01 / The AI has to guess

Ask your AI for net sales. It has seven numbers to pick from.

Nothing in your systems tells the AI which one you mean. So it picks one, and it can pick a different one next time. Try it on one order.

What were the net sales on this order?

Order 10482Berlin, Germany · prices include 19% VAT

Waxed field jacket119.00
Organic cotton T-shirtReturned29.00
10% off · WELCOME10−14.80
Shipping4.90

Paid by customer€138.10

Includes 19% VAT22.05

Not on this receipt

Product costNot in your store?
Shipping and the returnNot in your store?
Payment feesNot in your store?
MarketingNot in your store?

Contribution margin?

Example order

Order 10482Berlin, Germany · prices include 19% VAT

Waxed field jacket119.00
Organic cotton T-shirtReturned29.00
10% off · WELCOME10−14.80
Shipping4.90

Paid by customer€138.10

Includes 19% VAT22.05

Which number would your AI report?

Click a number to see what it counts

Not on this receipt

Product costNot in your store?
Shipping and the returnNot in your store?
Payment feesNot in your store?
MarketingNot in your store?

Contribution margin?

And none of these numbers is profit.

What this order actually earned depends on four costs your store doesn't have.

Your AI can't know which number is right. It picks one.

All seven are real sums from the same receipt, and it can explain every one of them. Nothing in the data says which one your business means, so the answer depends on the chat. In Dema the metric is defined once, so the AI gets the same number every time: in the app, the API, the MCP server and the agent.

See why your AI has to guess

02 / Why it guesses

Your AI doesn't know your words. It only sees field names.

Ask for ROAS, CM3 or “this product” and the AI has to work out which fields you meant, in every chat. Pick a word your team uses every day and watch it choose.

+ 200 more metrics in Dema

Claude + MCPs

7 fields could make “ROAS”

Sales from one system, spend from another. The AI picks the pair, in every chat.

Meta · its own ROASpurchase_roas
Meta · sales it claimsaction_values
Google Ads · sales it claimsmetrics.conversions_value
Shopify · store salestotal_price
GA4 · tracked salespurchaseRevenue
Meta · spendspend
Google Ads · spendmetrics.cost_micros

Each platform claims the same order, and store sales divided by ad spend gives a different number again. None of them counts returns or costs.

In each chat: Fetch, compare, pick
Kept working by: Whoever wrote the prompts

Claude OpenAI etc.

1 metric, already defined

Defined once. The AI asks for it by name.

Metric · 1 of 200+ROAS

Ad spend against profit, with attribution calibrated by tests.

The same definition sits behind the app, the API, the MCP server and the agent.

In each chat: 1 query to a defined metric
Kept working by: Dema

Claude + your own semantic layer

1 model, yours to run

Your team builds it, runs it and pays for the compute.

Your warehousemodels/roas.sql
The mappingWritten and kept right by your data team
The pipelinesFixed whenever an API changes
The computeSnowflake, BigQuery or Databricks, billed as it runs

A good route if you have the team for it. Dema reads from Snowflake, BigQuery and Databricks, so a warehouse you've built isn't wasted.

In each chat: 1 query to your model
Kept working by: Your data team

See the steps behind one question

03 / Ten steps or one

Same model, same question. Ten steps, or one.

Ask something that touches ads, orders and costs. On live APIs the model has to fetch and calculate its way to the answer, every time. On Dema that work is done before you ask.

+ MCPs

0 of 10 steps

Which campaigns made money after returns and costs last week?

Thinking…

Used a toolMeta Ads: Fetch spend per campaign
Used a toolGoogle Ads: Fetch spend per campaign
Used a toolShopify: Fetch last week's orders
Worked it outConvert SEK, NOK and DKK to EURPicks a rate
Worked it outRemove VAT, spread discounts and shippingPicks a definition
Used a toolShopify: Fetch edits, cancellations and refunds
Used a toolGA4: Match campaigns and sessions to ordersPicks whose credit
Used a toolNetSuite: Look up product costs
Worked it outApply fulfilment and payment feesEstimates the rules
Worked it outEstimate returns that haven't arrivedGuesses a return rate

An answer after 10 steps. All ten are redone the next time anyone asks.

etc.

0 of 1 step

Which campaigns made money after returns and costs last week?

Thinking…

Used a toolProfit per campaign, last week

An answer after 1 query. The same one the next time anyone asks.

Already done before you asked

Ad spend stored daily
Orders cleaned and deduplicated
Currencies converted
VAT, discounts and shipping handled
Campaigns matched to orders
Product costs applied
Fulfilment and payment fees applied
Expected returns included

Test your current setup

The AI readiness test for commerce

Five questions to ask your AI. Getting an answer isn't the same as passing: check each one against finance.

Add this to each question: Verify every number against the source data. Don't estimate and don't make any number up. If you're missing data or don't know how to calculate something, tell me instead of answering.

Net sales
“What were our net sales last month, excluding VAT and accounting for returns, by market?”
Every word is a definition: which VAT, which returns, which exchange rate and which countries make a market.
- ☐ VAT removed per country, including on shipping
- ☐ One exchange rate rule
- ☐ Returns linked to the original order
- ☐ Your market definitions
- ☐ Matches finance
Profitability
“Which campaigns generated positive contribution margin last week after product costs, fulfilment, payment fees and expected returns?”
Costs aren't in the store or the ad platforms, and most of last week's returns haven't arrived.
- ☐ Product-level cost data
- ☐ Ad spend from every platform
- ☐ A stated attribution approach
- ☐ Expected return rates
- ☐ One margin definition
Historical stock
“Which products did we spend advertising money on while they were out of stock last month, and how many website visits and product impressions did they get while sold out?”
Most systems show stock on hand now, ad platforms identify products differently from your store, and visits and impressions per product need your own site tracking.
- ☐ Stock on each past day
- ☐ Product IDs matched across store, ads and site
- ☐ Daily product-level ad spend
- ☐ Site visits and product impressions per product
- ☐ Warehouse and market availability
- ☐ Aligned dates and time zones
Customer value
“Which products attracted the most profitable new customers over the last six months?”
"New" depends on how far back you look, and six months of repeat orders after returns needs full customer history.
- ☐ Customer identity across emails
- ☐ A stated new-customer lookback
- ☐ Full purchase history
- ☐ Profit after returns
- ☐ Acquisition cost
- ☐ Cohorts old enough to judge
Incrementality
“If we increase Meta advertising spend by 20%, what is the expected incremental profit?”
Platform ROAS isn't incremental, and returns flatten as spend grows. This needs a model, not a query.
- ☐ Incremental impact, not platform ROAS
- ☐ Diminishing returns
- ☐ Spend and response history
- ☐ Profit, not revenue
- ☐ Stated uncertainty

Questions tested: 0 / 5

https://www.dema.ai/ai-readiness-test#challenge

06 / Choose your architecture

Three valid approaches. Different trade-offs.

None of these is always wrong. Pick the one that fits the questions you need answered, and how often.

Dimension	+ MCPs	+ your own layer	etc.
Initial setup	Often fast	Requires engineering	Integration and configuration
Simple exploration	Strong fit	Strong fit	Strong fit
Shared business definitions	You supply them in prompts	You own them	Built in, configurable
History	Depends on what each system keeps	You build it	Stored daily
Profit modelling	Needs cost logic from somewhere	You build it	Cost rules and returns built in
How the AI gets a number	Coordinates several sources	Depends on your architecture	Asks for a defined metric
Customisation	Prompt and tool dependent	Maximum control	Configurable within the product
Maintenance	Connectors, prompts and logic	Full infrastructure and models	Dema, plus your configuration
Marketing measurement	Needs separate models	You implement the models	MMM, geo tests, calibrated attribution
AI models	Claude, or the one you connect	Claude, or whatever you wire up	Claude, OpenAI or open-source models, or bring your own AI
Personal data (PII)	Raw customer records reach the model as the APIs return them	You decide what the AI can see	The AI gets computed metrics. GDPR compliant, with a DPA
Where it runs	Your AI provider's region	Your cloud	EU or US, your choice

Claude + MCPs is enough when

Questions are exploratory
Only a few systems are involved
Results are easy to check
Getting the same answer twice doesn't matter much
You don't need history or modelling

Your own layer fits when

You have a mature data team
Your business rules or data models are unusual
You need full control of the infrastructure
The priority is statutory reporting that ties to the ledger
You already run a strong analytics platform

Dema + the best models fits when

Ecommerce metrics drive decisions
Profit after returns and costs matters
You need product and customer history
Reports have to repeat and agree
You want MMM and incrementality
Your engineers have better things to build

You don't have to replace your stack. Dema reads from Snowflake, BigQuery and Databricks, and works as the intelligence layer behind the AI tools you already use.

Three ways to put AI to work on your commerce data