Can AI replace a data analyst for a small brand? Most of the typing, not the judgment

AI can do most of an analyst's pulling, joining and writing for a small brand. It can't set your definitions or own a wrong number. Here's the split, with a test.

By Justin Maddahi, Founder, Synthesis Intelligence

The short answer

For a small brand, AI can replace much of what a data analyst does by hand: pulling reports, joining them, checking totals and writing the weekly summary. It can't replace the person who decides what each number means and answers for it when it's wrong. Most small brands end up with AI doing the legwork and one owner making the calls.

Wanted: someone to pull five reports by Tuesday, tie them to the bank, explain why last month’s number changed, and be right when the owner repeats it to an investor. Pay is whatever a small brand can afford. Nobody on the team has quite the time, and half the owners reading this ad have written one like it.

So the question comes up fast: can AI just do this job? For most of the pulling, joining and writing, yes, but not for deciding what your numbers mean or standing behind them.

The split matters because it tells you what to hand over this month and what to keep. The rest of this post covers where an analyst’s week goes, what AI does well now, where it still needs you, and a test you can run in an afternoon.

Where does an analyst’s week really go?

Ask any analyst and you’ll hear the same thing: most of the week isn’t analysis. It’s getting the data into one place and making sure it’s right.

Task What it involves Who can do it now
Pull and join Export from each store, ad account and sheet, then line them up by date and product AI, once the accounts are connected
Check the numbers Tie totals to the source, spot a missing day, flag a number that moved oddly AI does the checks, a person sets the limits
Write it up Turn the tables into a weekly summary people will read AI drafts, a person edits
Decide what a number means “Revenue” means orders placed, or cash paid out? A person, written down once
Choose what to do Cut an ad, reorder stock, raise a price A person

Here’s a made-up example of a 30-hour analyst week for a small brand, and how many hours AI could take off each task.

The arithmetic: pulling and joining drops from 12 hours to 2, writing from 6 to 2, checking from 8 to 6, and the deciding stays at 4. That leaves 14 hours of the original 30. The saved hours aren’t the interesting part, though. The hours that stay are the ones where the job is really done.

What can AI do well today?

Quite a lot, as long as it can reach your real numbers. The reason is a plain standard called the Model Context Protocol, described in its own docs as an open-source standard for connecting AI applications to external systems. It’s what lets an AI app like Claude or ChatGPT look at your data instead of guessing.

Once that connection exists, a good AI can:

  • Answer a question across your store, your ads and your marketplace in one go
  • Line up two periods so a week is compared with a week
  • Work out profit after fees, cost and ad spend
  • Notice that a number moved and draft a few reasons why
  • Write the Monday summary in plain words

One operator story shows how far that goes. Katie Cirulli, who leads emerging brands at Grüns, told the Marketing Operators podcast about a member of her team. They turned their steps for diagnosing a drop in conversion rate into a Claude tool. Ask it why conversion is down today and it runs “the exact steps that I would take an hour of my day to do.” The employee also suggested an automatic alert if conversion falls below a set level. The episode gave no accuracy figures, so treat it as a sign of what’s possible, not a result.

Notice what happened there. A person wrote down the steps first, and then the AI ran them. That order is the whole story.

Where does AI still need a person?

The gaps aren’t about intelligence. They come from three things that live outside the data.

Definitions. Say you ask for last month’s revenue. Amazon alone can tell you what was ordered, what shipped and what was paid out, and those are three different numbers. Amazon’s own docs note that settlement reports can’t be requested or scheduled because Amazon schedules them itself. So “what settled last month” depends on when the report happened to land, and an AI can’t know which number you meant unless you’ve said so. Stanford’s classic guide to ontologies puts the fix simply: people and software need to share a common understanding of how information is structured. Our post on AI-ready data shows how to write those definitions down.

Checking against the real world. A model can spot that a total looks off. It can’t call your 3PL to ask whether a pallet went missing, and it can’t know a number is wrong because you changed suppliers last Thursday.

Owning the answer. When a number goes into a bank deck or a sale process, a person has to stand behind it. That isn’t something you can hand to software.

The makers of Claude give similar advice in their guide to building agents. They recommend finding the simplest solution possible and adding complexity only when needed. They also say agents add the most value where there are clear success criteria and meaningful human oversight. For a small brand, that’s a good rule of thumb: keep the setup simple and keep a person in the loop.

Why I care about the translation part

I started building Synthesis partly because of something I saw inside my own brand. People got upset at each other over differences of opinion, and I came to think that a lot of them had the right idea but couldn’t show it with analysis. Somebody, or something, has to translate between an idea and the numbers that prove it.

That’s a job an analyst has always done, and it’s the part AI does best when someone has written the numbers down clearly. If your data is clean and defined, an AI is a quick, tireless translator. If it isn’t, it’s a fast way to be confidently wrong. Our post on why AI gets your ecommerce numbers wrong is worth a read before you start.

Run this test: five questions you already know

Before you let AI stand in for anyone, find out how it does on your own data. It takes about an hour.

  1. Pick five questions where you already know the right answer. Make them different kinds, such as: - last month’s net sales - your best product by profit - ad spend per new customer - days of stock left on your top item - how many orders were refunded
  2. Ask each one in Claude or ChatGPT, connected to your real accounts. Don’t coach it.
  3. For each answer, write down whether the number matches yours.
  4. Also write down whether it told you its basis: the dates, the kind of sales, and what it left out.
  5. Ask the same five again tomorrow, worded differently, and compare.

What a pass looks like: at least four of five match, every answer states its basis, and tomorrow’s answers agree with today’s.

What a fail looks like: a number that’s close but not right with no explanation, a basis that changes between asks, or a confident answer to a question your data can’t actually answer. Any of those means the definitions aren’t written down yet, and that’s the thing to fix before you rely on it.

A missed question isn’t a verdict on the AI. It usually shows exactly which definition is missing.

When should you still hire an analyst?

AI isn’t the right answer for every brand, and here’s where you’d still want a person:

  • You need custom models, such as forecasting demand by region or testing price changes properly
  • You run experiments that need careful design and a real read-out
  • The data is messy enough that cleaning it is a full-time job
  • Investors or lenders want someone to sign off on the numbers

If none of those fit and your questions are mostly “what happened, why, and what should I look at,” AI on top of connected data covers a lot. It doesn’t have to be one or the other. An analyst who uses AI spends less time on legwork and more on questions only a person can ask.

Synthesis is built for the first case. It writes your business down once so every number means one thing, and it keeps the AI you already use on those numbers. There’s no chat of its own, since you ask in Claude, ChatGPT or another AI app. It’s the wrong tool if what you need is a bespoke statistical model.

A split that works for a small brand

Here’s a simple way to divide the week once your data is connected and defined.

  1. Let AI do the Monday pull, the joins and the first draft of the summary.
  2. Keep the definitions in one place, written by a person, and change them on purpose.
  3. Ask for the basis every time, so “sales” always comes with its dates and its kind.
  4. Spot-check one number a week against the source, and write down what you found.
  5. Keep the decisions with a person, and let the AI bring options, not orders.

The short version

AI can take over most of the pulling, joining, checking and writing that fills an analyst’s week. It can’t set your definitions or own a wrong number, and that stays with a person.

For most small brands the best setup is one owner who makes the calls and an AI that does the legwork on well-defined, connected data. Test it with five questions you already know before you trust it, and if it fails, fix the definitions first. You can read more about how we think about this in what Synthesis is.

Questions people ask

Can AI replace a data analyst?

It can replace a lot of the routine work, such as pulling reports, joining them and writing a summary. It can’t decide what your numbers should mean or take responsibility for a wrong one, so for most brands a person still owns the calls.

Should a small brand hire a data analyst or use AI?

If your questions are mostly “what happened and why,” start with AI on top of clean, connected data. Hire an analyst when you need custom models, deep experiments or someone to own the data full time.

What can AI do with my Shopify and Amazon data?

Once your accounts are connected, it can answer questions across them, line up periods, calculate profit and draft a summary. It needs your definitions written down first, or it will pick its own.

What are the risks of using AI instead of an analyst?

The main risk is a confident answer built on a choice you never saw, like which date range or which kind of sales it used. You reduce that risk by writing definitions down and asking the AI to show its basis.

How do I test whether AI can do my analyst's job?

Pick five questions you already know the answer to. Ask the AI, compare each answer with your known one, and check whether it told you its basis. Four right with the basis shown is a good start.

Does an analyst still help if I use AI?

Yes. A good analyst writes the definitions, runs the experiments and catches what the AI can’t see. With AI doing the legwork, they can spend more of the week on those.

Synthesis · Your Brand Brain

Those numbers were made up. Yours are waiting.

Every product, channel, customer and rival, and how each moves the others. Built from your own data the night you connect, then rebuilt every week.

Free for the first month. Then $999 a month, your whole team included.

One brain. Five ways in. Go on, press them.

Ask anything
Email, more than Meta. Its sales follow your email sends, tested against chance. A skipped send costs about $9,000.
Read from Klaviyo · Amazon · Meta

Answers in your AI

Ask in Claude or ChatGPT. Real numbers, with the sources.

Plugin

Build in your voice

Pages, ads and decks in your brand's own look, inside Claude Code.

Brand Brain

Money moves with what you sell and what it cost you to sell it.

A living map

Tap an area. Rebuilt every week from your own data.

Reports
Your week · Monday
Sales$412K +6%
Cost per order$18.40 +9%

Monday, handled

What moved, why, and what to do. In your inbox, with a live dashboard behind it.

Automations
Pause the ad set that keeps losingMeta · Active → Paused
Tag lapsed buyers for win-backShopify · 2,140 people

You approve. It acts.

It makes the change in Meta, Amazon or Shopify, and checks that it landed.

Examples shown. Yours are built from your own data.

$999a month after your free first month
  • 49 tools connected with you
  • Live in about 2 hours
  • Your whole team included
  • No annual contract. Cancel any time.
Book a demo