An AI Performance Analyst: Asking Plain-Language Questions of Your Data

You ask, "Why did cost per order rise last week?" A person with a spreadsheet takes an hour. An AI analyst can answer in seconds. The speed is real. So is the risk: a fluent answer built on the wrong numbers looks exactly like a right one.

This article explains what an AI performance analyst is, which questions suit it, and the checks that make its answers safe to act on.

What it actually is

It is a language model connected to your data: ad spend, orders, product costs, and perhaps delivery results. You ask in ordinary words. It turns the question into a query or a calculation, runs it on your numbers and writes the result in plain language.

The model does not know your business. It knows what you gave it and how you defined your terms. That is the whole difference between a useful analyst and a confident guesser.

Why the common approach fails

The usual first attempt is to paste a report into a chat and ask for "insights". Three things go wrong.

• The definitions are missing. "Revenue" may mean incoming orders to one source and delivered orders to another. The model picks one without telling you.
• The data is partial. A pasted export may be cut off, in another currency, or from a different date range than you think.
• The answer sounds causal. "Cost rose because creative fatigue set in" reads like a finding. Without evidence it is a hypothesis.

A model: question, answer, check

Treat every answer as a draft that has to pass a check before it becomes a decision.

• Plain-language question"Why did cost per order rise last week?"
• Your governed dataSpend, orders, costs, delivery status

↓

AI writes the answerWith the query and numbers it used

↓

You check itTotals match, period right, cause tested

An answer that cannot show its numbers is not ready to act on.

The check has three parts: do the totals match a source you trust, is the period and currency right, and is the stated cause supported or only suggested?

Questions that suit it, and questions that do not

Describe

Example: "What was spend and orders by campaign last week?"

Fit: Good fit

Compare

Example: "How did this week differ from last week?"

Fit: Good fit, verify totals

Explain

Example: "Why did cost per order rise?"

Fit: Gives hypotheses to test

Decide

Example: "Should I double the budget?"

Fit: Needs your margin and judgement

Description and comparison are where it is strongest. Explanation is useful when you treat it as a list of things to test. Decisions about money stay with you, because they depend on margin, cash and risk the data may not hold.

Illustrative example, not real results

Say last week's spend was 10,000 SAR for 100 orders, and this week's is 12,000 SAR for 96 orders. Both are invented numbers.

1. 1

   Last week

   10,000 / 100 orders

   100 per order
2. 2

   This week

   12,000 / 96 orders

   125 per order
3. 3

   Change

   Spend up 20%, orders down 4%

   +25%

Cost per order moved from 100 to 125 (12,000 / 96 = 125), a rise of 25%. A good analyst answer states that spend grew faster than orders, and then offers hypotheses: audience saturation, a creative that tired, or a slower week for demand. Each is something to test, not a conclusion. If delivered orders fell more than incoming ones, the real cost per kept order moved even more.

Trade-offs and when this changes

• Small stores. If you have a handful of orders a week, a spreadsheet is faster and easier to verify.
• Messy data. If tracking is broken, the analyst will analyse the breakage fluently. Fix measurement first.
• Sensitive data. Be deliberate about what you connect and who can ask what. Check your tool's current data-handling documentation.
• Actions. Reading data is low risk. Letting an AI change budgets is a different matter; see Safe Automation: Where a Human Must Approve.

Putting it into practice

1. Write definitions for your ten key terms: incoming order, delivered order, revenue, margin, and the rest.
2. Connect a single governed source rather than separate exports.
3. Require every answer to show the figures and period it used.
4. Keep a list of ten questions you ask every week, and test the analyst on those first.
5. Log answers that were wrong, and fix the definition or data behind them.

What you cannot know from the data

Correlation is not cause. The analyst can show that cost rose when a creative aged; it cannot prove one caused the other. Platform attribution is also incomplete, so a "channel result" is a reported figure, not a verified one. Use an experiment when the decision is large.

Next step

The AI Performance Analyst service is built around these checks. It works best on top of a dashboard that already joins ads, sales and cost, as described in How to Build a Dashboard That Connects Ads, Sales and Profitability. See also the KPIs an executive dashboard should show.