What is the best way to use AI for each kind of writing? Speech to text vs. AI automation vs. AI writing assistants
Published
AI can help with business writing in four main categories, and each pays off in a different kind of work. This guide shows the situations where speech-to-text tools such as WritingAnt work well, and the situations that other AI solutions are better suited to.
- Speech to text
You speak, AI writes it down
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.
- Agentic AI workflow
AI writes by itself
A workflow or AI agent writes and sends the text from data and rules. A person approves the draft, if anyone does.
- AI chat or writing assistant
You and an AI iterate on one text
You refine one text with an AI chat or writing assistant over several rounds, mixed with editing on the keyboard.
The three approaches at a glance
| Criterion | Speech to textModern AI-powered speech to text, driven by a personsuch asWritingAnt | Agentic AI workflowHighly automated, agentic workflows with AI-generated drafts | AI chat or writing assistantAI chats and writing assistants, in several rounds (mixed with keyboard-driven editing) |
|---|---|---|---|
| How it works | The writer says what they want to write. AI transcribes it and turns it into well-written text in the chosen style, right at the cursor: in a text field in any app on their computer, without leaving the app they are working in. | A workflow or AI agent writes and sends the text from data and rules. A person approves the drafts, or nobody does. | The writer gives an AI chat or writing assistant a prompt or a draft, reads the result and refines it over several rounds. |
| Who is the author | The person. Content, judgment and decisions are the writer’s own; AI handles the wording. The reader sees a person behind the text. How it reads reflects on the person who sent it, and on the company that person works for. | The AI. A person sets up the workflow and may check the output. The reader does not expect a person who takes pride in the message and responsibility for it. Where automation makes sense, a mistake is easily understood and forgiven as a slip in an automated chain. | Shared. The AI proposes content and structure; the person steers, chooses, and does more or less of the wording and editing themselves. The reader sees a person or a company behind the text, and expects the highest level of quality and correctness. |
| Best for | Daily internal and external communication where the right content depends on the writer’s judgment, coordination, timing and experience of complex topics. Especially the internal coordination behind an answer to a client request, and the answer itself. Typically business-critical, operational topics. Useful when a high degree of personalization is critical for the best business outcome. | Repetitive messages at high volume that carry information and follow rules. Personalization is less critical here: a medium level of personalization, which can be automated, is enough for a good business outcome. | A few important texts that are worth hours of work: long, crafted, published. The text itself is usually the product: it is sold in some form, or it is marketing copy for a wide or highly important audience. |
| Typical texts | Emails and chat messages to clients and colleagues, answers to client requests and the internal coordination behind them, follow-ups, feedback, status updates, ticket replies, and prompts for AI tools. | Order and appointment confirmations, reminders, status notifications, answers to simple frequent questions, first replies that route a request. Outreach, where it suits the product or service and a lower level of personalization is acceptable. | Books, news articles, white papers, long reports, tenders, marketing campaigns and web copy. |
| Strength | Around three times faster than typing, and the text says what the writer meant, because the writer said it. Personal by nature: every message is written for one reader. | Costs almost nothing per message and answers within seconds, at any hour. | Depth: research, structure, alternatives and several rounds of revision on the same text. |
| Weakness | Needs a person for every text, so it does not scale to thousands of messages a day. Not the tool for rounds of revision on a long document. | Generic text that readers recognize and ignore. Mistakes are repeated at scale, and nobody who fully knows the client is likely to have read the message, or truly cared about it, before it goes out. | Too slow for daily messages: copy the text over, write a prompt, wait, read, paste it back. The AI supplies content the writer did not write, so the result needs careful checking. |
| Effort to introduce | Minutes. Install, pick a writing style, press one key and speak. | Weeks to months: data connections, rules, testing, monitoring, and a way to hand over to a person. | Minutes to start. The skill lies in prompting and reviewing. |
| Accountability | Clear. The writer is the author and reads the text before sending it. | With the company, also for what the AI got wrong. Approval steps tend to become a formality when the volume is high. | With the person who publishes. Facts, quotes and sources from the AI have to be verified. |
| Expected return on investment | Time saved per person, from the first day, on writing that already takes place: usually 5–40 hours a month per user. A simple rule of thumb: speaking is roughly three times faster than typing. The biggest factors for a good return are management following up on adoption, since it is a new way of working, and how much of the working week people can actually speak instead of type. That depends on the environment: it is easier when working remotely than in an open-plan office, and easier where meeting rooms or phone booths are free to step into for a batch of emails. Low cost, no integration project. | Time saved and fewer FTEs (full-time equivalents) through reuse, weighed against the overall indirect cost per message. The project, the supervision and the damage of wrong or unwanted messages count against it. | Quality and reach of a few important texts, and time saved on research and drafting. |
| KPIs to monitor | Time saved, and how often people use it. WritingAnt measures the time saved for each user per day, week and month, and its value at an hourly rate you choose, so the business case stays visible and easy to follow up. Quality: occasional spot checks of the texts show that people read them through before sending. The risk of mistakes stays low because a person creates every text and can read it through, which makes this one of the lowest-risk ways to get a return from AI. | Time and money saved, and the return on investment after ongoing upkeep: the business and its underlying data change over time, and the automated answers have to be kept up to date with them, which takes continuous investment. Whether the expected outcomes are reached, such as resolved requests, conversions and customer satisfaction, and early signs of damage to goodwill and reputation. As AI has become far more powerful and easier to use over the last few years, people have been flooded with automated sales outreach and automated support replies. For some organizations the result was lost goodwill, missed deals and damaged client relationships (see “What companies tried” below). Whether automation fits, and in which situations, deserves careful analysis. | Not critical. Business success is usually measured the same way as when professional writers did this work before the AI boom. Beyond that, keep track of whether users are happy with the quality of the models they have access to, and budget through corporate accounts with clear monthly costs or allowances for AI chats and writing assistants. No bigger investment in monitoring is typically needed. |
What companies tried, what they learned, and our reflections
Many companies automated more of their client communication with AI agents than their clients accepted. Published findings, each with its source, show what came back.
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.
Return on investment: where AI pays off in writing
The return on an AI investment in writing depends more on the match between the approach and the kind of text than on the technology.
Speech to text has the shortest way to a return. People speak at around 150 words per minute and type at around 50. Someone who writes for two hours a day gets a large part of that time back from the first day, without an integration project, and the AI polishing raises the quality of the text at the same time. Typical WritingAnt users save around 20 hours per month.
Highly automated, agentic workflows have the highest return per message and the highest cost to get there: integration, testing, monitoring, and the cost of mistakes. They pay off where the volume is high and the content is simple. They lose money where a wrong or impersonal message costs a client.
AI chats and writing assistants pay off on a few important texts. Their return is quality and reach; per text they take the most time. Most organizations need all three approaches, each in its place. The use-case pages on this site include a calculator for the time saved, with your own numbers.
Voice prompting: speech to text for AI tools
Every role that works with AI chats, coding assistants or AI agents writes prompts, and the quality of the answer depends on the context in the prompt. People give more context when they speak: the background, the goal, the exceptions, and what was already tried.
This is where speech to text meets the other two approaches. The person who works on a long text with an AI writing assistant, or who sets up an automated workflow, also instructs the AI in words. Speaking is the fastest way to do that.
WritingAnt works at the cursor in every AI tool, with one key. On a modern computer the local Basic style usually returns the transcript in under 200 milliseconds, a recording can run for up to 20 minutes without stopping by itself, and the voice recording never leaves the computer.
How to decide, text by text
Five questions settle most cases.
- Does the content depend on your judgment, or on what you know about this reader? Then speak it yourself.
- Would you be uncomfortable if the reader learned that nobody had read the message before it was sent? Then do not automate it.
- Is it the same information for everyone, at high volume? Then automate it, and keep an easy way to reach a person.
- Is it one long or published text that is worth hours of work? Then use an AI chat or writing assistant over several rounds, and speak the first draft.
- Are you instructing an AI tool? Then speak the prompt. It will contain more context than the one you would have typed.