What is an AI content system?

The difference between having AI write something and being able to publish it. Most teams already generate drafts fine; what they lack is the part that decides whether a given draft is allowed out, which is what gets built here.

What is the difference between an AI content system and an AI writing tool?

A tool generates text and stops. Everything expensive happens after that — deciding whether a sentence is supported, whether it sounds like you, who signs it off, and what happens to it when the underlying facts change six weeks later.

What is the difference between AI content and content writing?

Content writing decides what should be said and writes it. This service builds the workflow a team uses to produce that content repeatedly under control. They work together, and the writing service is where the editorial judgement lives.

What is the difference between AI content and AI strategy?

AI strategy decides whether and where AI is worth using across a business, and in what order. This service implements one specific thing: the content workflow. Strategy first is often sensible; it is not a prerequisite if the content problem is already obvious.

Can the system write in our brand voice?

Yes, though the useful test is whether two of your own writers would mark the same sentence as off-voice. If they would not, the voice is not written down yet — and getting it written down tends to improve the human output as much as the generated kind.

Can it use our approved company information?

Yes — that is the point of it. The workflow is scoped to specific approved sources, such as product documentation, approved FAQs and brand guidelines, agreed at the start rather than pointed at everything a business has.

Can it check claims against our source material?

Yes, and the sentence worth worrying about is never the obviously wrong one. It is the plausible line nobody thought to question — which is exactly the kind a reviewer skims past, and the reason the checking is done per sentence rather than per document.

Do you use confidence percentages?

No. A number nobody can verify is not a control, and it tends to make people read less carefully rather than more. Supported, needs review and not in source are states somebody can act on.

Can people review content before it is published?

Yes, and the goal is to make review short rather than optional. A reviewer handed a marked draft is deciding about three sentences; a reviewer handed an unmarked one is re-reading everything, which is how review quietly becomes a rubber stamp.

Can different people have different approval rights?

Yes, and it is worth separating two things people conflate: who may produce a draft, and who may make it public. Opening the first widely while keeping the second narrow is usually what a business actually wants.

Can one source produce several different outputs?

Yes — a website announcement, a customer email, a sales note and social source copy can all run from the same approved facts, each written for a different job and depth. That is where a system earns its cost.

What happens when the source material changes?

You get a list of what to re-read, not an automatic correction. Deciding a published sentence has gone stale is a judgement — a changed price makes some pages wrong and leaves others fine — so the system finds the candidates and a person rules on them.

Can it help update old content?

Yes, and it is often the fastest win. Existing content can be checked against current source material to find the statements that have quietly gone out of date.

Can our existing content be used as source material?

Often, yes — but only the parts that are still accurate. Working out which of the existing material is safe to treat as a source is usually one of the first things a project does.

Can the workflow connect to our CMS or docs?

Usually, and it is checked first rather than assumed — the limit is normally what a system will let you read out, not what it will let you write in. Whatever cannot be reached gets named up front rather than discovered mid-build.

Will it produce a large volume of content quickly?

It will produce drafts quickly. Whether that becomes a large volume of published content depends on your review capacity, and we would not design a workflow whose main achievement was outrunning it.

What determines the scope of an AI content project?

The state of your source material, more often than anything technical. A team whose facts live in one maintained place is a contained build; a team whose facts live in six documents that disagree has a different project in front of it, and it is not really an AI one.

What do you need from us?

One thing above the rest: a recent piece of content that went out slightly wrong, and the story of how. That tells us more about where the workflow actually fails than any brief, because it shows which check was missing rather than which was intended.

Who owns the workflow and the content it produces?

You do. The part worth checking with any supplier is the configuration — the voice rules, the prompts, the review logic — because that is where the work actually accumulates, and it is handed over rather than held.

What happens after it is built?

The first weeks usually change it. Real jobs surface source material that is not as approved as everyone thought, and rules that need sharpening. How much of that is included is agreed per engagement.

AI content systems

Let AI draft. Keep the truth, the voice and the approval yours.

We design AI-assisted content workflows that run on your approved source material, your brand rules and a review step — so drafting gets faster without publishing getting looser.

See a draft checked against its source

Northstar LabsContent job #027

Source ready

Job Customer announcementVoice Clear · Confident · Practical

Approved product brief

01Shared project spaces are available on paid accounts.
02Admins can create a space.
03Members can be invited into a space.
04Existing projects can be moved into a space.

Draft

Shared project spaces are now available on paid accounts.
An admin can create a space and invite the people who need it.
Existing projects can be moved in without starting again.
Most teams have this set up in an afternoon.
Every customer gets unlimited spaces.

Approved

Shared project spaces are here

Shared project spaces are now available on paid accounts.

An admin can create a space and invite the people who need it, and existing projects can be moved in without starting again.

Open a space

Removed: Every customer gets unlimited spaces.
Confirmed: Most teams have this set up in an afternoon.

It starts with source material somebody has approved.

Illustrative content workflow · sample data

Before the system

Generating text stopped being the hard part some time ago.

Deciding what is allowed to become public never did.

An AI content system is a controlled workflow that helps a team turn approved source material into drafts or content variations using defined brand and review rules. It is worth building when generation is only one step inside a larger publishing process — which is almost always the case once more than one person is accountable for what goes out.

PromptWrite a customer announcement about our new shared project spaces feature.

DraftA perfectly reasonable draft, in about four seconds.

SourceWhich facts is it allowed to use, and which did it invent?
VoiceDoes this sound like the company, or like a model?
Claim checkWho confirms the parts nobody has verified?
ApprovalWho says this can go out?
VersionWhich of the six drafts is the real one?
OwnerAnd who answers for it afterwards?

None of those six are generation problems. They are workflow problems, and they are the reason a team that adopted an AI writing tool six months ago is often producing more drafts and publishing about the same amount.

The source lock

Every sentence has to come from somewhere.

Three states, because three are what a person can act on.

Yes, a workflow can check a draft back against the source material it was given. Each statement is traced to an approved fact, marked for review where nothing confirms it, or flagged as not present in the source at all. What that produces is not a score — it is a short list of sentences somebody has to make a decision about.

Approved source

01Shared project spaces are available on paid accounts.
02Admins can create a space.
03Members can be invited into a space.
04Existing projects can be moved into a space.

Choose a sentence to see where it came from.

Traces to: Shared project spaces are available on paid accounts.

A direct restatement of the approved fact.

Traces to: Admins can create a space. Members can be invited into a space.

Combines two approved facts without adding anything to either.

Traces to: Existing projects can be moved into a space.

“Without starting again” is a rephrasing, not a new claim.

Nothing in the source says this.

Probably true and nobody has checked. It needs somebody who knows to confirm it, or it does not go out.

Nothing in the source says this.

The source says paid accounts, and says nothing about limits. This is the sentence that would have to be corrected in public.

Selected sentence traces to 1 approved fact.

The fourth sentence is the interesting one. It is probably true, and it is exactly the kind of line that goes out unchallenged because it sounds reasonable — which is why "needs review" is a state and not a rounding error.

What the system does not do is put a confidence percentage next to a sentence. A number nobody can verify is not a control; it is a reason to stop reading carefully.

Brand voice

A voice a model can follow is a list of rules, not three adjectives.

“Clear, confident, practical” tells a person something. It tells a system almost nothing.

Yes, a workflow can be set up to write in a defined brand voice — but only once that voice exists as rules rather than as adjectives. Turning "clear and confident" into instructions a system can apply, and a reviewer can check against, is usually the most valuable hour of an AI content project.

Clear
Confident
Practical

What that means in practice

Short sentences. One idea each.
The product’s own words for its own features.
Plain verbs — “move”, “create”, “invite”.
Specific detail over general enthusiasm.

And what it rules out

Superlatives nothing supports — “seamless”, “effortless”.
“Revolutionary”, “game-changing”, “next-generation”.
Sentences that open by restating the topic.
Excitement doing the work a fact should do.

Off voiceWe’re thrilled to announce a game-changing new feature that revolutionises the way modern teams collaborate seamlessly across projects.

On voiceShared project spaces are now available on paid accounts. An admin can create one and invite the people who need it.

Same information, roughly a third of the words, and the second one could only have been written about this product. That last part is the test worth applying — if a competitor could publish the sentence unchanged, the voice rules are not doing their job.

The content job

Better output starts with a better brief, not a longer prompt.

The job is a structure. The prompt is just how it gets delivered.

The quality of what a content workflow produces depends far more on how the job is defined than on how the prompt is phrased. A brief that states the audience, the job to be done, the approved source, the format, what must be included and what must not be claimed gives a system something to be accountable to — and gives a reviewer something to check against.

Audience: Existing customers on paid accounts
Job: Explain what shared project spaces are and how to start
Source: Approved launch brief, version 3
Format: Customer email, around 120 words
Must include: Availability · the admin action · that existing projects can move
Must not claim: Anything about pricing, plan limits or a release date

The last field does most of the work. A brief that says what may not be claimed is the difference between a review that catches an invented sentence and a review that never thinks to look for one.

The review gate

Publishing can stay behind a person.

Fast drafting and controlled publishing are not in conflict.

Yes — a content workflow can require human review before anything is published, and in most businesses it should. The system’s job is to make review fast and specific: the reviewer opens a draft that is already checked against the source, already written to the brand rules, with the two sentences that need a decision marked. What they approve is what goes out.

01Draft readyGenerated against the brief and the approved sourceSystem
02Source checkEvery statement traced, flagged or marked for reviewSystem
03Brand checkMeasured against the voice rules, not against tasteSystem
04Editor reviewThe decisions a person actually has to makeEditor
05Approve or returnApprove, edit, or send it back with a reasonApprover
06PublishedAnd the version is recordedApprover

What the editor actually decided

Remove the generic opening: The first sentence restated the subject line and said nothing.
Cut the unsupported claim: “Unlimited spaces” is not in the source, so it does not go out.
Confirm or drop the unchecked one: Somebody who knows says the afternoon estimate is fair, so it stays.
Keep the source-backed detail: The migration note is the line customers will actually act on.
Sharpen the call to action: “Learn more” became “Open a space”.

Does this replace writers and editors?

No, and a workflow that assumed it did would produce exactly the content this page argues against. What changes is where the time goes: less of it on first drafts and formatting, more on the judgement calls — what to cut, what needs checking, and whether the thing is worth publishing at all.

Content

One source

The same approved facts, doing four different jobs.

And one process for when the facts change.

Yes — one approved source can produce several content variations, each written for a different job and a different depth, all tracing back to the same facts. That is where a system earns its cost: the source is agreed once, and every output is accountable to it rather than to whoever wrote that particular draft.

Website announcementExplain it to anyone who arrivesFull context, no assumed knowledge
Customer emailTell people who already pay youShort, and it assumes they know the product
Sales noteHelp a conversation that is already happeningThe objections, and what to say to them
Social source copyGive the social team something true to work fromThe facts and the angle, not the finished post

That last one is deliberately source copy rather than a finished post. Adapting a message to a platform — the format, the hook, the pace — is its own work and its own service.

Social media

And when the source changes

Source v1Spaces are available on paid accounts
Draft v1Four outputs published against it
Source v2Spaces now included on the team plan as well
FlaggedThe sentence that says “paid accounts” is now incomplete
Draft v2Corrected, reviewed and republished

The system does not decide that a sentence is wrong. It knows which outputs were written against which version of the source, so when the source moves, finding what needs re-reading takes minutes instead of somebody remembering.

Where this fits

Four neighbouring services, and the question each one answers.

The AI ones get confused most often, so it is worth being plain about it.

An AI content system is not the same as an AI writing tool. A tool generates text; a system connects that generation to the business context around it — the source material it is allowed to use, the brand rules it has to follow, the review it has to pass, and the versioning that keeps it current. It is also not AI strategy, which decides whether and where AI is worth using at all.

What should we say, and how should it be written?ContentStrategy, structure and the editorial work itselfContent
How does our team produce this repeatedly, under control?AI content systemsSource, brand rules, drafting, review, versioningThis page
Should we be using AI here at all, and where first?AI strategyAdoption roadmap, opportunity mapping, prioritiesAI strategy
Can AI move actual work through the business?AI agents & automationAgents and automations across operational workflowsAI agents & automation

And search

A content system can produce answer-first structure and keep it current, which helps. What it does not do is decide the search architecture — what should rank, what should exist, and how it all connects.

SEO & AEO

Should we just buy a tool instead?

Sometimes, and we would say so. If one person writes occasional content and nobody else reviews it, a good writing tool is the right answer and a workflow is overhead. A system starts paying when several people are involved, the same facts get rewritten repeatedly, and somebody is accountable for what goes out.

Control

The workflow only reads what it is meant to read.

Which sources, which people, and who signs it off.

Yes, a workflow can be scoped to specific approved sources rather than to everything a business has. Which material it can draw on, who can run a job, who reviews and who approves are configuration decisions made per project — and defining them is usually part of the build rather than an afterthought.

Sources this workflow reads

Product documentation
Approved FAQs
Brand and voice guidelines
Published service pages

Sources it does not

Unapproved drafts
Internal finance material
HR and personnel records
Anything not agreed as a source

What sits in each column is agreed at the start of a project. The point is not that a system is inherently safe — it is that a system reads what it was pointed at, and pointing it carefully is part of the work.

01Content ownerDefines the job and the source it runs against
02The workflowDrafts, traces every statement, applies the brand rules
03EditorMakes the decisions the draft surfaced
04ApproverSigns off, where a project needs sign-off

Shown as four for clarity. Plenty of projects run with two, and one person can hold more than one role — what matters is that each of these is somebody rather than nobody.

Who owns what: Your source material stays yours, and so does the approved content the workflow produces. The workflow itself, its configuration and its brand rules are built for you and handed over with the project, including the access needed to run and change it. What that looks like in practice is set out in each engagement rather than assumed here.
Connecting to what you already use: A workflow can read from and write to existing systems where those systems allow it — a CMS, a documentation tool, a content repository. We confirm what each one can actually expose before designing the connection, rather than promising an integration and discovering the limits later.

Scope

What makes one AI content project larger than another.

Mostly how many jobs it has to handle, and how good the source already is.

Scope depends on how many content workflows are involved, how many source systems they read from and what state that material is in, whether brand voice rules already exist, how many output formats and languages are needed, how complex approval is, how much source and claim checking is required, whether versioning and history are needed, which tools it has to connect to, and how many people use it in which roles.

Number of workflowsOne repeatable job, or several
Source systemsHow many, and whether they can be read at all
Source qualityWhether the material is approved and current
Brand voice rulesWhether they exist or have to be written
Output formatsHow many jobs one source has to serve
Approval complexityHow many reviewers, and with what rights
Claim checking depthHow rigorously statements must be traced
Versioning and historyWhether outputs must be tied to source versions
IntegrationsCMS, docs and repositories the workflow touches
Roles and permissionsHow many people, with which rights
Languages and marketsWhether the same job runs more than once
Custom interfaceWhether the team needs a surface of their own

What we need from you: The content jobs your team repeats, wherever the source facts actually live, whatever brand or voice material exists, the names of the people who review and approve today, and one recent example of content that took longer than it should have. That last one usually explains the whole problem.
Where projects usually start: With one content job rather than a platform. Getting a single repeatable job — the product update, the case study, the FAQ — running end to end teaches more about what the system needs than any amount of planning, and it is the fastest way to find out whether the source material is in good enough shape.

Selected work

Content and AI work we have made.

Published projects, with the scope shown taken from each project’s own record.

Selected content and AI work. The scope shown under each project is listed from its own record.

Content Strategy, structure and writing AI strategy Where AI is worth using AI agents & automation Agents across operations Selected work Everything we have published

Questions

AI content systems, answered.

What is an AI content system?: The difference between having AI write something and being able to publish it. Most teams already generate drafts fine; what they lack is the part that decides whether a given draft is allowed out, which is what gets built here.
What is the difference between an AI content system and an AI writing tool?: A tool generates text and stops. Everything expensive happens after that — deciding whether a sentence is supported, whether it sounds like you, who signs it off, and what happens to it when the underlying facts change six weeks later.
What is the difference between AI content and content writing?: Content writing decides what should be said and writes it. This service builds the workflow a team uses to produce that content repeatedly under control. They work together, and the writing service is where the editorial judgement lives.
What is the difference between AI content and AI strategy?: AI strategy decides whether and where AI is worth using across a business, and in what order. This service implements one specific thing: the content workflow. Strategy first is often sensible; it is not a prerequisite if the content problem is already obvious.
Can the system write in our brand voice?: Yes, though the useful test is whether two of your own writers would mark the same sentence as off-voice. If they would not, the voice is not written down yet — and getting it written down tends to improve the human output as much as the generated kind.
Can it use our approved company information?: Yes — that is the point of it. The workflow is scoped to specific approved sources, such as product documentation, approved FAQs and brand guidelines, agreed at the start rather than pointed at everything a business has.
Can it check claims against our source material?: Yes, and the sentence worth worrying about is never the obviously wrong one. It is the plausible line nobody thought to question — which is exactly the kind a reviewer skims past, and the reason the checking is done per sentence rather than per document.
Do you use confidence percentages?: No. A number nobody can verify is not a control, and it tends to make people read less carefully rather than more. Supported, needs review and not in source are states somebody can act on.
Can people review content before it is published?: Yes, and the goal is to make review short rather than optional. A reviewer handed a marked draft is deciding about three sentences; a reviewer handed an unmarked one is re-reading everything, which is how review quietly becomes a rubber stamp.
Can different people have different approval rights?: Yes, and it is worth separating two things people conflate: who may produce a draft, and who may make it public. Opening the first widely while keeping the second narrow is usually what a business actually wants.
Can one source produce several different outputs?: Yes — a website announcement, a customer email, a sales note and social source copy can all run from the same approved facts, each written for a different job and depth. That is where a system earns its cost.
What happens when the source material changes?: You get a list of what to re-read, not an automatic correction. Deciding a published sentence has gone stale is a judgement — a changed price makes some pages wrong and leaves others fine — so the system finds the candidates and a person rules on them.
Can it help update old content?: Yes, and it is often the fastest win. Existing content can be checked against current source material to find the statements that have quietly gone out of date.
Does this replace writers and editors?: No, and the workload tends to move rather than shrink: less time producing a first draft, more time on the decisions that used to get made hurriedly at the end because everyone was out of budget by then.
Can our existing content be used as source material?: Often, yes — but only the parts that are still accurate. Working out which of the existing material is safe to treat as a source is usually one of the first things a project does.
Can the workflow connect to our CMS or docs?: Usually, and it is checked first rather than assumed — the limit is normally what a system will let you read out, not what it will let you write in. Whatever cannot be reached gets named up front rather than discovered mid-build.
Will it produce a large volume of content quickly?: It will produce drafts quickly. Whether that becomes a large volume of published content depends on your review capacity, and we would not design a workflow whose main achievement was outrunning it.
What determines the scope of an AI content project?: The state of your source material, more often than anything technical. A team whose facts live in one maintained place is a contained build; a team whose facts live in six documents that disagree has a different project in front of it, and it is not really an AI one.
What do you need from us?: One thing above the rest: a recent piece of content that went out slightly wrong, and the story of how. That tells us more about where the workflow actually fails than any brief, because it shows which check was missing rather than which was intended.
Who owns the workflow and the content it produces?: You do. The part worth checking with any supplier is the configuration — the voice rules, the prompts, the review logic — because that is where the work actually accumulates, and it is handed over rather than held.
What happens after it is built?: The first weeks usually change it. Real jobs surface source material that is not as approved as everyone thought, and rules that need sharpening. How much of that is included is agreed per engagement.

Start here

Bring us the content your team keeps rewriting.

One repeatable job is a better starting point than a platform.

What your team writes repeatedly: The job that comes round every month
Where the source facts live: Docs, a wiki, somebody’s head
What the brand should sound like: Even if it is only adjectives so far
Who reviews and approves today: By name, not by department
Where the final content has to go: CMS, email, a deck, all of them

Content Strategy, structure and writing AI strategy Where AI is worth using AI agents & automation Agents across operations Selected work What we have made