Featuring Harshini Kumaran, founding marketer at Magi. Last updated: August 2026

In short

• The bottleneck in a multi-brand AI marketing workflow is re-briefing, and uploading a brand deck per task is re-briefing with extra steps.

• Harshini Kumaran runs Magi's LinkedIn, weekly newsletter and blog, plus LinkedIn for customers: 2 to 3 posts a week per brand, up to five voices on some.

• Ideas are assembled overnight from Fathom calls, Knowledge, LinkedIn monitoring, and weekly research, each with its source attached.

• The same chain- campaign to brief to outline to draft- produces every format. Only the template changes.

Harshini Kumaran is Magi's founding marketer. She runs the company's LinkedIn, writes the weekly newsletter, keeps the blog moving, and does LinkedIn for several Magi customers as well: 2 to 3 posts a week per brand, and on some brands three to five voices. Between all of that, she's on customer calls, which is where most of the good material comes from.

Drafting was never where the time went. Re-briefing was. Before every post: which brand, which persona, what did the customer say last Thursday, what have we already claimed, how does this voice sound when it's being funny. Across brands and voices, the context switching cost more than the writing.

She tried the obvious fix. Upload everything into a general-purpose AI tool, ask for a post, edit. The output came back generic enough that writing from scratch was competitive on time and better on quality.

Re-briefing AI is the hidden cost of scaling content

The pushback writes itself: the deck is context, so blame the model. And better models do help. The floor has risen every year since 2023.

The ceiling hasn't, because of what happens to the context. A deck uploaded per task is a transfer a human performs and then throws away at the end of the session. Nothing accumulates. And flattening a brand into a document strips the structure that made it usable: a model reading 40 pages can't tell a rule from an aspiration, so it returns an impression of the whole file. Averages are what generic sounds like.

That's the thread through this whole series. What got automated in every Magi workflow is the movement of information that already exists, from one container into another. We call the cost of that movement the transfer tax. Harshini's setup stops paying it on the way in, by keeping context in a structure the drafts are generated from.

Define the campaign once. Every asset inherits the context.

A Magi campaign is the container. Take the one running Magi's security-vertical work: market, product, brand, two ICPs (scaleups, enterprise), three personas (Head of or VP Marketing, VP Demand Generation, Global Head of Campaigns), and a brief written for people who respond to proof and not hype. It names their four core pressures and what Magi solves for each.

Then the two fields that do the downstream work. Tone guardrails: precise, outcome-focused, grounded in real proof, no vague AI positioning. Research topics: new Gartner, Forrester and IDC findings on cybersecurity GTM, the metrics analysts are emphasising, and what CISOs are actually discussing on LinkedIn, Reddit and industry forums.

The same campaign, further down: tone guardrails and the three research topics the Research Agent runs against every week.

Setup takes minutes and Magi pre-populates most of it. Everything generated under that campaign inherits all of it. This is the difference between AI in B2B marketing as an operating system and as a chat box.

Good ideas come with receipts

The Ideas view is where the week starts, and nothing in it was generated on demand.

Four inputs feed it. Customer and prospect calls via Magi's native Fathom integration, so a buyer's phrase from Tuesday afternoon is a post angle Wednesday morning. Knowledge, which takes meeting recordings, Notion pages, Drive files and URLs. LinkedIn monitoring of the competitors and influencers configured per brand. And the Research Agent's weekly deep research against the campaign's topics, which the Ideation Agent converts into angles.

The Ideas view. Each row carries its campaign, source knowledge, content type and relevance score.

Each row shows title, excerpt, campaign, source knowledge, content type and relevance score. Harshini can check where an idea came from before committing a post to it. The ideas she reviews at 9am were assembled overnight from conversations she was in at 3pm.

Approve the brief. Approve the outline. Then let AI write.

From an idea: pick a voice, pick a template, send. Voices are defined in BrandOS with example sentences, which is what most brand voice AI gets wrong; "confident" gives a model nothing to imitate, six real sentences do. Magi has five voices configured and a brand can add as many as it needs. Templates carry information hierarchy: LinkedIn default, event recap, announcement.

The first thing back is a content brief for review. Then an outline for review. Then the Content Agent drafts. A wrong angle caught at the brief costs half a minute. Caught in a finished draft, it costs a rewrite.

The draft pulls from the template for shape, the campaign for situation, and BrandOS for how the brand sounds and what it can claim.

A Google Cloud stat became a founder-facing post, and nobody typed the persona into a prompt

The idea read 97% Plan to Spend More on AI Before They Add Another Tool, under the brand awareness campaign, surfaced by the weekly research run.

The brief framed it for founders. The draft opened on the statistic and turned it: most of that spend will make marketing messier, because a new AI licence does nothing about context that lives in the founder's head and in decisions nobody wrote down. It closed by asking where the reader's brand context lives today. That's the campaign's persona and pressure landing in the copy unprompted.

The Design tab produced the visual, a padlock with the line One source of truth. Deekshaa gave design input through comments on the asset, in the editor. Harshini published to LinkedIn from the same screen.

LinkedIn, newsletters and blogs shouldn't need separate workflows

None of the above is LinkedIn-specific. The campaign, the Ideas queue, the voice, the two gates and the publish step are the same for every format Harshini ships.

The weekly Team Magi newsletter runs on a newsletter template and a newsletter skill switched on during generation. Its two inputs each week, the product release note and a resource or take, come out of the same Knowledge and Ideas queue. The blog runs on a blog template with the same brief and outline gates, and publishes from the editor to WordPress or HubSpot rather than LinkedIn. Skills stack: a humanizer skill runs on any format where the voice audit matters.

So the marginal cost of a new format is a template, and the marginal cost of a new brand is a campaign. The re-briefing that used to scale with both now scales with neither.

Your BrandOS sets the ceiling for your AI marketing

Across Magi's customers, output quality varies, and the variable is upstream. Where BrandOS is full and voices are defined with examples, drafts often publish as they come. Where BrandOS is thin or voices are adjectives, drafts drift toward the generic register Harshini was trying to escape, and editing happens by chatting with Copilot in the editor, which beats rewriting but is still work.

AI marketing automation removes the transfer tax. It doesn't remove the need to have decided what the brand is. The hours saved on re-briefing get spent getting the definitions right, because every draft after that compounds them.

The first step to automating marketing is finding what you're still repeating

Pick one brand you write for. Count how many times last week you re-explained who it is and who it's for, to a tool, a person, or yourself after switching brands. If the number is small, a better prompt is enough. If it's most of every post, the fix is a container the context lives in and a marketing workflow that generates from it.

Next: Deekshaa, on the transfer tax in design, and where the line between what a machine can move and what a designer must decide actually sits.

Frequently asked questions

What is an AI marketing workflow? An AI marketing workflow is a repeatable sequence in which AI agents gather signals, draft, format and schedule marketing assets while a person holds the review points. The distinguishing feature of a good one is that context persists in a structure between tasks instead of being re-uploaded each time.

How do AI agents transform content marketing? AI agents transform content marketing by moving labour from production to review. In Magi, a Research Agent gathers signals weekly, an Ideation Agent turns them into angles, and a Content Agent drafts from a campaign and BrandOS, so the marketer approves briefs and outlines rather than assembling them.

Why does AI-generated LinkedIn content sound generic? AI-generated LinkedIn content sounds generic when brand context arrives as a flat document and the model averages it. A brand book uploaded per task loses the difference between a rule and an aspiration, so the output reflects the whole file rather than applying any part of it.

What is the best AI for LinkedIn posts for a B2B team? The best AI for LinkedIn posts for a B2B team is one that holds brand and campaign context persistently rather than per prompt. Magi does this through campaigns and BrandOS; a general-purpose chat tool starts from zero context on every request.

Can the same workflow produce newsletters and blogs? Yes. In Magi the campaign, Ideas queue, voice and review gates are shared across formats. A newsletter uses a newsletter template and skill; a blog uses a blog template and publishes to WordPress or HubSpot. Only the template changes.

How does Magi keep brand voice consistent across multiple LinkedIn profiles? Magi keeps brand voice consistent across profiles by defining each voice in BrandOS with example sentences and generating every draft from that definition. Harshini runs up to five voices for a single brand this way, and the Audit tab checks finished posts against the same rules.

Does Magi integrate with Fathom? Yes. Magi's native Fathom integration pulls customer and prospect call recordings into Knowledge, where the Research and Ideation Agents draw on them. A phrase from an afternoon call can surface as a post idea the next morning.

What is BrandOS? BrandOS is Magi's marketing operating system. It holds a brand's identity, voices, tones, messaging, language rules, visual identity, keywords, content guidelines and legal constraints in one structured place, and every agent generates from it.

About this post. First in The 100x Marketer, a series on how Magi's own team uses the product. Sourced from Harshini Kumaran's recorded workflow episode, Magi's product as configured for its own marketing, and the security-vertical campaign brief shown in the episode. Published by Magi HQ, an agentic marketing automation platform for lean B2B teams. Customers include 100ms, Payactiv, Lyric and Accuknox.

TL;DR. Harshini Kumaran, founding marketer at Magi, runs one AI marketing workflow across LinkedIn, the weekly Team Magi newsletter and the blog, plus LinkedIn for customers at 2 to 3 posts a week per brand in up to five voices. A Magi campaign holds market, ICPs, personas, brief, tone guardrails and research topics once. AI agents for marketing assemble ideas overnight from Fathom calls, Knowledge, LinkedIn monitoring and the Research Agent's weekly research, each with its source attached. From an idea she picks a BrandOS voice and a template, approves a content brief and an outline, and the Content Agent drafts. Edits run through Copilot, visuals through the Design tab, publishing direct to LinkedIn, WordPress or HubSpot. Newsletters and blogs run the same chain with a different template and skill. Output quality tracks how well BrandOS is defined. This AI content automation removes the transfer tax on the way in; it does not decide what the brand is.