AI for PR Firms: Press Release Drafts, Pitch Lists and Client Approvals

A small PR firm can use AI to draft releases and build pitch lists without sending anything unapproved by treating approval as a gate, not a courtesy: drafts live in a status folder, the client's written yes is logged against a version, facts and quotes are checked line by line, and only one named person sends. AI drafts and sorts. People approve and press send.

The risk in this work is rarely the draft itself. It is a release that goes out with a quote the spokesperson never said, a figure nobody verified, or a pitch to a journalist whose beat changed last year. Each of those is a process failure that AI makes faster. Below is a workflow that uses AI where it is strong, drafting and sorting, and puts human checks where it is weak: facts, quotes, contacts and sign-off.

Follow me on Instagram@sagnikteaches

Four states every release passes through

Most near-misses in agencies come from ambiguity about status. Someone says "the client is happy with it" in a call, and a junior reads that as approval. Define four states and make them visible in the file name or folder.

Connect on LinkedInSagnik Bhattacharya

1. Draft. Written by AI, edited by a person. Not shown to anyone outside the firm except the account lead.
2. Checked. The fact-check table is complete, every quote is confirmed by the speaker, and a second person has read it.
3. Client-approved. The client has replied in writing, naming the version, with any changes incorporated.
4. Sent. Released by one named person, with the date and recipient list logged.

A release can only move forward one state at a time, and it moves forward only when a person records why. If a client says "looks good" by phone, the account lead writes the email to confirm: "Confirming you have approved version 3 for release on Thursday at 09:00." A reply of "yes" closes the gap. No reply, no send.

Subscribe on YouTube@sagnikteaches

This also answers the contractual side. Say in your proposal that all releases require written client approval and that AI tools may assist drafting, with confidential client material handled as described in your terms. For wording, how to mention AI use in client contracts and proposals has examples. Do not rely on goodwill when a release has gone wrong; rely on the log.

Set up the tools and the data rules first

Client briefs, embargoed news and unreleased figures are confidential, sometimes market-sensitive. Before drafting anything, settle three things. First, the account: ChatGPT Business and Claude Team do not train on business content by default, according to both pricing pages in November 2026, and list prices are $20 per seat a month billed annually or $25 monthly with a two-seat minimum. Check the training setting on the vendor's own page. Second, what goes in: strip anything the client has marked as embargoed or price-sensitive unless your contract and the client's rules allow it. Third, who may use which tool, which is a policy question covered in how staff request a new AI tool.

For a one-person firm, a paid individual plan with training off may be enough. For a team, put the prompts below into a shared library so everyone uses the same wording; how to build a shared prompt library for your team shows how.

Drafting the release from a fact sheet, not a vibe

The best input to a drafting prompt is a structured fact sheet that the client has already signed off. It makes the draft quick and it limits what the model can invent, because every statement in the release should trace to a line on the sheet. Here is an illustrative fact sheet for an independent bookshop announcing an author event.

FACT SHEET v2 (client-approved 4 March)
Who: independent bookshop, [shop name], owner [owner name]
What: evening reading and signing with a local children's author, [author name]
When: Thursday 17 April, doors 18:30, reading 19:00 to 19:45
Where: the shop, [address withheld in draft]
Tickets: free, booking needed, 40 places
Why now: new paperback release; shop's tenth year trading
Approved quote (owner): "[exact words approved by owner]"
Approved quote (author): "[exact words approved by author]"
Contact for media: [name, role]
Do not say: sales figures, rival shop names, anything about the author's earlier publisher

Then the prompt sets the rules for the draft.

Write a press release of 300 to 350 words from the fact sheet below. Use inverted-pyramid structure:
the news in the first paragraph, then detail, then quotes, then a boilerplate and a media
contact. Use only facts and quotes in the sheet; copy quotes exactly. Do not add statistics,
superlatives, dates or claims. If something needed for the release is missing, write
[MISSING: what]. Respect the "Do not say" line. Plain tone, house spelling.
Fact sheet: [paste]

Illustrative output: a headline such as "Independent bookshop hosts evening reading with local children's author", a lead paragraph with who, what, when, where and why, two paragraphs of detail, the two approved quotes in quotation marks, and a short boilerplate. The model also adds a [MISSING: boilerplate text] for the shop's description. What you would fix: look for "popular", "beloved" or "much-anticipated", which the sheet never claimed; delete them. Check that "tenth year trading" has not become "ten years of serving the community", a claim with a different meaning. For tuning the release itself for journalists, how to write a press release with AI that journalists will read goes deeper on structure.

The fact-check pass: a table, not a glance

Models write confident prose, and an invented detail looks like a real one. The fact-check pass makes every claim earn its place. Build a table with one row per checkable statement: the claim, where it appears, the source, who verified it and the date. Here is the beginning of an illustrative table for the bookshop release.

You can use AI to build the claims column. Paste the draft and ask for a list of every checkable statement, numbers, dates, names, superlatives and quotes, as a table with a blank source column. It is quicker than reading for them, and it rarely misses one. The checking of sources stays with people. Do not ask the model whether the claims are true; it will say yes, politely.

List every statement in the draft that could be checked against a source: numbers, dates,
names, places, titles, superlatives, "first" or "only" claims, and quotes.
Output a table: claim, exact words in draft, type, source (leave blank). Do not judge
whether any claim is true.
Draft: [paste]

Treat "first", "only", "largest" and "leading" as red flags. They are claims that need proof, and drafts from models often add them for sparkle. For more on checking copy before it goes live, how to fact-check AI marketing copy before it goes live offers a general routine.

Quotes need their own rule

No quote is ever generated. Every quotation mark in a release surrounds words the named person said or approved. If a spokesperson is short of time, use AI to draft three interview questions, run a ten-minute call or email exchange, tidy their words lightly, and send the exact final text back for approval. Log the approval. A fabricated quote, even a bland one, is the single fastest way to lose a client and a journalist's trust. Add this rule to your terms, and make the check part of the "Checked" state.

Journalist list research: let AI suggest, let people verify

List building is where AI gives the most plausible-looking wrong answers. Ask a model for "ten journalists who cover independent retail" and you may get names that left a publication, beats that never existed and email addresses that are guesses. Use AI for the part it does well: brainstorming outlet types and angles, and sorting a list you have built.

A sound method has four steps. First, ask for categories, not names: "What kinds of outlets would cover a children's author event at an independent bookshop?" The answer might include local news sites, regional magazines, parenting publications, book trade titles, community calendars and local radio. Second, find real entries yourself from the outlets' own pages, recent bylines and your licensed media database. Third, record why each fits: the article you read, the date, the angle. Fourth, have AI sort and tidy the sheet, not generate it.

Below is my media list as a table (outlet, contact, beat, recent relevant article I read, date).
Tasks: 1) flag rows with no recent article; 2) group by outlet type; 3) suggest one angle
per group in a sentence, using only the details in the table; 4) do not add any contact, beat or
article not in the table.
Table: [paste]

Illustrative result: the model groups 22 rows into five outlet types, flags four rows that have no recent article, and proposes angles such as "a parent-facing 'things to do' listing" and "a trade story about independent shops marking milestones". You keep what fits. Contact data of individuals is personal data, so keep your list to business contacts, store it securely and follow the data-protection rules that apply to you; if you contact people in the EU, that includes the GDPR.

Pitches: personal, short and true

A pitch lives or dies on the first two lines, and that is where a model's generic enthusiasm hurts most. Use AI to draft from your list sheet, with the real article reference, and then personalise by hand.

Write a 90-word pitch email to [first name] at [outlet]. Use only these facts: [three facts from the
approved sheet]. Open with one sentence referring to this article I read: [title and date].
No flattery, no claims not in the facts. End with one clear question. Subject line under 8 words.

Illustrative output: a subject line "Children's author reading, 17 April", an opening that names the journalist's recent piece about reading habits, three lines of facts, and a question: "Would you like the release and an author interview slot?" What you would fix: models love "I hope this finds you well" and "I came across your wonderful work". Delete both. Check the article reference is correct and that nothing in the pitch goes beyond the approved facts. Then send nothing until the release is in the "Client-approved" state, because a pitch is a public statement too.

A worked example: what a release costs in time, before and after

Take an illustrative three-person PR firm producing about 10 releases a month for four small clients, such as a bookshop, a florist, a removals firm and a photography studio. Before AI, a release takes roughly 12 hours across drafting, fact checking, client rounds and list building, so 120 hours a month. Assume drafting takes about 4.5 of those hours and list building about 2.5.

With the workflow above, drafting from a signed-off fact sheet takes about 1 hour including edits, the claims table takes 45 minutes, list building drops from 2.5 hours to about 1.5 because the sorting is handled, and client rounds are unchanged because the approval gate stays. The sum per release is 12 hours, minus 3.5 saved on drafting and 1 on lists, plus 0.75 for the claims table: about 8 hours 15 minutes, so roughly 82 hours a month and a saving of about 38. Those are my own assumptions, labelled illustrative, and the saving can vary widely. The honest way to find your figure is to time two releases each way.

Notice what did not shrink: the client approval round and the human fact-check. Both are the point of the exercise. If a firm tries to shorten them, the time saved is borrowed against a future correction.

Before and after: what an unchecked draft looks like

A short contrast shows why the fact-check table exists. Suppose a junior asks a model for a release about a removals firm opening a second depot, giving it only "second depot opening in May, three new vans". An unchecked first draft might read: "Leading removals firm expands to meet soaring demand, opening a flagship second depot in May with a fleet of three state-of-the-art vans. Customers have praised the firm's industry-leading service."

Run the claims list on it and the problems appear at once. "Leading" and "industry-leading" are unsupported superlatives. "Soaring demand" is a cause the client never stated. "Flagship" and "state-of-the-art" are decoration that a journalist will cut and a competitor may challenge. "Customers have praised" implies evidence nobody supplied. Only two facts survive: a second depot in May and three new vans. The repaired lead reads: "A removals firm will open a second depot in May, adding three vans to its fleet." Shorter, true, and approvable. Everything else needs a source from the client or comes out.

The lesson for the team is that model drafting adds a layer of flattering language by default. Make "delete anything the fact sheet does not support" the first edit on every draft, before style or length.

A holding statement when the news is bad

Not every release is good news. When a client faces a complaint, a recall or an outage, speed tempts people to paste the situation into a tool and send whatever comes back. This is the moment the approval gate matters most. A holding statement is a short, factual line issued while the facts are gathered, and AI can help draft options, provided the facts come from the client and nobody outside the firm sees it before sign-off.

Draft three versions of a holding statement of no more than 60 words. Facts supplied by the client:
[list only confirmed facts]. Rules: state only those facts; no blame, no speculation, no promise of
an outcome; include one line saying when the next update will be given; plain tone.
If a fact needed is missing, write [MISSING: what].

Illustrative output: three short options, each stating what happened, what the client is doing, and when the next update comes. What you would fix: any apology wording the client's lawyer has not seen, any sentence that implies a cause, and any promise such as "this will not happen again". For sensitive matters, add a legal or insurer review state before "Client-approved". Keep the draft in the same locked-file routine, because a holding statement revised in a hurry is the version most likely to be sent in the wrong state.

Where approvals slip, and how to stop each one

• The verbal yes. A client says "go ahead" on a call. Fix: the account lead sends a confirming email naming the version, and only a written reply counts.
• The late edit. A client approves version 3, then a colleague tweaks a number before sending. Fix: after approval the file is locked; any change creates version 4 and goes back through approval.
• The embargo slip. A release marked for Thursday 09:00 goes to one journalist on Wednesday. Fix: embargo time sits in the first line of every file name and the send log, and the sender checks it.
• The wrong list. The pitch goes to last month's list. Fix: each release has its own list sheet tied to the version.
• The silent AI edit. Someone pastes the approved release into a tool to "tighten" it, and a word changes. Fix: AI work happens before approval, never after. Post-approval changes are by hand and tracked.

The last one is easy to miss. A model asked to shorten a paragraph will happily alter a number or turn "approximately 40" into "40". Make the lock a rule, not a habit.

What happens after sending, without repeating monitoring

Once the release is out, coverage tracking is its own job and has its own tools. AI media monitoring for small PR agencies covers alerts, sentiment and reports. From a process view, the link back to this workflow is one sheet: record which outlet picked up which version, so that the next release's list gets better. Over three or four releases for the same client you will see which outlet types respond, and your list research becomes quicker and more accurate.

Keep a short post-mortem for any release with a correction. A single line, "what slipped and which state should have caught it", is enough. After ten of these you will have a record that tells you which of the four states needs reinforcing.

Auditing the send log once a quarter

A workflow like this only works if the gate holds when nobody is watching, so test it. Once a quarter, pick five sent releases at random and pull three things for each: the version that went out, the client's written approval, and the fact-check table. Then answer four questions. Does the sent file match the approved version word for word? (Paste both into any document-compare tool; a single changed figure is a finding.) Does every quote have a logged approval from the speaker? Does every row in the claims table have a source and a name? Was the embargo time in the send log met?

An illustrative result for a three-person firm: five releases checked, four clean, one where the sent version carried a corrected date that was added after approval without a new version number. Nothing went wrong in public, but the date change skipped the gate. The fix is procedural, not technical: the account lead re-reads rule 6 below with the team, and the next quarter's audit checks that file type again. Keep the audit to under an hour. Its value is the habit, and a written record you can show a client who asks how their releases are controlled.

A one-page rule set for the team

1. Every release starts from a client-approved fact sheet.
2. AI drafts, sorts and lists. It never supplies facts, quotes or contacts.
3. Every claim goes in the fact-check table with a source and a name.
4. Every quote is confirmed verbatim by the speaker.
5. Approval is a written reply naming the version.
6. Approved files are locked; changes make a new version.
7. One named person sends, and logs the date, list and version.

Put the rule set in your proposals and on your wall. If you would rather map the workflow to your own client mix with someone, that is what my AI implementation consultation covers, through the contact page. The AI guide lists the rest of the series.

Questions PR firms ask about AI drafts and approvals

Can AI write quotes for a client spokesperson?

No. A quote must be something the named person said or has approved in their own words. Let AI draft a question list for the interview, or tidy a quote the spokesperson gave you, then send the final wording back to them for sign-off before it appears in a release.

Can AI find journalist email addresses and beats?

It can suggest outlets and angles, but it cannot be trusted for contact details or current beats, and it may invent both. Build the list from the outlet's own pages, the journalist's recent bylines and your licensed database, and check each entry yourself before it is used.

Do we need to tell clients we use AI?

Yes, in plain words in the contract or proposal, along with what is and is not entered into tools. Many clients already ask. Disclosure also supports your approval gate: they know a human checks every release and that they sign off before anything goes out.

How do we stop an unapproved release going out by mistake?

Make approval a visible state, not a feeling. Keep drafts in a folder named for status, make sending depend on a written yes with the date and version, and allow only one person to press send. A release with no recorded approval does not leave.

Further reads

• How Small Agencies Use AI to Write New-Business Pitches — Using AI for your own new-business pitches.
• AI Social Media Monitoring: A Monthly Insight Report for Clients — A monthly client insight report built with AI.
• Is It Safe to Put Customer Data Into ChatGPT? — The data basics behind your confidentiality rules.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in a tool's privacy terms.
• AI Ethics for Small Businesses: A Practical Checklist — A practical ethics checklist for a small business.
• AI Hallucinations Explained for Business Owners: Causes and Fixes — Why models invent details, in plain English.
• AI PR Pitch Emails: Personalise for Journalists With a Review Step — A five-step pitch workflow for small PR agencies: beat note, hook line, three-line pitch, manager sign-off and a follow-up cadence that stops on time.
• Crisis Communication Plan Template: Draft Holding Statements With AI — A filled-in crisis plan: scenarios, holding statement templates, an approval chain and a media Q&A drill, with AI drafting and named people approving.
• Media List Building With AI: Find, Check and Segment Journalists — A four-stage media list method: AI finds candidates, you verify beat, recent work and contact data, then segment and keep opt-outs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business and Claude Team pricing pages, checked November 2026; the workflow is general practice guidance, not a vendor claim.