# What leading companies tried, what they learned, and our reflections

**AI in client communication: lessons learned**

This page collects published articles on situations where companies used AI agents to automate too much of their written communication with potential clients, existing clients and consumers, and what came back. It also includes studies of where AI in writing works well. Each finding comes with its source and our take.

Part of the guide: [Speech to text vs. AI automation vs. AI writing assistants](https://www.writing-ant.com/writing-ant/speech-to-text-vs-ai-automation-vs-ai-writing-assistants)

1. **You speak, AI writes it down** (Speech to text): You say what you want to write. A speech-to-text tool such as WritingAnt turns it into well-written text instantly or within a few seconds. The content and the judgment stay yours.
2. **AI writes by itself** (Agentic AI workflow): A workflow or AI agent writes and sends the text from data and rules. A person approves the draft, if anyone does.
3. **You and an AI iterate on one text** (AI chat or writing assistant): You refine one text with an AI chat or writing assistant over several rounds, mixed with editing on the keyboard.

Source: https://www.writing-ant.com/writing-ant/ai-in-client-communication-lessons-learned  
Published: 2026-09-30  
Last updated: 2026-09-30

**Our reflection (WritingAnt):** It pays to be deliberate about which parts of internal and external communication are automated with AI agents, and to keep monitoring and evaluating that choice. Where judgment matters, and judgment is still a key difference between people and AI, communication is better served when a person is in the driver’s seat of the content.

For that part, we recommend a modern, AI-powered speech-to-text tool — such as WritingAnt — that automatically polishes what a person has spoken. It combines human judgment and proofreading with the speed of AI, usually around three times faster than the keyboard, and keeps the AI risk far lower than when AI agents draft content from historical data, context and rules, because a person drives the content by speaking it.

### Fortune: MIT report: 95% of generative AI pilots at companies are failing (2025)

An MIT study of generative AI in companies found that about 95% of pilot projects showed no measurable effect on profit and loss, and that purchased tools succeeded far more often than systems built in-house. The study defines success narrowly and has been criticized for it.

Source: [Fortune: MIT report: 95% of generative AI pilots at companies are failing](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* We believe speech to text with AI-powered editing is probably the best return on investment a company can get from AI in internal and external communication: the person drives the content by speaking, with their own judgment, and proofreads the message before it is sent. AI does the detailed writing, around three times faster than a person would, and likely with better quality and a more consistent writing style. Consider starting with a ready-made tool that fits into the writing people already do, in the apps they already use, before a large integration project. A speech-to-text tool such as WritingAnt is one example. It is free to download and start using, so a team can judge the benefit and the fit for itself from the first day. The app keeps track of the time saved by speaking instead of typing, month by month, and an hourly value can be added, so every team member sees what their saved time is worth. Paid plans are usage-based tiers, so the cost stays small next to a value that can be measured. The included writing styles work well out of the box for most companies. For a company’s own tone of voice, the styles learn from the corrections people make as they write, with no significant extra effort, and carry what they learned into the next texts.

### Entrepreneur: Klarna CEO Reverses Course By Hiring More Humans, Not AI (2025)

Klarna announced in 2024 that its AI assistant did the work of 700 customer service agents. In May 2025 its CEO said the company was hiring people for customer service again, and that customers should always be able to reach a person.

Source: [Entrepreneur: Klarna CEO Reverses Course By Hiring More Humans, Not AI](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* Consider keeping people in customer service for everything beyond routine questions, and making those people faster rather than replacing them. With a speech-to-text tool such as WritingAnt, the agent still decides and speaks the answer; AI only turns the spoken words into well-written text, in about a third of the time typing takes.

### 20VC with Harry Stebbings: The Untold Story of Higgsfield, CEO Alex Mashrabov (YouTube; B2B support from 36:10) (2026)

Alex Mashrabov, founder and CEO of the AI video company Higgsfield, said in September 2026 that probably over 60% of customer support requests, especially first-line ones, can be handled with AI, but that in B2B support AI automation just doesn’t work. Yet the company spends more than 10,000 dollars per employee per month on AI models, on average across its roles.

Source: [20VC with Harry Stebbings: The Untold Story of Higgsfield, CEO Alex Mashrabov (YouTube; B2B support from 36:10)](https://www.youtube.com/watch?v=jszn8rFtxm4), 2026

*Our take (WritingAnt; our interpretation, not part of the source):* When one of the fastest-growing AI companies of 2026, one that knows the value of AI and applies it skilfully to its own core business, still deliberately keeps people in B2B client communication, and does not fully automate consumer support either, that is a lesson many companies can follow. We believe AI-powered speech to text, such as WritingAnt, fits B2B support particularly well, because it keeps the person in the driver’s seat: they drive the content by speaking and proofread it before it is sent. That is around three times faster than a keyboard, while dramatically reducing the risks that come with AI-generated content and heavier automation.

### a16z: Narrative Violation: In B2B customer support, AI is a Copilot, Not a Replacement (2026)

Andreessen Horowitz reported that about 15% of B2B support requests are resolved end to end by AI, against about 35% in B2C. Two-thirds of the time AI quietly triages tickets to people, and hard tickets took 5.3 messages with AI triage against 3.9 when handled by people alone.

Source: [a16z: Narrative Violation: In B2B customer support, AI is a Copilot, Not a Replacement](https://www.a16z.news/p/narrative-violation-in-b2b-customer), 2026

*Our take (WritingAnt; our interpretation, not part of the source):* Consider AI as a copilot for the support person, not a replacement. A speech-to-text tool such as WritingAnt is a copilot of that kind for writing: the person stays in the conversation, and AI speeds up and polishes the reply.

### NBER Working Paper 34255: How People Use ChatGPT (Chatterji, Cunningham, Deming and others) (2025)

A study by OpenAI’s economic research team and the Harvard economist David Deming, based on more than a million ChatGPT conversations, found that writing is the most common use of ChatGPT at work: about 40% of work-related messages in June 2025. About two-thirds of those writing messages ask ChatGPT to modify text the user already wrote (editing, critiquing, translating) rather than to create new text from scratch.

Source: [NBER Working Paper 34255: How People Use ChatGPT (Chatterji, Cunningham, Deming and others)](https://www.nber.org/papers/w34255), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* This is where AI in writing is most mature: improving a text whose content comes from a person. WritingAnt uses GPT-6, the model people already trust to edit their text. You get it built closer into your workflow and with a lower threshold: you speak, and automatically edited text appears in any text field, in any application on your computer. Messy text you have already typed can also be fixed by WritingAnt. Highlight it, press a simple keyboard command, and it is replaced with an improved version, without leaving the app you are writing in.

### Noy & Zhang, Science: Experimental evidence on the productivity effects of generative artificial intelligence (2023)

In an experiment published in Science, 453 college-educated professionals did writing tasks typical of their jobs, 20 to 30 minutes each; half of them were given ChatGPT. With ChatGPT, the average time taken fell by 40% and the rated quality of the output rose by 18%.

Source: [Noy & Zhang, Science: Experimental evidence on the productivity effects of generative artificial intelligence](https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf), 2023

*Our take (WritingAnt; our interpretation, not part of the source):* AI help with writing is well proven where a person is responsible for the text and sees it before it goes out. Consider that setup for daily communication. With a speech-to-text tool such as WritingAnt, the person speaks the content and proofreads the AI-polished result, around three times faster than typing.

### Gartner: 64% of Customers Would Prefer That Companies Didn’t Use AI For Customer Service (2024)

64% of customers would prefer that companies did not use AI in customer service, and 53% would consider switching to a competitor because of it. Their top concern: that it becomes harder to reach a person. Survey of 5,728 customers.

Source: [Gartner: 64% of Customers Would Prefer That Companies Didn’t Use AI For Customer Service](https://www.gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Make sure customers can always reach a person, and consider using AI behind that person rather than in front of them. With a speech-to-text tool such as WritingAnt, AI works for the agent: it turns a spoken answer into a fast, well-written reply.

### Gartner: 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI (2025)

Many companies planned to cut their customer service teams sharply and let AI take over. Gartner expects that by 2027, half of them will have dropped that plan.

Source: [Gartner: 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI](https://www.gartner.com/en/newsroom/press-releases/2025-06-10-gartner-predicts-50-percent-of-organizations-will-abandon-plans-to-reduce-customer-service-workforce-due-to-ai), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* If the team stays, its time is what counts. Consider using AI to speed up and polish what each person writes, for example with a speech-to-text tool such as WritingAnt, on exactly the complex and emotional cases that agentic AI workflows do not take over.

### Hunter: The State of Cold Email (2026)

69% of decision makers are uncomfortable with a cold email that is obviously written by AI. The average reply rate was 4.5%, and small, targeted recipient lists performed best. Based on emails sent through Hunter’s own platform and a survey.

Source: [Hunter: The State of Cold Email](https://hunter.io/the-state-of-cold-email), 2026

*Our take (WritingAnt; our interpretation, not part of the source):* Consider smaller lists and messages that the seller decides and speaks personally, so outreach stays personal. A speech-to-text tool such as WritingAnt makes that practical: speaking an individual message to each prospect takes about a third of the time typing it would, and AI only tidies the wording.

### TechCrunch: a16z- and Benchmark-backed 11x has been claiming customers it doesn’t have (2025)

ZoomInfo, the US sales-data company, ran a one-month pilot of an AI sales development agent from the startup 11x in early 2025. ZoomInfo said the product “performed significantly worse than our SDR employees”, and it did not move forward. A former 11x employee told TechCrunch the startup was losing 70–80% of the customers that came through the door; 11x put its retention rate at 79%.

Source: [TechCrunch: a16z- and Benchmark-backed 11x has been claiming customers it doesn’t have](https://techcrunch.com/2025/03/24/a16z-and-benchmark-backed-11x-has-been-claiming-customers-it-doesnt-have), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* Consider testing AI outreach agents against your own people in a short pilot before replacing anyone. Where a person’s messages perform better, a speech-to-text tool such as WritingAnt makes that person faster: the rep speaks each message with their own judgment, and AI turns it into well-written text.

### Gartner: By 2030, 75% of B2B Buyers Will Prefer Sales Experiences that Prioritize Human Interaction Over AI (2025)

Gartner predicts that by 2030, 75% of B2B buyers will prefer sales experiences that prioritize human interaction over AI. After several years of growing interest in self-service and AI-driven sales, buyers want authentic human engagement again, especially in complex or high-stakes purchases. This is a forecast, not a measured result.

Source: [Gartner: By 2030, 75% of B2B Buyers Will Prefer Sales Experiences that Prioritize Human Interaction Over AI](https://www.gartner.com/en/newsroom/press-releases/2025-08-25-gartner-says-by-2030-that-75-percent-of-b2b-buyers-will-prefer-sales-experiences-that-prioritize-human-interaction-over-ai), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* Consider where in the sales conversation a person should be the one who writes: follow-ups, replies and offers in complex deals. With a speech-to-text tool such as WritingAnt, the rep speaks those messages personally, around three times faster than typing, and AI only polishes the wording.

### Harvard Business Review: AI-Generated “Workslop” Is Destroying Productivity (2025)

40% of employees said they had received “workslop” in the past month: AI-generated work that looks finished but lacks the substance to move a task forward. Survey of 1,150 full-time employees in the US.

Source: [Harvard Business Review: AI-Generated “Workslop” Is Destroying Productivity](https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* Consider letting AI shape the wording, not the substance. With a speech-to-text tool such as WritingAnt, the content is the person’s own thinking, spoken, so what reaches the reader has substance.

### McCarthy Tétrault: Moffatt v. Air Canada (2024)

A Canadian tribunal held Air Canada liable for wrong information that the chatbot on its website gave a customer. The argument that the chatbot was responsible for its own words was rejected.

Source: [McCarthy Tétrault: Moffatt v. Air Canada](https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Consider which answers should never go out unseen. Where the company has to stand behind the content, a person can speak the answer and read the AI-polished text before it is sent, for example with a speech-to-text tool such as WritingAnt, so the company knows what goes out in its name.

### Bitkom: Kundenservice beim Online-Shopping: Mensch schlägt Chatbot (2025)

In a Bitkom survey of 1,006 internet users in Germany, 62% of online shoppers wanted a human contact they can reach quickly when there is a problem; only 36% wanted a chatbot. 86% were satisfied with human contact, 50% with chatbots.

Source: [Bitkom: Kundenservice beim Online-Shopping: Mensch schlägt Chatbot](https://www.bitkom.org/Presse/Presseinformation/Kundenservice-Online-Shopping-Mensch-Chatbot), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* Consider giving customers the person they ask for when something goes wrong, without slowing the team down: with a speech-to-text tool such as WritingAnt, the agent speaks the reply and AI turns it into well-written text.

### Ipsos BVA: Observatoire des services clients 2025 : l’humain (toujours) au cœur de la relation client à l’ère de l’IA (2025)

In the 2025 Ipsos BVA customer service observatory, 90% of French respondents said they prefer to wait for a human advisor, and 64% refused even an advisor assisted by AI. 31% used a chatbot, and only half of them were satisfied. Representative sample of 5,000 adults in France, the UK, Germany, Italy and Spain.

Source: [Ipsos BVA: Observatoire des services clients 2025 : l’humain (toujours) au cœur de la relation client à l’ère de l’IA](https://www.ipsos.com/fr-fr/observatoire-des-services-clients-2025-lhumain-toujours-au-coeur-de-la-relation-client-lere-de-lia), 2025

*Our take (WritingAnt; our interpretation, not part of the source):* A caution that applies to speech-to-text tools such as WritingAnt too: the answer has to come from the person. Consider AI only where the advisor still decides and speaks every word, AI only polishes the wording, and the customer is still served by a person.

### Computer Sweden: AI i all ära – många svenska kunder vill fortfarande prata med en människa (2024)

In ServiceNow’s 2024 consumer study, run by Opinium, 59% of Swedes wanted customer service to move towards more human contact and less automation. Of those who avoid AI services, 35% did not want AI to replace human contact and 32% doubted the credibility of AI answers. 1,000 respondents in Sweden, 15,000 in ten countries.

Source: [Computer Sweden: AI i all ära – många svenska kunder vill fortfarande prata med en människa](https://computersweden.se/article/2086449/ai-i-all-ara-manga-svenska-kunder-vill-fortfarande-prata-med-en-manniska.html), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Consider letting a person who knows the case give the answer, so the customer gets human contact and a credible reply. A speech-to-text tool such as WritingAnt only speeds up the writing.

### Hochschule Luzern, IFZ Retail Banking Blog: Kundenservice bei 61 Schweizer Banken 2026 (2026)

A Lucerne University of Applied Sciences review of customer service at 61 Swiss banks found that the AI chatbot of the digital bank Yuh handles about 45% of inquiries and works well for standard questions on features, prices and general topics, while about 45% of contacts still come by phone. Customers with problems prefer to call and speak to a person.

Source: [Hochschule Luzern, IFZ Retail Banking Blog: Kundenservice bei 61 Schweizer Banken 2026](https://hub.hslu.ch/retailbanking/kundenservice-bei-61-schweizer-banken-2026-zwischen-automatisierung-und-persoenlicher-beratung/), 2026

*Our take (WritingAnt; our interpretation, not part of the source):* Consider automating the routine half and investing in the other half: the written replies to problem cases. A person decides and speaks them, and a speech-to-text tool such as WritingAnt polishes the text.

### Netzwoche: Schweizer Bevölkerung lässt sich lieber von Chatbots als von Menschen helfen (comparis.ch survey) (2024)

comparis.ch surveys found that 72% of the Swiss population would rather use a chatbot to find information on a website faster (56% in 2021), and 72% to check a delivery status. For mental-health matters fewer than 15% would, and for physical-health matters 21%.

Source: [Netzwoche: Schweizer Bevölkerung lässt sich lieber von Chatbots als von Menschen helfen (comparis.ch survey)](https://www.netzwoche.ch/news/2024-04-16/schweizer-bevoelkerung-laesst-sich-lieber-von-chatbots-als-von-menschen-helfen), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Consider highly automated, agentic workflows for simple information, and for personal matters a person who writes, made faster by a speech-to-text tool such as WritingAnt.

### EU AI Act, Article 50: Transparency obligations (2024)

From 2 August 2026, the EU AI Act requires AI systems that interact directly with people to tell them they are interacting with AI, unless that is obvious. AI-generated text published to inform the public must be disclosed as such, except where it has undergone human review or editorial control and a person holds editorial responsibility.

Source: [EU AI Act, Article 50: Transparency obligations](https://artificialintelligenceact.eu/article/50/), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Consider staying on the human side of that line: the person speaks the content, reads the AI-polished text and is responsible for what is sent. That is how a speech-to-text tool such as WritingAnt works.

### EUR-Lex: Directive (EU) 2023/2673 on financial services contracts concluded at a distance (2023)

Under the EU directive on financial services contracts concluded at a distance, applied from 19 June 2026, consumers dealing with online tools such as chatbots or robo-advice have a right to request and obtain human intervention before the contract, and in justified cases after it.

Source: [EUR-Lex: Directive (EU) 2023/2673 on financial services contracts concluded at a distance](https://eur-lex.europa.eu/eli/dir/2023/2673/oj/eng), 2023

*Our take (WritingAnt; our interpretation, not part of the source):* When the customer asks for a person, that person has to answer well and fast. Consider a speech-to-text tool such as WritingAnt, which lets the advisor speak the answer and send well-written text.

### Börsen-Zeitung: Trade Republic stellt Kundenservice neu auf (2026)

Trade Republic said its AI-driven dialogue systems reached their limits during heavy trading, and invested an amount in the tens of millions of euros in more than 1,000 trained service agents, reachable by phone, live chat and callback around the clock in eight languages. Complaints to German consumer centres, mainly about not reaching customer service, had more than doubled.

Source: [Börsen-Zeitung: Trade Republic stellt Kundenservice neu auf](https://www.boersen-zeitung.de/banken-finanzen/trade-republic-hat-den-kundenservice-komplett-neu-aufgestellt), 2026

*Our take (WritingAnt; our interpretation, not part of the source):* When many people answer customers, their writing speed decides the waiting time. Consider a speech-to-text tool such as WritingAnt, which lets each agent speak the reply and send well-written text in about a third of the typing time.

### TIME: AI Chatbot Curses at Customer and Criticizes Company (2024)

After a system update in January 2024, the customer service chatbot of the parcel company DPD swore at a customer and called DPD the worst delivery firm in the world. DPD disabled the AI element immediately.

Source: [TIME: AI Chatbot Curses at Customer and Criticizes Company](https://time.com/6564726/ai-chatbot-dpd-curses-criticizes-company/), 2024

*Our take (WritingAnt; our interpretation, not part of the source):* Consider having every reply spoken by a person and read before it is sent, so nothing goes out in the company’s name without someone seeing it. A speech-to-text tool such as WritingAnt keeps that fast.

The findings show one pattern. Automation works for information. Where a relationship, a judgment or a liability is involved, the text should come from a person who stands behind it, and that person should be able to write it fast.

The full guide: speech to text vs. AI automation vs. AI writing assistants: https://www.writing-ant.com/writing-ant/speech-to-text-vs-ai-automation-vs-ai-writing-assistants