We Let the Agent Write the Blog Post

TL;DR. We shipped autonomous blog writing in 2 modes at one price: interview mode asks you 5 questions, autonomous mode runs without you. Both are grounded in your audited pages and 1 open opportunity. The agent is blocked from 7 claim types, thin research fails the run, and publishing still needs you.

We resisted shipping a write-the-whole-post button for a long time. Not because it was hard to build, but because the first version we built was very good at producing something that looked exactly like a useful article and contained nothing that was true. Autonomous mode shipped only once we had a list of things the agent is not allowed to do, and that list turned out to be the actual product.

What did we actually ship in the Blog Creator?

Two modes at the same price. "Interview me" asks you 5 questions and writes from your answers. "Write it for me" runs without you. Both are schedulable, both are grounded in your own audited pages, and neither costs more than the other.

Interview mode came first and it works because the questions are small. Five prompts, sized so a founder can answer them on a phone between meetings: what happened, what you think about it, what you would tell someone in your position, what you would not do again, and what you want the reader to do next. The writing agent handles the structure, the question headings, the schema and the internal linking. You supply the part that cannot be sourced anywhere else, which is what you think.

Autonomous mode is the same writer with the interview replaced by evidence the agent can gather on its own. That sounds like a small substitution and it is not. Removing the founder from the loop removes the only input that was guaranteed to be true, so every other input had to get much stricter to compensate. The full workflow, including scheduling, is written up in our guide to the autonomous blog workflow.

What is autonomous mode grounded in?

Four inputs, all of them yours. Your audited pages, which are the only URLs the post is permitted to link to. Your measured standing across the assistants. Live topic research. And 1 open opportunity that the post exists to close.

The audited-pages constraint does the heaviest lifting. A model that is allowed to write a link will write a link, and it will write one that reads perfectly and returns a 404, because a plausible URL is the easiest thing in the world to generate. Restricting the link set to pages we have actually crawled and scored means every internal link in the draft resolves, and every claim the post makes about your own product can be traced back to a page that exists and that you wrote. It also has a side effect we did not anticipate: the constraint makes the agent notice when you have no page to link to, which is usually a more interesting finding than the post itself.

Your measured standing keeps the tone honest. An agent that does not know how you are doing will default to the house style of the internet and open with a sentence about being the leading platform in your space. An agent that can see you are named occasionally in your category, and rarely recommended, writes something a reader will recognize as true. It is difficult to overstate how much of the cringe in AI-written marketing copy comes from a model guessing at a company's stature and guessing high.

The fourth input is the one people skip. Every autonomous run is attached to a single open opportunity from your findings, and the post's job is to close it. That gives the draft a success condition before it is written. Content that exists because it was Tuesday cannot be evaluated. Content that exists to close a specific gap can be checked against that gap later, which is the only reason to publish it in the first place.

What is the agent forbidden from writing?

Seven categories, blocked outright: percentages, study results, survey findings, quotes, named customers, case-study outcomes and dates. Every one of them is a claim a reader will trust on sight and a language model will produce on demand.

Almost every generic AI article you have read contains a sentence of the shape "studies show that most buyers now do X". There is no study. The number and the citation-flavored phrasing are a stylistic tic the model learned from a corpus in which confident writing has numbers in it. It is not lying in any meaningful sense, it is imitating the texture of authority, and the result is indistinguishable from a fabricated statistic because that is precisely what it is.

Quotes and named customers are the ones that can actually hurt you. A fabricated percentage embarrasses you. An invented customer testimonial, published under your logo, is a different category of problem entirely, and an agent with no concept of legal exposure will write one cheerfully if the paragraph feels like it needs social proof. Dates go on the list for a quieter reason: a model will confidently place your funding round, your launch or an industry event in the wrong year, and that error propagates into the AI answers that read your page afterwards.

The same discipline runs elsewhere in the platform. The scan review is forbidden from writing a number at all: the server materializes every figure and a single unbacked number invalidates the whole summary. Traffic estimates on content ideas are always a range and never a point estimate, because a point estimate is a promise we cannot keep. We wrote about the general approach in how we stop our AI from inventing numbers.

Why does thin research fail the run instead of shipping something generic?

Because a generic article is worse than no article. If live topic research comes back thin, the run stops at that step and tells you, rather than filling the gap with the plausible middle of the internet. Credits are refunded when a run fails.

This is the behavior we argued about most and the one we are most confident about now. A language model asked to produce fifteen hundred words with nothing to say will produce fifteen hundred words. It will not stall, it will not complain, and it will not signal that it had nothing. The output has the same headings, the same rhythm and the same length as a good post. That is the whole problem with generative writing tools: the failure mode is invisible at exactly the moment you are least likely to be reading carefully, which is when you asked a machine to do the reading for you.

So we made the tool refuse. If the research step cannot assemble enough grounded material on the topic, the run does not degrade gracefully into filler, it stops. Founders occasionally find this annoying, and we understand why, because a failed run feels like a broken product. Our position is that a tool which tells you the topic is not ready has given you real information, and a tool which hands you a competent-looking article about nothing has taken something from you and charged you for it.

The refusal also protects the thing the post was supposed to help. Publishing generic content to improve how AI assistants describe you is actively counterproductive, because assistants are increasingly good at recognizing pages that restate the category consensus and add nothing. Volume was a viable strategy in search. In answers it is closer to noise you have to maintain.

Which mode should you pick, interview or autonomous?

Interview mode sounds like you. Autonomous mode sounds like your brand. If the post depends on a story only you were in the room for, answer the 5 questions. If the topic is one your category needs covered, let the agent run.

We should be blunt about the ceiling, because most people selling autonomous content are not. Autonomous mode cannot write the sentence you would have written after a bad customer call. It has no opinions it did not source from your own pages, it has no grudges, and it will never tell a reader that something widely recommended in your industry is a waste of money. Those sentences are the ones people remember and share, and no amount of grounding produces them. This post, for instance, is an interview-mode post, because the argument in it is an argument we had.

What autonomous mode is genuinely good at is the work that is not blocked on your best story. Coverage of the questions your buyers ask that you have simply never written a page for. The comparison your category expects to exist. The explainer that every competitor has and you do not. Those posts do not need a voice, they need to exist, be accurate, and answer the question in the first paragraph. Most content calendars are not stalled for lack of insight. They are stalled because the founder is the bottleneck on every single item.

Our own rule of thumb: if you would be disappointed to see the post published without your name on it, use interview mode. If you would be fine with it appearing under the company name, let the agent write it and spend your time on the review instead of the draft.

Why do both modes cost the same?

Because pricing autonomy at a premium would push people toward the mode that produces the worse post for their situation. Same price, either way, with 100 free credits to try both, and every run shows its exact price before it starts.

We have watched enough tools do the opposite to be sure about this. Charge more for the automatic mode and you turn a craft decision into a budget decision. Founders on a tight month pick the cheap mode regardless of which one suits the post, and the tool quietly starts producing the wrong kind of content for the wrong reasons. Charge the same and the only question left is the honest one: does this post need you in it?

Scheduling makes the choice matter more, not less. On Growth and Business plans the agent workforce runs on a schedule, so autonomous posts can be produced without anybody opening the app that week. When credits run out, paid schedules pause and resume automatically rather than failing, and pack credits roll over and never expire. Full plan detail sits on our pricing page.

What does an autonomous draft still need from you?

A read. The agent researches, drafts, structures and links without you, and it stops at publish. High-impact actions route back to the founder for approval, which is the same rule every agent in the platform follows: it works autonomously, you stay in control.

In practice the review is short, because the constraints have already removed the failure modes that take the longest to check. You are not verifying statistics, because there are none the server did not produce. You are not checking whether links resolve, because they can only point at pages we crawled. What you are reading for is judgment: is this the argument we want to make, is the framing right, and is there a sentence only you could add. That last one takes two minutes and it is usually the difference between a post that is correct and a post that is worth reading.

If you want to see the shape of it on your own brand, run free discovery first so the agent knows your competitors, your personas and the questions your buyers actually ask. Then pick one open opportunity and let it write. Read the draft with a hard eye, and if the topic was not ready, the run will have told you so before you spent the credits.

Frequently asked questions

What did DataEase AI ship in the Blog Creator?

Two modes at the same price. Interview me asks you 5 questions and writes from your answers. Write it for me runs without you. Both are schedulable, both are grounded in your own audited pages, and neither mode costs more than the other, so the choice is about the post rather than the budget.

What is autonomous blog writing grounded in?

Four inputs, all of them yours. Your audited pages, which are the only URLs the post is permitted to link to. Your measured standing across the AI assistants. Live topic research run at write time. And 1 open opportunity that the post exists to close, so the draft has a job you can evaluate it against.

What is the DataEase AI content agent forbidden from writing?

Seven categories are blocked outright: percentages, study results, survey findings, quotes, named customers, case-study outcomes and dates. Every one of them is a claim a reader will trust on sight and a language model will produce on demand, which is exactly the combination that makes fabricated detail dangerous rather than merely wrong.

What happens if topic research comes back thin?

The run fails at that step and tells you, rather than filling the gap with generic filler. DataEase AI treats a failed step as the correct outcome, and credits are refunded when a run fails. A model asked for 1,500 words with nothing to say will always produce 1,500 words, and that output is the hardest kind to spot.

Should you use interview mode or autonomous mode?

Interview mode sounds like you. Autonomous mode sounds like your brand. If the post depends on a story only you were in the room for, answer the 5 questions. If it is a topic your category needs covered and your calendar is the blocker, let the agent run and review the draft afterwards.