What is Nrth Star? AI automation for agencies, explained

Nrth Star is automation software for marketing agencies. It audits how your agency runs and ranks where the hours go, writes each automation as a plain-English plan you approve, replays that plan against your last 30 days with everything switched off so you can see what it would have done, then hosts it and tells you when something breaks.

That is the product in a sentence. The rest of this is why it is built that way, because the shape is unusual and the reasons are the interesting part.

The problem it starts from

Ask an agency owner what they want to automate and the honest answer is usually some version of “I am not sure, but last month felt busy.”

They are not being vague. They genuinely cannot say whether the time went on onboarding, on chasing quotes that nobody answered, or on rebuilding the same client report for the ninth month running. Nobody tracks it, and the people who could tell you, the account managers and the coordinators doing the work, are not usually the people in the conversation.

Every automation tool on the market opens on an empty editor and asks what you would like to build. That is a reasonable question to ask a developer and a useless one to ask an agency owner, and it produces a specific failure that nobody ever notices: you build the automation you happened to think of, rather than the one costing you the most. There is no way to tell the difference from inside an editor, so nothing ever tells you that you picked wrong.

So the first thing Nrth Star does is not build anything.

How it works, start to finish

Six steps. The order matters more than any of them individually, particularly where the connecting happens.

1. The audit. A conversation about how the business runs, plus a short anonymous survey your team fills in about last week rather than about how things go “in general”, because people are poor at generalising and good at recall. It comes back as a ranked report of where the hours go, with a number against each one. One person saying reporting takes ages is an anecdote a founder discounts. Most of the team independently naming it is not arguable. This step is free, connects to nothing, and needs no card.

2. The plan. You pick something off that list. Nrth Star writes back what the process looks like now, what should happen instead, and what it still needs to know. You approve it or correct it in ordinary words. There is no canvas and there are no boxes to join up.

3. Connect. Only now do you connect any accounts. One click per app, the standard consent screen, no keys to copy. That ordering is deliberate: connecting first would mean granting write access to the system your business runs on in order to find out what might be possible, which is the wrong way round.

4. Shadow mode. Before anything is live, the automation is replayed against your last 30 days of real data with everything switched off. Nothing is sent and nothing changes. You get back what it would have done, client by client and date by date. More on this below, because it is the part that does the most work.

5. Arm it. A separate, deliberate action. Nothing drifts into being live, and one press turns it off again.

6. It runs. When something eventually breaks, you are told what stopped, why, and what would fix it, and then asked. It does not repair itself behind your back.

What an automation actually looks like

This is the part most people do not believe until they see it. An automation here is not a diagram. It is a short list of English sentences:

WHEN    a deal moves to Proposal Sent in HubSpot
WAIT    3 business days
UNLESS  the client has replied or the deal has moved on
SEND    a follow-up in your voice
STOP    after two follow-ups

That is the whole thing. No flowcharts, no boxes joined by lines, no field mapping.

It reads that way because that is genuinely how it is stored, rather than being a friendly summary generated on top of something more complicated. The practical consequence is the one that matters: you can read it, so you can tell whether it is right. A diagram with fourteen boxes cannot be checked by the person who most needs to check it, and “approved” on something nobody can read is not approval.

The part that gets harder, not easier

Here is the argument behind all of this, and it is about where automation is going rather than about us.

Software that writes automations is getting rapidly better. Describing a job in a sentence and getting something that runs is close to a solved problem, and it will keep getting cheaper. If building were the bottleneck, this whole category would be finished within a couple of years.

Two things do not get cheaper at the same rate.

The first is knowing which automation is worth having. That is a question about your business, not about software, and no amount of generation ability answers it. An agency that can now build fifty automations in an afternoon has a harder problem than one that could build none, not an easier one.

The second is knowing that the thing you built is right, and this is the one that gets genuinely worse. The failure people brace for is an automation that errors. Those are the safe ones: nothing happens, you find out, you fix it. The expensive failure is the automation that works perfectly and is quietly wrong. It runs green for six weeks while it emails the wrong list, chases clients who already paid, or writes to the wrong field. Nothing alerts, because from the software’s point of view every run succeeded.

The faster automations get written, the more of that there is. So the useful question stops being can it build this and becomes how do I know it is right before it touches anything. That is the question this product is organised around, which is why the audit comes before the building and the shadow run comes before the arming.

Shadow mode, which is the point

Testing an automation normally means running one step against a sample record. That tells you the step works. It does not tell you what the whole thing would do to your actual client list, which is the only question you have at 5pm on the day you turn it on.

A shadow run replays the entire automation across your last 30 days of real history with everything switched off, and hands back four lists:

• What it would have done. Every client it would have contacted, and on what day. This is the list to read line by line, because every line is a real person who would have received a real email.
• What it would have left alone. Usually the longest list, and it carries the value it actually read each time, so you can catch a rule that is technically doing what it was told and is not doing what you meant.
• What it could not decide. Records where the history does not contain enough to say. Reported rather than assumed, because a confident wrong answer is worse than an admission.
• What is still in flight. Something that started inside the window and has not finished yet. A three-day wait on a deal that moved yesterday resolves tomorrow, which has not happened.

The question that page asks you is deliberately narrow. Not “is this automation correct”, which nobody can answer by looking, but “do any of these look wrong to you”, because you are the only person who knows that the client on line four went on retainer in March and should never have been chased.

Nothing else in this market does this, and it is the difference between “trust us” and “here is what this would have done to your business last month.”

When something breaks

It will. The systems automations sit on top of change: a field gets renamed, a permission is tightened, a version is retired. Any tool that implies otherwise is selling you something.

What differs is what happens next. Nrth Star pauses the automation rather than letting it run half-working, works out what stopped, and sends you what broke, why, and the proposed fix in plain English. Then it waits.

It never applies a repair on its own, and that restraint is deliberate rather than cautious. An automation that fixed itself incorrectly does not fail. It succeeds at doing the wrong thing to real clients, and it does so with no error anywhere to notice. The same argument as shadow mode, arriving at a different moment.

What it will not do

The honest list, because a page like this is worth nothing without one.

It will not let you tune every step. There is no editor. If you want a rule that fires only on Tuesdays for clients in one pipeline stage whose name starts with a letter before M, you want a builder, and Zapier is the better tool.

It will not connect to everything equally well. A couple of systems are supported first-hand and understood in detail. Most others connect and work, with thinner diagnostics, which mostly matters when something breaks and there is less to go on. If your agency runs on something unusual, check before you switch.

It will not promise that nothing can go wrong. It promises that you see what it would do before it does it, that it stops rather than limping, and that it asks before repairing itself. Those are different claims and the second set is the one that can actually be kept.

It will not decide every case in a shadow run. Sometimes the history genuinely does not say. Those land in their own list and are counted against the result rather than quietly rounded up.

It is not for every business. It is built for marketing agencies of roughly ten to thirty people, which is a deliberately narrow market: it means the audit already knows what a retainer is, what a proposal cycle looks like, and that reporting week is the worst week of the month. Advice that suits everybody is rarely specific enough to act on.

What it costs

The audit is free and stays free. No card, no accounts connected, and you can run it again whenever you want.

After that you pay for the number of automations you are running, not for how often they run. Not tasks, not credits, not tokens, not seats. Starter is $49 a month for 2 active automations, and the AI writing is included rather than billed separately.

This is the one place a competitor comparison is worth making, because every alternative prices in credits or tasks and every review of every one of them complains that working out next month’s bill needs a spreadsheet. Counting by automations means the number you are billed for is a number you already know, and a busy month costs the same as a quiet one.

Who it is for

Use it if you know you are losing hours and cannot say where, if nobody at your agency wants to become the automation person, or if you have a pile of half-working connections that somebody has been babysitting for a year.

Do not use it if you enjoy building this yourself and want control over every step, or if what you need is genuinely two steps long. Moving a form into a spreadsheet does not need an audit and a rehearsal, and pretending otherwise would be silly.

If you are not sure which of those you are, the audit answers it and costs nothing. It tells you where your hours are going and what each one is worth, and that is a useful thing to know whichever tool you end up using, including none. There is a longer walk-through of the whole process, and a page about who is building this and why.