AI for Financial Publishers: What Pays and What Crosses the Line

A financial publisher should put AI to work for its own team first. Readers come later. Here is the order.

The first system that pays is a research assistant. It works over the firm's archive. Ask it what the house said on bank margins last spring. It answers in seconds. It uses your published work only. It links to each piece it used.

The second is reuse. Think charts and tables built from cited figures. Think newsletter blocks, and recaps for a new format.

Q&A for paying readers comes after that. It can be a real product. But it has to stay impersonal. The Supreme Court drew that line in 1985. Say it starts to answer "should I sell?" for one reader's portfolio. Then it has crossed that line.

Where AI pays for a financial publisher, ranked

1. The research copilot, for your own team. A copilot is an AI helper that works next to your staff. Analysts and editors ask the archive questions. They get answers with sources. Most of the hours are here. It is also the safest place to start. The readers are your own staff.
2. Reuse of published work. A chart or table drawn only from figures in the cited pieces. A newsletter block in your house style. A recap of last quarter's work for a new audience. Each is fast, and easy to measure. Each is checked before it goes out.
3. Checks before you publish. The AI flags a missing disclaimer. It flags a forward-looking statement, a claim about the future, with no caveat. It flags a number that doesn't match the source it cites. The editor decides. The machine points.
4. Q&A over the archive, for paying readers. Readers ask your research a question. They get an answer with sources. Arizent launched one in January 2026. It is for paid subscribers of American Banker and The Bond Buyer. It draws "exclusively" from each brand's own journalism and research archives. Each answer links back to the articles. Here the impersonal line shapes the design.
5. Personal picks. Answers on a reader's own holdings, or picks made for them. Don't build this, unless the firm is registered to give advice. Then it must supervise that advice as such.

Our estimate for a tool like this is $130,000 to $450,000 a year. That is for a firm with 5 to 50 analysts and editors. The basis is time. Each analyst spends 5 to 10 or more hours a week on rote work. It goes on lookups, pulling data and formatting. Faster work on what gets published adds to that.

The Impersonal Line

The SEC is the US watchdog for stocks and bonds. The Advisers Act has a carve-out for publishers. It is called the publisher's exclusion. It lets a newsletter comment on stocks. The newsletter need not register as an adviser.

The case that set its terms is Lowe v. SEC. The Court read the carve-out. It said it covers bona fide publications of general and regular circulation. In plain words, real publications. They go out to the public, issue after issue. The Court said the Act aims at "personalized advice." A publication stays outside the Act while its dealings with readers "remain entirely impersonal."

The diagram at the top of this page sorts AI uses by side.

A static newsletter stays impersonal by its nature. Each reader gets the same issue. A chatbot is not like that. It answers each reader on their own, on what they ask. And nothing stops a reader from typing their portfolio into it.

We call this design rule The Impersonal Line. A system for paying readers holds it with five rules:

• It answers from what was published. Each claim cites a piece that went out to all readers. It makes up nothing that isn't in the archive.
• It sums up. It doesn't advise. "Our March coverage argued that..." is research. "You should..." is advice.
• It turns down the personal question. A reader asks about their own holdings. It says what it can't do. Then it points to what you've published on the firms they named.
• It doesn't time itself to the market. No alerts built around one reader's holdings.
• It carries the disclaimer each time. Code adds it. The model can't forget it.

None of this is legal advice. The facts of your firm matter. Take the five rules to your lawyer. Let them draw the line for your firm. But draw it before you launch. Don't wait for the first reader to post a screenshot of an answer.

How a cited research answer is assembled

The pattern we build for these firms runs in a fixed order. The rules check always comes last.

First, the whole archive goes in. That means the website, read through its sitemap. A sitemap lists each page. It also means the PDFs, spreadsheets and decks behind it. Each scanned page goes through OCR. That turns a picture of a page into text.

Then a question searches two ways at once. One path goes through each piece's summary, then down to its best passages. The other goes straight to the passages. The two sets of results are merged and reranked. Reranking sorts the passages so the best come first. Those are the ones the answer will rest on.

Next, a lead agent hands out the work. It goes to one agent each for research, data, charts and market data. A compliance step runs last on each output. It adds the disclaimer. It checks that nothing reads as advice.

If the archive has no sure answer, the system says so. Any fact from outside the archive is marked as such.

We built one for a publisher that covers capital markets. It has 10,000+ documents and a 50-person editorial team. It answers only from the firm's own archive. Each answer has a source. It adds the disclaimer to each output, so no answer can go out without it. It runs across four business units. The general method is in how to make a RAG system cite its sources.

What a good answer looks like

The test is simple. Each sentence should take an editor under a minute to check.

Each claim has a numbered source. Each source has a title, a date and a link. The date matters more than people think. Take a correct answer from 2023, shown as current. In finance research, that is a wrong answer. A chart, if there is one, uses only figures from the cited pieces. And the disclaimer sits in the same place each time.

FINRA, licensed content and the other rules

If you are a broker-dealer, FINRA's rules apply as written. A broker-dealer buys and sells stocks and bonds for clients. FINRA put out Regulatory Notice 24-09 in June 2024. It says FINRA's rules don't change with the tool. Take Rule 2210, the rule on what firms send out. It applies "whether member firms' communications are generated by a human or technology tool." A firm that uses AI to write must also supervise it. It needs a sound system for that, one built for its own business. So a research recap that AI wrote and sent to clients is covered too. It gets the same review as any other.

Your archive is yours. Your data feeds may not be. Wire stories, licensed market data and outside research come with terms. Do the terms allow indexing? Do they allow AI-made versions? Check before they go into the knowledge base. The cleanest systems answer from what the firm published itself. They pull live prices from a feed licensed for display.

A policy before the reader product. Who approves a new data source? Who reviews flagged answers? Who can change the disclaimer? A one-page version is in AI governance policy.

What it costs and takes to run

We scoped the copilot above at $190 to $785 a month in tools and model costs. That pays for the language models and the hosting. It also pays for the search parts: embeddings, reranking and a database. Embeddings are number codes that let a system search by meaning. Our clients with 1,000+ daily AI interactions typically see all-in model costs of $200 to $1,200 a month. Those bills sit on their own accounts. More is in what it costs to run an AI system.

The model bill is small next to three jobs that never stop:

• Ingestion that keeps up. Ingestion means adding new work to the index. Each new piece goes in the day it is published. Each correction flows through too. If not, the system answers from the version you took back.
• The queue of questions it couldn't answer. An admin screen lists them. It is the best planning tool your editors will ever get. Someone should read it each week.
• A test set. Two or three dozen real questions. Each has the right answer and the right sources. Run it again after each change to the prompts, the model or the index.

What goes wrong with AI at a financial publisher

The answer cites the wrong piece. It's close on topic, but wrong on the number. Reranking and a source check catch most of it. Editors catch the rest.

Old research reads as current. The sources have no dates. Or the system picks the most-cited piece over the latest one.

The model fills the gap. Ask it about a topic the archive doesn't cover. It answers from its own training. The rule is the reverse: no source, no answer.

The disclaimer lives in the prompt. Prompts drift. Code doesn't. Add it after the answer is written, each time.

A chart from nowhere. The model works out a figure. It shows it as if it came from your data. Chart only what was cited.

Licensed content in the index. It's fine until the license holder asks.

What to ask a vendor

• Does each answer cite a published piece, with its date and a link? Show us, on a question we choose.
• What happens when the archive has no good answer?
• Where is the disclaimer added? Is it in the prompt? Or is it added by code after the answer is written?
• How does the system turn down a question about a reader's own holdings? Show us the exact reply.
• How soon can we search a new piece after it goes out? How are corrections handled?
• Can we see the test set and run it ourselves?
• Whose accounts do the model, index and logs run on? Who owns the code? The full list is in who owns the code an AI vendor writes.

When not to build a research copilot

A thin archive. Fewer than about 200 published pieces is too little to search well. Good search does more, for less.

A team of one or two. Say fewer than three people do research or content work. The hours saved won't cover the build. A lighter tool will.

You want stock picks. Making new investment advice is a compliance question first. It is a build question a distant second.

Your best content isn't yours. A system built mostly on outside feeds is a licensing project first. It is an AI project second.