On this page13
- What does a PR workflow actually involve?
- The three-lane map: draft, assist, human
- Where AI drafts well: releases, reports and repurposing
- Where AI should only assist: research, lists and localisation
- What stays human, and why it's the expensive part
- The verification rule set I use on every AI output
- How do you roll AI into an existing PR workflow?
- When this map doesn't fit
- FAQ
- Which AI tools are best for PR work?
- Should I tell reporters I used AI?
- How much time does AI actually save in PR?
- Can AI build a media list for me?
Use AI in a PR workflow by sorting every task into one of three lanes. In the draft lane, AI produces the first version and a person edits it. In the assist lane, AI speeds up a person's research or thinking but never produces the final output. In the human lane, AI stays out. Every AI output then gets one named verification step before it touches a reporter, a founder or the public.
That's the whole method. The rest of this post is the map I use to decide which lane each task belongs in, stage by stage, and the specific check I run on each output.
What does a PR workflow actually involve?
People who haven't run PR tend to picture one task: asking a publication to write about you. I've pushed back on that for years.
what it actually is - creating a storyline, working on the messaging & narrative, pitching to targeted journalists, reaching out to them with well-curated pitches, following up & more! Shilika Jain on X, 13 Apr 2023
Broken into stages, a typical campaign for a launch or a funding announcement runs roughly like this:
- Research and monitoring
- Narrative and angle
- Media list building
- Pitch writing
- Outreach and follow-up
- Interview and spokesperson prep
- Releases, op-eds and owned content
- Localisation for other markets
- Coverage reporting
- Repurposing coverage into owned and search-friendly content
Each stage has a different mix of production work and judgment. That mix decides the lane.
The three-lane map: draft, assist, human
This is how I'd sort each stage today. The verification column matters most, because "AI did it" is never a finished state.
| Stage | Lane | What AI does | Verification before it's used |
|---|---|---|---|
| Research and monitoring | Assist | Summarises a reporter's recent articles, flags mentions | Open the original articles; never pitch off a summary alone |
| Narrative and angle | Human | Lists options you then reject | Founder and comms lead agree on one angle in writing |
| Media list building | Assist | Formats, de-duplicates, suggests beats | Check each reporter's last three bylines yourself |
| Pitch writing | Draft (lightly) | Subject-line variants, tightening | Read it aloud; remove anything you wouldn't say on a call |
| Outreach and follow-up | Human | Nothing, or reminders only | Personal send, personal follow-up |
| Interview prep | Assist | Generates likely hostile questions | Spokesperson rehearses with a person |
| Releases and op-eds | Draft | Structure and first pass from approved facts | Fact sheet check, every number and name traced to a source |
| Localisation | Assist | Rough translation for internal review | Native-speaker edit plus local outlet fit |
| Coverage reporting | Draft | Compiles links, summaries, tables | Click every link, confirm earned vs paid |
| Repurposing | Draft | Turns one interview into posts and FAQs | Founder approves anything in their voice |
A few of these need more explanation than a table cell allows.
Where AI drafts well: releases, reports and repurposing
The draft lane works when the inputs are already fixed. A press release built from an approved fact sheet, a coverage report built from a list of links, a LinkedIn post built from a transcript the founder already said out loud. The thinking happened earlier, so the model is only arranging it.
My rule for the draft lane: AI never gets to add a fact. It can reorder, shorten and restructure. If a number, name, date or quote appears in the draft that wasn't in the input, it comes out until someone can trace it. Invented quotes and confident wrong figures are the main way AI-drafted releases go wrong, and I go through those risks properly in the risks of using AI to write press releases.
Repurposing is where the time savings are largest. One founder interview can become a thread, a newsletter section, an FAQ block for your site and a short op-ed outline. That last step also helps with AI search, since answers that live on your own pages in clear question-and-answer form are easier for assistants to pick up. I cover that side in why PR drives AI search citations.
Where AI should only assist: research, lists and localisation
The assist lane is for tasks where AI speeds you up but a wrong output is expensive.
Research. Asking a model to summarise a reporter's recent work is a fine starting point. Pitching from that summary is not. Models blur reporters together, miss beat changes, and occasionally attribute articles to the wrong byline. Always open the originals before you write a word of the pitch.
Media lists. A good list is a set of judgments about who covers what this quarter, not a spreadsheet. AI can clean the spreadsheet. I've written about building one properly in how to build a media list, and none of that process is skippable because a tool exists.
Localisation. This one I feel strongly about because so much of my work is APAC. Machine translation gives you a rough internal draft. It doesn't tell you that Korean reach often runs through Naver-syndicated mainstream outlets and Telegram-driven news apps like CoinNess, which is something we learned placing coverage across Korea, and it doesn't rework the angle for what a local editor cares about. Translating an English release and hoping is the failure mode, with or without AI. When Tangem's self-custody survey was turned into a data story, one stat travelled across English, Korean, Japanese and Dutch editions without a separate pitch for each market. That travelled because the stat was strong, not because the translation was fast.
What stays human, and why it's the expensive part
The human lane is short but it's where campaigns succeed or fail.
- edit pitches - eat lunch - pitch pitches - eat dinner - follow up on pitches Shilika Jain on X, 23 May 2024
I posted that as a joke about the life of a PR manager, but it's accurate about where the hours go. Editing, pitching and following up are relationship work. A reporter can tell when the follow-up came from a sequence tool, and a templated nudge on a pitch they already ignored makes the next pitch less likely to be opened.
The narrative and angle stage also stays human. AI is useful for generating options so you can see what you're rejecting. The choice itself depends on what reporters are tired of, what competitors already said and what the founder can credibly defend in an interview. I'd never let a tool pick the angle for a funding announcement.
The verification rule set I use on every AI output
Each lane has its own checks in the table, but five rules apply to everything that comes out of a model:
- Trace every fact. Any number, name, date, title or quote must point to a source document a person has opened.
- Read it aloud. If a sentence sounds like nobody in particular wrote it, a reporter will hear that too.
- Check the person, not the summary. Reporter details, beats and recent articles come from the outlet's site, not the model's memory.
- Mark earned vs paid. In coverage reports, every link gets checked so sponsored posts aren't counted as earned coverage.
- Get the voice owner's sign-off. Anything published under a founder's name gets the founder's approval, line by line if needed.
How do you roll AI into an existing PR workflow?
If your team isn't using AI yet, don't change everything at once. A sequence that works:
- Week 1: Reporting. Start with coverage reports, the lowest-risk draft-lane task. You'll learn how the tool fails without any external exposure.
- Week 2: Research summaries. Add reporter and topic summaries, with the rule that nobody pitches without opening the originals.
- Week 3: Repurposing. Turn existing interviews and coverage into owned content, with founder sign-off.
- Week 4: Release first drafts. Only from a locked fact sheet, with the trace-every-fact rule enforced.
Pitches and outreach stay out of the rollout entirely. Those are the steps where a bad AI output costs you a relationship rather than an hour.
When this map doesn't fit
Two situations change the lanes. If you're a solo founder with no comms person, you'll lean on the draft lane more heavily for pitches too, and that's reasonable as long as you rewrite them in your own words before sending. And if your company works in a regulated area, such as finance, health or anything with securities questions, move releases and op-eds from the draft lane to the assist lane, because every sentence will need legal review anyway and a model's confident phrasing creates more redlines, not fewer.
The map also assumes someone senior is reading the output. AI in a PR workflow saves real hours, but only for teams that already know what good looks like.
FAQ
Which AI tools are best for PR work?
General assistants like ChatGPT and Claude handle most of the drafting and summarising described here. The tool matters less than the lane rules: a better model still shouldn't pick your angle or send your follow-ups. For a wider view of startup tooling, see my notes on the AI marketing tool stack for startups.
Should I tell reporters I used AI?
You don't need to announce that AI helped tighten a pitch. What matters is that every fact is accurate and the pitch reads as written by someone who knows their beat. If AI generated a quote or a statistic, that's not a disclosure question, it's an accuracy problem to fix before sending.
How much time does AI actually save in PR?
It saves the most on reporting, repurposing and first drafts, and close to nothing on relationship work. Teams that try to save time on pitching and follow-up usually lose it back in lower response rates. Measure it on your own workflow for a month before deciding.
Can AI build a media list for me?
It can suggest names and clean up a sheet. It can't confirm who's actually covering your topic this quarter or who has moved outlets. Treat its list as a starting draft and verify every reporter against their recent bylines.
Want a second pair of eyes on where AI fits in your comms setup? Book a 30-minute PR workflow teardown, or see how I run Web3 PR campaigns end to end.
