ChatGPT Prompt Success Stories: 21 Verified Cases, Not 25 Invented Ones

ChatGPT Prompt Success Stories: 19 Sourced Cases, Plus the Lab Evidence Nobody Quotes

Field verification log · Updated September 2026

By Tom Morgan · ~12 min read

Disclosure: No sponsorship, no affiliate links, no product placements, no template libraries. Every source is linked at the point of use and labeled by how independent it actually is. An earlier version of this article linked to a prompt-template site I run; a reader pointed out — correctly — that it sat awkwardly next to a “no sponsorship” byline. It’s gone.

Read this first: what “sourced” means

Each case below carries one of three labels:

• Tier A — Independently corroborated. Peer-reviewed research, a working paper from a named institution, or established business press confirming the company’s own account.
• Tier B — Company-published. The company’s own case study or a named executive on record. These are primary sources, but self-interested: the numbers are company-reported estimates unless the text says otherwise.
• Tier C — Self-reported. Individual accounts, treated as anecdote and flagged as such. None remain in this version.

Two structural notes before the list:

1. Fourteen of the nineteen company cases link to OpenAI’s customer-story pages. That’s a limitation of the evidence base, not an editorial choice — the companies above are the ones publishing named, attributable accounts, and OpenAI hosts most of them. It’s why Tier A cases are marked separately, and it’s why the most valuable section of this article isn’t company cases at all: it’s the controlled research from MIT, Harvard, and Microsoft Research near the end, including two studies that failed to reproduce the headline gains.
2. The genre’s canon is here because it’s load-bearing, not because it’s fresh. Klarna, the BCG study, and Duolingo appear in every serious write-up on this topic. They’re included with the complicating context attached. The research section is where this article tries to contribute something the listicles lack.

The index

Company-reported figures unless the source column says otherwise.

Headline figures worth holding onto: Indeed +20% applications · BBVA Peru −80% query time · Notion ~−95% build time on one feature · Travelers 85–90% self-completion · Klarna ~⅔ chats automated (Feb 2024 estimate, revised within sixteen months — see Case 17).

Cases 1–4 · No team, no budget, no margin for error

These are the cases with no PR team between the person and the claim. None of these people “wrote a prompt” in the engineering sense — they described a problem in plain language until the answer was usable.

Case 01 — Sharp & Sharp Certified Seed · Agriculture · Tier B

Rachael Sharp is taking over her family’s farm in Allendale, South Carolina, and inheriting her father’s handwritten planting records going back to 1971. Her own description of the stack: “It’s so much data, it almost scares you away.” She had ChatGPT turn the archive into something she could query on demand — past yields, planting dates, what worked on which field — in seconds rather than a drawer search.

The part that’s easy to miss: she’s not typing this at a desk. Riding the combine or walking a soybean field, she logs loads and checks details by voice, and photographs a stressed crop to ask what’s wrong with it on the spot. South Carolina has gone from over 200 certified seed farms to seven. Acting on decades of inherited judgment without breaking stride is, in that context, a survival trait.

“Before, I would wonder, where can I get this, who can help me with that. And now it’s like, okay, I can do this.”
— Don Sharp, Rachael’s father

54 years of paper records made queryable.
Sourced: OpenAI, “Small business stories,” Feb. 4, 2026

Case 02 — The Original Tamale Co. · Food / retail · Tier B

A third-generation family operation working hundreds of farmers markets across California. Two uses of the same tool, from two people running the same company — counted once, because it’s one company and one source page.

Use one — shipping a real feature with zero code. Customers kept losing track of which market the family would be at on a given day. Christian Ortega, who handles marketing and operations, had never written a line of code. He described the tool he wanted — searchable by ZIP code, near-me lookup, an editable market list — and had a working version live on the website the same afternoon. Not a demo: a functioning map-and-search tool, no developer hired, nobody waited on.

“I had the idea, made it that same afternoon, and put it on the website. I didn’t need to wait on anyone. I could just do it.”
— Christian Ortega

Use two — judgment in a second language. Xochitl Ortega co-runs the business in English, her second language. Addressing employees, handling a tense moment with a vendor, speaking publicly — all carried the extra weight of translating instinct in real time. Her approach: start in Spanish, work the phrasing with ChatGPT until it sounds like her, not like a translation app. Narrower than a coded feature, and the use most likely to apply directly to a reader running a business in a language that isn’t their first.

“I feel like I just went to university… now I can speak to anyone — from the person who cleans, to the CEO of a company. I’m confident.”
— Xochitl Ortega

Sourced: OpenAI, “Small business stories,” Feb. 4, 2026

Case 03 — Reno Salvage · Industrial · Tier B

An 86-year-old salvage yard run on institutional memory that lives in one or two heads. Manager Richard Lane’s job is triage: a plasma cutting table fails mid-shift, a customer needs a spec nobody has memorized. When the cutter broke, instead of waiting days for a technician, he described the fault to ChatGPT and got a troubleshooting step that fixed it on the spot.

The bigger fix was a part-numbering system for more than a thousand products the yard had postponed for years because doing it by hand looked like weeks of work. Lane built it in an afternoon, organized so the crew could remember the scheme rather than just look it up.

“I don’t necessarily see AI as the answer to all the questions. It’s more like a partner helping you find answers in yourself.”
— Richard Lane

1,000+ parts catalogued in one afternoon.
Sourced: OpenAI, “Small business stories,” Feb. 4, 2026

Case 04 — Morgan Stanley · Financial services · Tier A

Morgan Stanley built an internal research assistant on GPT-4 — a system its ~15,000–16,000 financial advisors could query in plain English against the firm’s research library, roughly 100,000 documents. The rollout was deliberately cautious: a pilot with a few hundred advisors, internal evaluation against the firm’s own standards, and only then a firmwide launch in September 2023.

This case is Tier A, and the corroboration is now linked rather than asserted. The launch was independently reported by CNBC and Fortune, which confirm the tool’s existence, the advisor count, the document corpus, and the GPT-4 basis. What the independent coverage does not confirm: the later performance figures. The 98% adoption rate across advisor teams, the jump in document access from 20% to 80%, and the claim that the system went from answering 7,000 questions to “any question” in the corpus all come from Morgan Stanley’s own 2024 press release and OpenAI’s updated case page — company-reported, and labeled as such here.

Two reasons this case matters beyond its numbers. First, it’s the one deployment in this list where skeptical financial press covered the launch contemporaneously rather than recycling a vendor page. Second, the pattern — pilot, internal eval, staged rollout — is the discipline the rest of this article keeps pointing at, executed at a scale where “ship it and see” was never an option.

Sourced: OpenAI case page · CNBC, Sept. 18, 2023 · Fortune, Sept. 20, 2023

Cases 5–7 · The spec replaced the prompt

At three software companies, the unit of work shifted from “write the function” to “describe the result precisely enough that something else can write the function.” That distinction is the closest thing to a transferable skill in this list: specificity does the work, not cleverness. For the mechanics, the open resources beat any template library — DAIR.AI’s Prompt Engineering Guide and the OpenAI Evals repository, which is what “watch it fail in a specific way” looks like as engineering practice.

Case 05 — Notion · Software · Tier B

Ryan Nystrom leads AI Product Engineering at Notion. A voice-input feature existed on mobile but not on web or desktop; the historical cost of porting it was three coordinated teams and roughly two weeks. Nystrom pointed a coding agent at the mobile codebase, described how the feature needed to look and behave on web, and gave it a way to verify its own output. It returned a complete first cut that matched Notion’s conventions closely enough to ship the next day.

His framing of what changed: “I’ve almost found myself spending a lot more time writing these spec documents that I can hand off and let it work on. Honestly, I don’t really write code by hand anymore.” Not a productivity hack — a different job description for the same title.

~2 weeks → 3–4 hours for one cross-platform feature — a single, well-scoped feature, which is exactly the kind of task where this kind of figure is believable.
Sourced: OpenAI, “What Codex unlocks for Notion,” June 9, 2026

Case 06 — Nextdoor · Software · Tier B

Cory Dolphin, head of engineering at Nextdoor, named the shift directly: “away from iteratively prompting an agent, and towards outcome engineering — engineers start to think about the result they want to see and work with an agent to engineer that result.” On the Opportunity Alerts feature, an engineer wanted service providers shown on a map — historically a request needing mobile, frontend, and backend coordination, and one that might never have cleared the backlog. One person built it end to end.

His team also uses the same tooling for the unglamorous work: debugging Kubernetes pods that won’t start and tracing race conditions in Rust that are genuinely hard to reproduce by hand. The honest line from his own account is worth keeping: “a lot of the team are addicted to it” — an admission, not a slogan.
Sourced: OpenAI, “How engineers at Nextdoor use Codex,” June 9, 2026

Case 07 — Datadog · Software · Tier B

The case to show a skeptic first, because Datadog didn’t take the result on faith. Brad Carter’s team built a replay harness: they pulled pull requests that had already caused real production incidents, ran the agent against each one as if it were the original reviewer, and asked the engineers who’d lived through those incidents whether the feedback would have changed anything.

The result, stated plainly rather than rounded up: the agent flagged something useful in roughly 22% of the incidents tested — on pull requests that had already passed human review. More than 1,000 engineers now use it regularly. The internal success signal isn’t a dashboard metric; it’s engineers posting in Slack when a comment changed how they thought about a problem.
Sourced: OpenAI, “Datadog uses Codex for system-level code review,” Jan. 9, 2026

The loop every enterprise case below actually ran

1. Ship a small change — never a big-bang rollout.
2. Test it on real past cases — replays, A/B tests, historical data. Not vibes.
3. Ask the people who’d know — the engineers, advisors, or clinicians who lived through the original cases.
4. Keep it, revert it, or say so — including in public, including when it’s expensive. Case 17 is what step 4 looks like when it costs you.

Cases 8–16 · Where a wrong prompt has a dollar value

At consumer scale, in a regulated industry, or in front of a patient, “good enough on first try” is not an option. Each of these ran an A/B test, replay, or staged rollout before anyone called it a success.

Case 08 — Indeed · HR tech · Tier B

The technique was naming the why, not just the what. “Invite to Apply” matched job seekers to roles but didn’t explain matches, and people apply more often — and more successfully — to jobs they understand. Using few-shot prompting (worked examples of good explanations), the team got GPT to generate that context at scale, then fine-tuned a smaller model to match the quality at 60% fewer tokens once the approach was proven. In Indeed’s own A/B test: +20% started applications and +13% further down the funnel — interviews and hires, not clicks. “Regardless of how good our matching is, explainability is key,” said CEO Chris Hyams.
Sourced: OpenAI · Indeed

Case 09 — Lowe’s · Retail · Tier B

Two assistants on one foundation: Mylow for customers on the website, Mylow Companion for associates in 1,700+ stores. Company-reported: conversion more than doubled among shoppers who used the assistant, and in-store customer satisfaction rose roughly 200 basis points on assisted interactions.
Sourced: OpenAI

Case 10 — Intercom · SaaS / support · Tier B

Intercom’s support agent, Fin, sits directly in front of customers, so a model migration is a live-fire exercise: swap the model wrong and thousands of conversations degrade at once. Their practice is to validate a candidate model against historical conversation data before any traffic shift — outcome: a full migration validated and shipped in 48 hours, because the evaluation harness, not the prompt, did the risky work. The prompt is the easy part; knowing the new model won’t regress on last month’s hardest tickets is the product.
Sourced: OpenAI

Case 11 — BBVA · Banking · Tier B

One bank, two data points. In Spain, an internal program led to 20,000+ employee-built assistants, with the bank estimating ~2.8 hours saved per employee per week (a company estimate — treat it as one). In Peru, query handling time fell from 7.5 minutes to about one. The Spain figure illustrates the governance model; the Peru figure is the cleaner, harder number of the two.
Sourced: OpenAI

Case 12 — LSEG · Financial data · Tier B

London Stock Exchange Group’s data-products team compressed release cycles from 3–6 months to about two weeks by pairing engineers with coding agents on well-specified work. Note the boundary: this is software delivery inside a data company — no part of LSEG’s published account involves LLMs touching trading infrastructure, and any account implying otherwise would be a different, much more demanding claim.
Sourced: OpenAI

Case 13 — Preply · Edtech · Tier B

A language-tutoring marketplace’s version of the “why” problem: tutors couldn’t give written feedback on every lesson at scale. Preply uses GPT to generate per-lesson personalized feedback from lesson content — grounded in the actual lesson’s material rather than a generic “praise the student” template, which tutors and students can both smell.
Sourced: OpenAI

Case 14 — Singular Bank · Private banking · Tier B

Portugal-based private bank; bankers reported reclaiming 60–90 minutes per day — the softest number in this list, a self-report survey, so weigh it accordingly. It’s the pattern (assistants for document-heavy knowledge work, inside compliance guardrails) rather than the precise figure that’s transferable.
Sourced: OpenAI

Case 15 — Boston Children’s Hospital · Healthcare · Tier B

Clinicians used GPT-4-based tooling to help surface candidate diagnoses in 40+ rare-disease cases that had previously gone undiagnosed, with physician review central to the workflow — the model proposes, the clinician disposes. Nothing here is autonomous diagnosis, and the published account doesn’t claim otherwise. If a version of this story you read elsewhere dropped the physician-in-the-loop detail, that tells you more about the reteller than the case.
Sourced: OpenAI

Case 16 — Travelers · Insurance · Tier B

An AI assistant for first-notice-of-loss: 85–90% of claims in supported categories are self-completed by policyholders, with humans taking the exceptions. The highest percentage in this article, and the most context-dependent — “claims the assistant was allowed to handle” is not “all claims.” The honest framing isn’t “90% of insurance claims are now AI-handled”; it’s “in the categories where Travelers deployed it, most eligible claims no longer need a human to start.”
Sourced: OpenAI

Cases 17–18 · The complicating cases

Case 17 — Klarna · Fintech · Tier A

Act one (February 2024). Klarna announced its AI assistant had handled 2.3 million conversations in its first month — work equivalent to roughly 700 full-time agents — with resolution times around 11 minutes versus 25 for humans, and projected a ~$40 million profit improvement. The “two-thirds of customer chats automated” figure still circulating in listicles dates from this press release. It was a company estimate at a moment of maximum enthusiasm, and the listicles never carried what happened next.

Act two (2025). Bloomberg reported that Klarna was hiring humans again and shifting premium customers back to people, with executives acknowledging quality, brand, and complexity costs the automation push had undervalued. Klarna’s messaging changed accordingly — away from headcount framing, toward AI as an assistant to human teams.

Both acts are true at once. The efficiency gains were real; so were the costs, discovered the expensive way, in public. A listicle that prints act one is marketing. One that prints act two alone is anti-AI marketing.
Sourced: Bloomberg, May 8, 2025

Case 18 — Duolingo · Edtech · Tier B

The most instructive “success story” here is really a story about refusing to ship. Duolingo evaluated GPT-3 and concluded it wasn’t good enough — not accurate enough, not controllable enough for a product whose entire value is correctness. They waited for GPT-4, then shipped Duolingo Max. The product exists because the team was willing to say “not yet” to a model the industry was racing to deploy.

Then, in 2025, Duolingo went further: an “AI-first” internal memo, reduced contractor work in favor of AI workflows — a decision that drew real criticism and landed differently than the measured 2023 story. Both episodes are the same company learning in public: first, “don’t ship until it’s ready”; then, “maybe we overshot.” The through-line isn’t a prompting technique. It’s a willingness to revise the plan when the evidence moves.
Sourced: OpenAI · 2025 “AI-first” memo as reported by Bloomberg and others.

Part 5 · The lab evidence — including the studies that failed

Company cases tell you what organizations claim. Controlled studies tell you what happens when people are randomly assigned to a tool and measured by someone with no stake in the outcome. Four studies do most of the work in this debate, and an honest review has to include the two that complicate the picture.

Study A — Noy & Zhang (MIT) · Science, 2023 · Tier A

A preregistered randomized experiment: 453 college-educated professionals given occupation-specific writing tasks, half with ChatGPT access. Results: 40% less time, 18% higher quality as graded by blinded professionals. Inequality between workers decreased — the weakest writers gained the most. Published in Science, not a vendor blog: doi.org/10.1126/science.adh2586.

Limitations worth stating: short, incentivized writing tasks are not a job. Nobody’s quarter-end report has a grader paying bonuses for above-average work. This is evidence about a class of tasks, not about employment.

Study B — Peng et al. (GitHub / Microsoft Research) · 2023 · Tier A, with a flag

95 professional developers, one standardized task (an HTTP server in JavaScript). The group with GitHub Copilot finished 55.8% faster — 71 minutes versus 161 — with a 95% confidence interval of [21%, 89%], and a higher completion rate (78% vs 70%): Microsoft Research.

The flag: the authors were GitHub-affiliated, studying GitHub’s own product. The methodology is sound and the task was standardized, but vendor-adjacent research deserves a discount on interpretation — same standard this article applies to OpenAI case pages.

Study C — the replication that didn’t reproduce · 2026 · Tier A

A 2026 study analyzing two years of commit-based developer activity found Copilot users showed a small average increase in weekly activity that was not statistically significant after adoption — a null result in real organizational settings where Study B found 55.8% in a lab task: arXiv:2509.20353. A large-scale study of Danish workers (Humlum & Vestergaard, 2025) similarly found time savings of only ~3% on average — real, but an order of magnitude below the lab figure, with gains concentrated among novice users.

Both results are true at once. Study B measured a motivated developer on a well-specified task with the tool in hand. Studies C measured years of ordinary organizational work, where tasks are ill-specified, tooling is unevenly adopted, and most of the jagged frontier’s danger zone is. The gap between 55.8% and 3% isn’t anyone lying — it’s the difference between a task the model is good at and a job, which is mostly tasks the model is unevenly good at. This is Case 18’s frontier finding at population scale, and it should sit next to every “2 weeks → 3 hours” figure in this article, including the ones above.

For a broader synthesis — including the field experiments finding ~26% more completed tasks at Microsoft and Accenture — the OECD’s 2025 review is the most balanced single document available.

Case 19 — The jagged frontier study · Research · Tier A

The paper that ties all of it together: Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon — HBS Working Paper 24-013, 758 consultants at Boston Consulting Group using GPT-4.

• Inside the model’s competence boundary: consultants using AI finished 12% more tasks, 25% faster, at 40% higher quality than their own baseline. Below-average performers gained the most.
• Outside it — the finding every summary skips: on tasks just past the frontier, consultants with AI were worse than consultants without it. Blind trust in fluent, confident, wrong answers pulled correct performance down roughly 19 percentage points. The authors call it “falling asleep at the wheel.”

This single paper explains almost every other entry in the article: why Datadog replays history, why Morgan Stanley piloted with hundreds before sixteen thousand, why Klarna’s act one was oversold, and why Studies B and C disagree. The danger was never the model’s failures. It’s that the failures look identical to the successes.
Sourced: HBS Working Paper 24-013 · If you read one outside writer on applied AI, make it co-author Ethan Mollick’s One Useful Thing.

What survives scrutiny

Strip away the vendor adjectives and three patterns hold across the nineteen cases, the four studies, and three years:

1. Nobody won with one magic prompt. They won with a prompt tested against real cases, watched until it failed in a specific way, and fixed that one thing. The Datadog replay harness and the BCG jagged frontier are the same lesson wearing different clothes.
2. The unit of work became the spec. Describe the outcome precisely — including how to verify it — and the wording takes care of itself. Prompt templates are downstream of this, not a substitute for it.
3. The winners budgeted for being wrong. Pilot groups, A/B tests, staged rollouts, public reversals. The genre’s worst stories — Klarna’s 2024 headline, the “700 agents” stat echoing through listicles to this day — all skipped this step or had it skipped on their behalf by people summarizing them.

One last habit, worth more than any single case above: the genre’s most-recycled numbers traveled listicle to listicle for over a year without anyone linking the original claim. Check the link, or treat the number as folklore.

People who will sharpen your judgment on this topic

None of these links benefit me:

• Ethan Mollick — One Useful Thing (Wharton; co-author of Case 19). The most consistently honest applied-AI writing anywhere.
• Simon Willison — the sharpest working practitioner writing on LLMs, evals, and what vendors’ claims look like from inside real projects.
• AI Snake Oil — Arvind Narayanan and Sayash Kapoor on what’s real and what isn’t, with an unusually good nose for inflated benchmark claims.
• DAIR.AI — Prompt Engineering Guide — the open, maintained reference for the actual mechanics.
• OpenAI Evals — the “test it on real past cases” loop, as open-source practice. If you take one habit from the enterprise cases, take this one.
• OECD, “The effects of generative AI on productivity, innovation and entrepreneurship” (2025) — the best single synthesis of the lab-and-field evidence, positive and null.

Changelog

• v3 (Sept 2026): Added independent launch coverage for Case 04 (CNBC, Fortune); separated company-reported from independently corroborated figures within the case. Added Part 5 (controlled studies, including two failed replications). Removed unverifiable summary claims and merged the duplicate summary tables. Count unchanged at 19 cases plus the research section.
• v2: 21 → 19 cases; The Original Tamale Co. and BBVA consolidated from two entries each into one; Tier C case cut.
• v1: 25 cases.

About the author: Tom Morgan. This article is updated in place when sources move or claims fail. Corrections welcome; the standard for inclusion is a named person, a named organization, and a link you can check.