The Internet Ninjathe internet ninja.
Get cited
guideAI in operations · 17 min read

AI agents examples: what deploying companies actually report

by TIN research desk updated uses 19 case files read as markdown
Contents · 13 sections
The short answer

AI agent examples with a public record include Klarna’s customer-service assistant, Otto’s stock-ordering system and GitHub Copilot in randomised trials at three companies: named organisations’ AI systems doing or drafting real work, each with a published result or finding.

Klarna said its assistant handled “two-thirds” of customer-service chats in its first month source source; in 2025 its CEO said “what you end up having is lower quality” source source. ING says chats its chatbot handled without human support rose from 46% at end-2023 to 75% at end-2025 source source. Field trials of GitHub Copilot at three companies found a “26.08% increase (SE: 10.3%) in completed tasks” source source. The SEC found Presto’s drive-thru voice AI “required human agent intervention, including entering the order, in all instances” source source.

What counts as an AI agent here

TIN’s working definition for this page: an AI agent is a system that carries out or drafts a task on an organisation’s real work, such as answering a customer, drafting a reply, writing code, ordering stock or adjusting a patient’s oxygen, rather than only scoring a case or flagging it for a person. The definition is about the work, not the product category, so the examples of AI agents on this page are deployments: each names the organisation, the work and the result.

The selection rule: a live TIN case file is in the table if a named organisation put an AI system on its own real work, or in a field trial on real work; the system carries out or drafts the task itself; and the case file records a result or a finding. Two rows are in the table for a different reason: Air Canada, whose tribunal decision does not describe the chatbot as AI, and Nate, where prosecutors allege the automation rate was effectively zero; both are there because the record turns on what the system did or was claimed to do: a tribunal held Air Canada to its chatbot’s answer, and the founder of Nate is charged, and presumed innocent. Case files where the system only flags cases for a person, where the figure measures AI spend rather than work done, or where the result is a research benchmark rather than a deployment were left out.

The table holds 19 rows, one per case file, and each row links to its file, so the count can be checked by counting the links.

This page does not rank agent products. Textbook taxonomies sort agents by how they are built; this page groups them by what they do on the job, because that is what the deploying companies report against:

  1. Agents that talk to customers.
  2. Assistants that work beside employees.
  3. Systems that act on their own in operations.
  4. Cases where the agent turned out to be people, or where the company was held to what its chatbot said.

The “who measured” column uses one key throughout.

Who measuredWhat it means
CompanyThe deploying company’s own statement, filing or press release. Press that relays it does not change this.
VendorThe company that sold the system.
ResearchersA published trial or study, with any vendor tie stated.
RegulatorA regulator’s order, a prosecutor’s charge or a tribunal’s decision, with its status.
PressAn independent outlet’s report, not an audit.

Agents that talk to customers

Six case files, all on customer service. The depth version is TIN’s review of agentic AI in customer service, and the banking cases are covered in AI in banking customer service.

CaseWhat the agent doesResult, as the case file words itWho measured
Klarna source, February 2024OpenAI-powered assistant for refunds, returns, payment issues, cancellations, disputes and invoice queriesFirst month: “2.3 million conversations”, “two-thirds” of customer-service chats; errands resolved in “less than 2 mins compared to 11 mins previously”; a “25% drop in repeat inquiries”Company
Notion with Decagon sourceSupport agent that automates repetitive inquiries and routes the rest to the right expert”34% faster ticket resolution” (in the case study’s text, resolution time “improved up to 34%”), “2x increase in deflection”, “3.4% ask for human rate”Vendor (Decagon’s case study; Notion’s Global Head of Customer Experience is named and quoted)
Octopus Energy source, 2023Kraken “Magic Ink” drafts email replies; a person reviews and sends each oneCEO, May 2023: by the end of April it was “answering 34% of all customer queries”, with “an 80% satisfaction rating. Humans get 65%”; Kraken’s own case study puts satisfaction “around 70%” sourceCompany (a CEO statement relayed by Sifted, City AM and TheWrap source)
ING source, 2023 to 2025Chatbot, with GenAI chatbots live in seven countriesChats handled without human support, at year end: 46% (2023), 54% (2024), 75% (2025); “>80% of chats resolved without human support in 4 countries, rolled out in 7 countries”Company (investor presentation and annual report)
Nordea source, October 2024 to September 2025Nova chatbot on boost.ai’s platform, live since 2017; in Sweden, chats reach it before a human”5.2 million chats; ~78% handled by Nova Chatbot”; separately, “76% of all chat sessions are now handled by our chatbots” source; boost.ai: “a 91% in-scope resolution rate for private banking customers” in Nordea Sweden sourceCompany; the 91% is Vendor
MakeMyTrip source, Q1 FY27 (quarter to 30 June 2026)Myra 2.0 conversational travel assistant, including bookings”Myra engaged in over 85,000 daily conversations in Q1 FY27”; “resolving over 50% of post-booking queries in relation to flights and hotels during Q1 FY27”Company (SEC Form 6-K exhibit and earnings call)

What the headline figures count

These figures count different things, so they cannot be compared or averaged. ING’s metric is “chatbot deflection”, which ING defines as “the percentage of chats handled by the chatbot without human support, at year end” source source. The slide does not say which countries the figure covers or label it GenAI-only source. ING’s annual report also says it “reduced inbound contacts to contact centres by 43 percent”, but ING credits several digital changes for that, so it is not a chatbot figure source source. McKinsey, which built the 2023 pilot, says that “within the first seven weeks of use” the chatbot was “helping 20 percent more customers avoid long wait times” compared with the previous solution; no ING filing or release found for the case file repeats that figure source source.

Nordea’s wordings do not define “handled”, so TIN’s case file does not read them as “resolved without a human” source. The boost.ai figure counts “conversations that are marked as ‘solved’”, which leaves chats outside the bot’s topics out of the denominator, and it covers Nordea Sweden only source source. When FinTech Futures repeated it in November 2022, it dropped both the in-scope qualifier and the Sweden scope source. MakeMyTrip’s filing scopes its resolution figure to post-booking flight and hotel queries, while on the earnings call the CEO put it more broadly, saying the bot “independently resolves over 50% of customer calls”, as MediaNama reported source; the filing’s narrower wording is the one to repeat source source. The Notion case study gives no baseline for its deflection figure source.

What happened to the jobs

Klarna said its assistant was “doing the equivalent work of 700 full-time agents”, which TIN’s case file records as a modeled equivalence, not a headcount cut source source. By May 2025 CNBC reported that Klarna “will soon look to recruit more human customer service agents” (Klarna’s 2025 reversal, explained) source source.

Octopus Energy’s CEO said the AI was “doing the work of 250 people by answering customer emails”, a workload estimate, and said the technology was “unlikely to lead to job cuts” source source. In 2024 Octopus announced 4,000 new jobs (the Octopus Energy case) source source.

Assistants that work beside employees

Four case files where the AI helps a person who still does the job.

CaseWhat it doesResult, as the case file words itWho measured
JPMorgan Chase, LLM Suite source, launched summer 2024In-house generative-AI assistant staff use to draft presentations, analyse earnings transcripts and compare documents”went from zero to 200,000 onboarded users within eight months”; the chief analytics officer: “a little under half of JPMorgan employees use gen AI tools every single day” sourceCompany (trade press relays the bank’s count)
Moderna, ChatGPT Enterprise source, April 2024Company-wide assistant; staff build custom GPTsWithin two months, “Moderna had 750 GPTs across the company”; “each user has 120 ChatGPT Enterprise conversations per week on average”; OpenAI states the legal team “boasts 100% adoption”Vendor (OpenAI’s case study)
GitHub Copilot, three company trials sourceCode-completion assistantPooled across trials at Microsoft, Accenture and an anonymous Fortune 100 company, 4,867 developers: “26.08% increase (SE: 10.3%) in completed tasks”; recent hires and junior developers “27% to 39%”, senior developers “8% to 13%“Researchers (peer-reviewed in Management Science, and post-registered as AEARCTR-0014530; two of the six authors are Microsoft employees)
Tutor CoPilot, with FEV Tutor sourceSuggests moves to human tutors during live sessions; the tutor, not the AI, works with the studentStudents “4 percentage points (p.p.) more likely to master topics (p<0.01)”; students of lower-rated tutors “improving mastery by 9 p.p.”; a trial of “900 tutors and 1,800 K-12 students”Researchers (a preregistered trial by the Stanford team that built the tool)

Usage is not an outcome

JPMorgan’s and Moderna’s figures count users, GPTs and conversations: they show the tool is used, not what the use produced. JPMorgan’s productivity estimate has no single figure: The Digital Banker reports “an estimated three to six hours per week” source source, while AI News quotes the chief analytics officer describing “an hour saved here and three hours there” and cautioning that such snips “often just shift bottlenecks” source. Tearsheet reports a “10% to 20% productivity increase” for coding only source source. Moderna’s usage figures come from OpenAI’s case study, and TIN’s case file states plainly: “They are not independently audited.” source source

Copilot and Tutor CoPilot were measured in randomized trials against a control group, which is why their figures say more source source. They also carry their own limits. A separate METR trial found the opposite sign for a different group, slowing experienced open-source developers by 19% source source. Tutor CoPilot rests on one working paper that had not completed journal peer review, and its tutors flagged “suggestions that are not grade-level appropriate” source source. The trial record for coding assistants is in TIN’s review of AI coding agents, and assistants that help a person are covered in agent assist.

Systems that act on their own

Six case files where the system does or drafts the task, from ordering stock to adjusting oxygen flow; in the Boston Children’s study the model proposed diagnoses that experts confirmed. Automation that runs the work is covered in TIN’s reviews of AI workflow automation, AI back office automation and supply chain automation.

CaseWhat it doesResult, as the case file words itWho measured
Otto, Germany source, reported 2017Blue Yonder deep-learning system that forecasts demand and orders stock; not a language model”predicts with 90% accuracy what will be sold within 30 days”; buys “around 200,000 items a month from third-party brands with no human intervention”; surplus stock “declined by a fifth”; returns cut “by more than 2m items a year”; “Otto did not fire anyone”Press (The Economist); the accuracy definition is Otto’s Michael Sinn’s, quoted by Retail Systems in 2017 source
Omega Healthcare with UiPath sourceUiPath Document Understanding automates documents across billing, medical coding and insurance correspondenceDocumentation time cut “by 40%”; turnaround cut “by 50% with a process accuracy of 99.5%”; “monthly savings of 6,700 worker hours” (October 2024) and “more than 15,000 hours a month” (June 2025), two dates rather than two measurementsCompany and Vendor (the case file: “repetition, not independent measurement”)
JPMorgan COIN source, 2017Machine learning that interprets commercial-loan agreements; not a language modelThe work had “consumed 360,000 hours of work each year by lawyers and loan officers”; “12,000 new wholesale contracts per year”; “It has never been independently measured or updated.”Company (one 2017 Bloomberg article citing the bank’s designers)
Boston Children’s Hospital with OpenAI o3 Deep Research sourceA reasoning model proposed explanations for unsolved rare-disease cases; experts and a certified lab confirmed each diagnosisThe NEJM AI abstract: “overall diagnostic yield of 18 of 376 (4.8%, [CI, 2.9 to 7.5])” in previously unsolved cases; the case file records 7 of the 18 as rediscoveries and notes that reviewers were not blinded to model confidenceResearchers (peer-reviewed in NEJM AI; OpenAI supplied the model and OpenAI-affiliated authors co-wrote the paper)
SAVE-O2 AI trial source, 4 US hospitals”a closed-loop, trend-based, autonomous oxygen titration system”; not a language modelShare of time in the target oxygen range: “85% [1%] vs 63% [2%]” against usual care; 300 adults; a single, “unblinded” trialResearchers (peer-reviewed in JAMA Internal Medicine; sponsored by the Defense Health Agency)
John Deere See & Spray sourceComputer vision fires only the nozzles over a weed; not a language model”saved a range of 28.4% to 62.4% on postemergence herbicides” (peer-reviewed, Weed Technology) source; “43 to 59 percent reduction in herbicide use compared to broadcast applications” (University of Arkansas); “average product savings across all fields was 76%” (Iowa State) source; at a low sensitivity setting, the pigweed population rose 280% each year (University of Arkansas)Researchers (independent universities)

Otto’s figures are roughly 2017-vintage and unaudited, and COIN’s have not been updated since the 2017 article source source. The three John Deere figures come from three separate studies and are not one range source source source. At Boston Children’s, the case file records that the model did not diagnose any patient or make any clinical decision, and a finding counted only after expert review and laboratory confirmation source.

When the agent turned out to be people, or a company was held to its chatbot

Three case files where a regulator, a prosecutor or a tribunal decided what the system did. Each carries its status. The longer record is in TIN’s review of AI agent fraud.

CaseWhat happenedFinding or charge, as the case file words itWho measured
Presto Automation source, SEC 2025Drive-thru voice AI that filings said “eliminat[es] human order taking”Off-site human agents entered the orders; the 2023 pilot “required a human agent to enter the orders approximately 70% of the time”; reported automation rates of 95% to 99% counted only orders completed without restaurant-staff involvementRegulator (SEC order; Presto consented “without admitting or denying the findings”)
Nate source, DOJ and SEC 2025Shopping app that claimed it was “able to transact online without human intervention”The DOJ alleges “the app’s actual automation rate was effectively zero percent”, with “hundreds of contractors, or ‘purchasing assistants,’ in a call center located in the Philippines”Regulator (charges, not findings; the founder “is presumed innocent unless and until proven guilty”)
Moffatt v. Air Canada source, 2024 BCCRT 149Website chatbot gave wrong bereavement-fare adviceThe tribunal held that “It makes no difference whether the information comes from a static page or a chatbot” and found “Air Canada did not take reasonable care to ensure its chatbot was accurate”; it ordered damages, interest and feesRegulator (tribunal decision)

Presto’s order is explicit that the original system “was designed to rely on this human support and was not capable of processing orders without it” source source. Air Canada is the case where the company was held to what its chatbot said. The tribunal decision never describes that chatbot as AI, and TIN’s case file makes no claim about its architecture source.

Who measured these results

Counted from the “who measured” column of the four tables above.

Who measuredRowsCases
Company or its vendor10Klarna, Notion with Decagon, Octopus Energy, ING, Nordea, MakeMyTrip, JPMorgan LLM Suite, Moderna, Omega Healthcare, JPMorgan COIN
Researchers5GitHub Copilot, Tutor CoPilot, Boston Children’s, SAVE-O2, John Deere See & Spray
Regulator, prosecutor or tribunal3Presto, Nate, Air Canada
Independent press1Otto

TIN’s inference from these case files, not a measured finding: outside the companies and their vendors, the outcome figures in this record come from trials and university studies; the regulator, prosecutor and tribunal rows add findings and charges about what a system did, not outcome figures; and the production figures for agents that talk to customers are the companies’ own. The one press row is no exception, because The Economist’s Otto figures were sourced from Otto and Blue Yonder source.

Questions to ask before you repeat an AI agent example

Who measured it? Read the column before the number. A company’s investor presentation, a vendor case study and a randomized trial are three different kinds of evidence, and each of these AI agent examples says which one it rests on.

What does the denominator count? ING counts chats handled without human support at year end source. Nordea’s “handled” is undefined, and the boost.ai figure leaves out chats the bot was not built for source. Two “resolution” figures can describe different sets of chats.

What window? Klarna’s headline is its first month source. MakeMyTrip’s is one quarter source. Otto’s and COIN’s are from 2017 source source. A figure without its window cannot be checked against anything later.

Is a person still in the loop? At Octopus Energy a person reviews and sends every reply source. In Tutor CoPilot the tutor works with the student source. At Presto, the SEC found that off-site people entered the orders the voice AI was credited with source.

Is it still running as described? Klarna’s 2024 figures were followed by its CEO’s own account of lower quality and a move back to human agents source.

What this record cannot tell you

  • It is not a sample of AI agents in use. The table holds what reached the public record: a company’s own publication, a vendor case study, a trial, a regulator or a tribunal. Deployments that published nothing are absent.
  • Many production figures are self-reported. The tally above gives the count, and each row says which.
  • The figures cannot be compared or averaged. Deflection, handled, in-scope resolution and post-booking resolution are different measures over different windows, so this page gives no average and no ranking.
  • Some figures are old or short. Otto and COIN date from 2017; Klarna’s is one month and MakeMyTrip’s one quarter.
  • The record leans American. 12 of the 19 case files are tagged as US cases.
  • The selection is TIN’s. These 19 cases were chosen by the rule stated above, from TIN’s own case files.

Sources

  1. TIN case file, Klarna. https://theinternetninja.com/stories/klarna-ai-assistant-customer-service/ Primary record: https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/ Reversal: https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-ai-customer-service-buy-now-pay-later/747586/ and https://www.cnbc.com/2025/05/14/klarna-ceo-says-ai-helped-company-shrink-workforce-by-40percent.html
  2. TIN case file, Notion with Decagon. https://theinternetninja.com/stories/decagon-notion-ai-cx/ Primary record: https://decagon.ai/case-studies/notion
  3. TIN case file, Octopus Energy. https://theinternetninja.com/stories/octopus-energy-kraken-ai-support/ Primary record: https://sifted.eu/articles/octopus-energy-ceo-chatgpt-news, https://www.cityam.com/ai-doing-the-work-of-over-200-people-at-octopus-chief-executive-says/, https://www.thewrap.com/ai-satisfies-customers-better-than-humans/, https://www.techuk.org/resource/case-study-kraken-tech-s-generative-ai-tool-for-customer-service.html and https://octopus.energy/press/octopus-energy-4000-jobs/
  4. TIN case file, ING. https://theinternetninja.com/stories/ing-netherlands-generative-ai-customer-service-chatbot-the-before-after-on-custo/ Primary record: https://ing.com/binaries/content/assets/documents/investors/20260318---driving-value-through-digitalisation-scalability-and-genai.pdf and https://ing.com/binaries/content/assets/documents/annual-reports/2025-ing-bank-nv-annual-report.pdf McKinsey’s account of the pilot: https://web.archive.org/web/20260114105101/https://www.mckinsey.com/industries/financial-services/how-we-help-clients/banking-on-innovation-how-ing-uses-generative-ai-to-put-people-first
  5. TIN case file, Nordea. https://theinternetninja.com/stories/nordea-ai-agent-resolves-91-of-in-scope-private-customer-chats-and-cuts-fraud-lo/ Primary record: https://www.nordea.com/en/doc/cio-presentation-cmd25-nordea.pdf, https://www.nordea.com/en/news/relationships-before-robots-how-nordea-is-leveraging-ai-in-digital-banking and https://www.boost.ai/case-studies/nordea-employs-comprehensive-conversational-ai-strategy-to-scale-customer-service
  6. TIN case file, MakeMyTrip Myra 2.0. https://theinternetninja.com/stories/makemytrip-myra-2-0-conversational-booking-q1-fy27-2026/ Primary record: https://www.sec.gov/Archives/edgar/data/0001495153/000119312526329581/mmyt-ex99_1.htm Trade press: https://www.medianama.com/2026/08/223-makemytrip-ai-myra2-0-bookingq1-fy27-earnings-call/
  7. TIN case file, JPMorgan Chase LLM Suite. https://theinternetninja.com/stories/jpmorgan-chase-llm-suite-genai-200000-employees-eight-months-2025/ Primary record: https://www.jpmorganchase.com/about/technology/blog/llmsuite-ab-award Productivity reporting: https://thedigitalbanker.com/jpmorgan-chases-llm-suite-drives-ai-transformation-across-the-enterprise/, https://www.artificialintelligence-news.com/news/jpmorgan-chase-ai-strategy-2025/ and https://tearsheet.co/artificial-intelligence/jpmorgan-chases-gen-ai-implementation-450-use-cases-and-lessons-learned/
  8. TIN case file, Moderna and ChatGPT Enterprise. https://theinternetninja.com/stories/moderna-openai-chatgpt-enterprise/ Primary record: https://openai.com/index/moderna/
  9. TIN case file, GitHub Copilot field experiments. https://theinternetninja.com/stories/generative-ai-copilot-three-field-experiments-software-developers-26pct-tasks-2024/ Primary record: https://economics.mit.edu/sites/default/files/inline-files/draft_copilot_experiments.pdf and https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  10. TIN case file, Tutor CoPilot. https://theinternetninja.com/stories/tutor-copilot-stanford-fev-rct-4pp-mastery-9pp-lower-rated-tutors-2024/ Primary record: https://arxiv.org/abs/2410.03017
  11. TIN case file, Otto. https://theinternetninja.com/stories/otto-germany-ai-stock-replenishment/ Primary record: https://www.economist.com/business/2017/04/12/how-germanys-otto-uses-artificial-intelligence and https://www.retail-systems.com/rs/OTTO_Blue_Yonder_AI_Delivery_Times.php
  12. TIN case file, Omega Healthcare and UiPath. https://theinternetninja.com/stories/omega-healthcare-uipath-document-automation/ Primary record: https://www.businessinsider.com/omega-healthcare-uipath-ai-document-processing-health-transactions-2025-6 and https://www.prnewswire.com/in/news-releases/omega-healthcare-processes-60-million-transactions-with-enterprise-ai-and-automation-from-uipath-302285442.html
  13. TIN case file, JPMorgan COIN. https://theinternetninja.com/stories/jpmorgan-coin-contract-intelligence/ Primary record: https://www.bloomberg.com/news/articles/2017-02-28/jpmorgan-marshals-an-army-of-developers-to-automate-high-finance
  14. TIN case file, Boston Children’s Hospital and OpenAI o3 Deep Research. https://theinternetninja.com/stories/boston-childrens-manton-openai-o3-deep-research-rare-disease-reanalysis-18-diagnoses-376-cases-nejm-ai-2026/ Primary record: https://europepmc.org/article/MED/42529462
  15. TIN case file, SAVE-O2 AI trial. https://theinternetninja.com/stories/save-o2-ai-autonomous-oxygen-titration-jama-85pct-vs-63pct-normoxemia-2026/ Primary record: https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2852401
  16. TIN case file, John Deere See & Spray. https://theinternetninja.com/stories/john-deere-see-spray-ai-spot-spraying-cuts-herbicide-use-43-59-independently-con/ Primary record: https://www.cambridge.org/core/journals/weed-technology/article/comparing-herbicide-application-methods-with-see-spray-technology-in-soybean/05967C11D1BF0B174A13BE687B865AF2, https://aaes.uada.edu/news/see-and-spray-research/ and https://crops.extension.iastate.edu/cropnews/2024/08/precision-spraying-technology
  17. TIN case file, SEC and Presto Automation. https://theinternetninja.com/stories/sec-ai-washing-presto-automation-voice-ai-drive-thru-human-agents-cease-and-desist-2025/ Primary record: https://www.sec.gov/files/litigation/admin/2025/33-11352.pdf
  18. TIN case file, DOJ and SEC, Nate. https://theinternetninja.com/stories/doj-sec-nate-saniger-ai-shopping-app-effectively-zero-percent-automation-40m-securities-wire-fraud-2025/ Primary record: https://www.justice.gov/usao-sdny/pr/tech-ceo-charged-artificial-intelligence-investment-fraud-scheme and https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26282
  19. TIN case file, Moffatt v. Air Canada. https://theinternetninja.com/stories/air-canada-chatbot-bereavement-misinfo-tribunal-orders-812-cad-negligent-misrepresentation-2024/ Primary record: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html
  20. TIN, AI in operations topic hub. https://theinternetninja.com/guides/ai-in-operations/
  21. TIN, the case file registry. https://theinternetninja.com/stories/
  22. TIN, “Agentic AI customer service”. https://theinternetninja.com/blog/agentic-ai-customer-service-what-the-verified-record-shows-2026/
  23. TIN, “AI in banking customer service”. https://theinternetninja.com/blog/ai-in-banking-customer-service-what-the-record-shows-2026/
  24. TIN, “Klarna AI layoffs: the 2025 reversal explained”. https://theinternetninja.com/blog/klarna-ai-layoffs-the-2025-reversal-explained-2026/
  25. TIN, “Does AI customer service cut jobs? The Octopus Energy case”. https://theinternetninja.com/blog/does-ai-customer-service-cut-jobs-the-octopus-energy-case-2026/
  26. TIN, “AI coding agent”. https://theinternetninja.com/blog/ai-coding-agent-what-the-measured-record-shows-2026/
  27. TIN, “Agent assist”. https://theinternetninja.com/blog/agent-assist-what-the-record-shows-when-ai-helps-the-human-2026/
  28. TIN, “AI workflow automation”. https://theinternetninja.com/blog/ai-workflow-automation-what-the-documented-record-shows-2026/
  29. TIN, “AI back office automation”. https://theinternetninja.com/blog/ai-back-office-automation-what-the-numbers-show-2026/
  30. TIN, “Supply chain automation”. https://theinternetninja.com/blog/supply-chain-automation-what-the-documented-record-shows-2026/
  31. TIN, “AI agent fraud”. https://theinternetninja.com/blog/ai-agent-fraud-what-the-adjudicated-record-shows-2026/

Questions

What are some real-world examples of AI agents?

Klarna's customer-service assistant had 2.3 million conversations, two-thirds of its customer-service chats, in its first month, by Klarna's own figures. Otto's demand-forecasting system bought around 200,000 items a month with no human intervention, as The Economist reported in 2017. At Boston Children's Hospital, an OpenAI reasoning model reanalysed 376 unsolved rare-disease cases and physicians established diagnoses in 18, in a peer-reviewed study that OpenAI-affiliated authors co-wrote. The page's table holds 19 rows, each with who measured its figure: 17 deployments and two cases, Air Canada and Nate, where the record turns on what the system did or was claimed to do.

Is ChatGPT an AI agent?

Under this page's working definition, which is TIN's, an assistant counts when it carries out or drafts work for someone. Moderna rolled out ChatGPT Enterprise company-wide, and what it reported is usage, from OpenAI's own case study, not an outcome.

What are the top AI agents?

This page does not rank agent products. It is organised by deployment, and the useful question is which deployments published a result and who measured it.

What are the types of AI agents?

Textbook taxonomies sort agents by how they are built. This page groups them by what they do on the job: agents that talk to customers, assistants that work beside employees, and systems that act on their own in operations, plus the cases where the agent turned out to be people. That is what the deploying companies report against.

Which AI agent results were measured by someone other than the company?

Of the 19 rows, 10 were measured by the company or its vendor, five by researchers (GitHub Copilot, Tutor CoPilot, Boston Children's, the SAVE-O2 trial and John Deere's See & Spray), three by a regulator, prosecutor or tribunal (Presto, Nate and Air Canada), and one by independent press (Otto, with figures sourced from Otto and Blue Yonder). It is TIN's inference from these case files, not a measured finding, that outside the companies and their vendors the outcome figures in this record come from trials and university studies, while the production figures for agents that talk to customers are the companies' own.

method and author
Written by
TIN research desk. No single byline: the brand is the work.
Data used
19 case files, each linked where it is used.
Rules
Every figure carries its source link, and nothing is rounded or invented. How TIN works →
Updated
2026-10-16
next

Where to go from here