushroom consultants, beekeepers, pest control, interpreters, bookkeepers, insurance brokers, grant writers, travel advisors, tattooists, photographers, musicians, piano teachers, choir directors, wedding celebrants, genealogists, archivists, crossword constructors, tabletop operators, tailors, and shoemakers. English and Japanese material produced retained cases. An accessible Spanish interview was used for the recruiter.

Late searches repeatedly produced generic prompt collections, AI-service sales pages, hobby experiments, undisclosed operators, and future-tense suggestions. The last creative-work batch produced no further qualifying case. A separate sewing-and-workshop search did produce the Nigerian shoemaker, so the catalogue was expanded. A final batch covering puppeteers, chimney sweeps, piano tuners, and signwriters produced no further qualifying case.

This is a broad, bounded search with diminishing returns. It is not proof that no other examples exist. Lack of a retained case in an occupation says more about available public evidence than about whether people in that occupation use LLMs.

What these cases suggest for your own work

These are interpretations of the cases, not additional verified examples.

The next useful experiment is to choose one task you already repeated this week. Supply its actual input. Define the output you would use. Compare total completion time, including correction, against your normal method.