An AI growth agency is a growth service in which software does most of the reading, analysis and drafting across a company’s marketing and revenue data, and people approve what goes out. A real one shows its working: the sources behind each finding, how confident it is, what it drafted, and who approved it.
The term is new, and new terms attract relabelling. Some firms that use the words are built differently from a traditional agency. Others are traditional agencies whose staff now use AI tools, which is sensible and common, but is a different thing with a new name on it.
This note gives two short definitions, then the four things a real one should be able to show you, the red flags of a relabelled one, and a checklist you can take into a first call. We use the term for ourselves, so we have tried to define it in a way that would hold up if a competitor had written it.
What does “AI growth agency” mean?
Two definitions: one short enough to remember, one precise enough to check.
The short definition: an AI growth agency is a growth service where software does most of the work of reading data and drafting the next piece of work, and people decide what ships.
The precise definition has three parts, and each one rules something out. It is a service, so it is accountable for doing the work, not just for supplying a tool you operate yourself. It is growth work, so it spans the path from a person first reaching you to that person paying, not one channel in isolation. And the software is the main worker, not an accessory: it reads the sources, finds the patterns and writes the drafts, at a volume a team of people could not match by hand.
That makes it different from two neighbours it is often confused with.
- A traditional agency that uses AI tools internally. Most agencies now do. The people still do the reading and deciding, and the tools make them faster. That can be excellent work. It is not a different kind of service.
- An AI marketing tool. Software you buy and operate yourself. You do the deciding, you own the output, and nobody is accountable for the result but you.
Neither neighbour is worse. They are different things, and a buyer deserves to know which one they are paying for.
What should a real AI growth agency show you?
If software does the reading, the reading can be shown. A real one shows it without being asked.
A traditional agency’s reasoning lives in people’s heads, and the report is a summary of it. When software does the reading, there is no reason for the reasoning to be hidden. Every finding was computed from specific data, and that data can be named. So the bar should be higher, not lower.
Four things, for every finding and every piece of work.
- Its sources. Which systems the finding was read from, over which dates. “Read from your email tool and your calendar, the last thirty days” is a source. “Our analysis shows” is not.
- Its confidence. How sure it is, and why, with the sample size printed next to the number. A pattern seen in six people should say six.
- What it drafted. The actual draft, the email or page or brief, not a line in a report saying a campaign was prepared.
- Who approved it. A named person on your side, and when, for anything that reached a customer. If nothing needs approval, nobody is accountable.
There is a fifth thing, which is a property of the findings rather than a field on them. Patterns in marketing data are correlational. The people who read a guide and the people who bought may overlap for reasons the guide had nothing to do with. A real service says so, in plain words, and treats its findings as good reasons to try something, not as proof that something worked.
If nothing needs approval, nobody is accountable.
What are the red flags of a relabelled agency?
The word on the door changed. Check whether anything behind it did.
None of these proves bad faith. Each is a reason to ask a harder question.
- The deliverable is the same monthly deck agencies have always sent, with new vocabulary in the headings.
- Nobody can tell you which of your data sources a given finding was read from, or the answer is “our proprietary model”.
- Results are stated as causes, that a campaign drove revenue, with no method and no mention of what else changed that month.
- A score or index is presented with no working behind it: no inputs, no sample, no date.
- Autonomy is the selling point. “It runs itself” and “no approvals needed” mean nobody on your side sees what goes out in your name.
- The access request is vague. A real service can list exactly which systems it will read, and why each one is needed.
- Case results are quoted with no way to check them, and no one will say where the numbers came from.
The most useful single test comes from the score flag: ask for a score’s working. A service that built the number can show it in a minute. A service that cannot has either not built it or would rather you did not look.
How do you check one in a first call?
Seven questions, and what a real answer tends to sound like.
Take this into any first call with a service that calls itself an AI growth agency. The point is not to catch anyone out. It is to find out what you would be buying before you sign.
| Ask | A real answer sounds like | A relabelled answer sounds like |
|---|---|---|
| Which of our systems will you read? | A named list, with the reason for each | “Everything, it all feeds the model” |
| Show me one finding traced to its source | The finding, the systems and dates it came from, the sample size | A chart with no source, or a promise to send it later |
| How confident are you in that finding? | A level, a reason, and the count it rests on | “Very”, or a percentage with no sample |
| Is it a cause or a correlation? | Correlation, and what would test it | “It drove the result” |
| What do you draft, and who approves it? | Show a draft; a named person on our side approves before anything is sent | “It runs automatically” |
| What does the software do, and what do people do? | A clear split, with people on judgement and approval | Vague, or the people do everything |
| What happens to our data if we leave? | A plain answer on deletion and export | An answer that changes the subject |
Two or three weak answers is not unusual in a young category. A service that cannot trace one finding to its source, in the call, is the one to walk away from.
What do AI assistants say when you ask for one?
The assistants are as unsure where the category’s edges are as buyers are.
In twelve answers we collected on 29 September 2026 (ChatGPT, Claude and Gemini, with and without web search), to the question of which AI growth or marketing agency suits a B2B SaaS company, the answers mixed two different things. The names that came up most were software tools: HubSpot in four of the twelve answers, Clay and Jasper in three each. When search was on, the sources cited included several agency websites alongside the sites of software vendors.
That mix is the category problem in miniature. Asked for an agency, the assistants reached for tools. Asked to cite something, they often cited agencies. Neither the assistants nor the market have settled what the term covers, which is one reason a working definition is worth writing down.
Does the label matter if the work is good?
The label does not matter. What it lets you check does.
A traditional agency doing excellent work is a better buy than an AI growth agency doing poor work. Nothing in this note argues otherwise. If you have a good agency that reports honestly and traces its claims, keep it.
What the label should buy you is a higher standard of evidence. When the reading is done by software, the sources, the confidence and the drafts all exist as records. A service that has those records and does not show them is choosing not to. That choice, more than the word on the door, is what to judge.
An AI growth agency is not defined by the AI. It is defined by what it will show you: where each finding came from, how sure it is, what it drafted, and who said yes.
See the use case for an agency for franchise brands: How does an agency sell AI search visibility to franchise brands under its own name?
Questions people ask next
What is an AI growth agency?
A growth service in which software does most of the reading, analysis and drafting across a company’s marketing and revenue data, and people approve what reaches customers. A real one shows the sources, confidence, drafts and approvals behind its work.
How is an AI growth agency different from an agency that uses AI?
In an agency that uses AI, people do the reading and deciding and tools make them faster. In an AI growth agency, software does most of the reading and drafting, and people concentrate on judgement and approval. Both can do good work; they are different services.
Should an AI growth agency send campaigns without approval?
No. Anything that reaches a customer in your name should be approved by a named person on your side. A service that sells the absence of approval is removing the one check that catches a wrong fact or a wrong audience.
Can an AI growth agency prove what caused revenue?
Usually not from marketing data alone. Patterns across tools are correlational. A good service says so, and treats a finding as a reason to test something rather than as proof.
What is the quickest way to vet one?
Ask it, in the first call, to trace one finding back to the systems and dates it was read from, with the sample size. A service that built its findings can do this in minutes.
Does Throughline meet its own definition of an AI growth agency?
It is built to. Findings carry their sources, sample and confidence, drafts are staged for review rather than sent, and a named person on your side approves anything that reaches a customer. Ask it to trace one finding to its source in the first call.