AI reporting tools: what they can and cannot tell you about your marketing

AI reporting tools come in four kinds: a general AI assistant you hand a spreadsheet to, AI features built into business intelligence tools, automated insights inside ad platforms and agency dashboards, and an assistant that sits on a joined model of your sales, ad and finance data. All four can summarise what happened and write it up in plain English. Only the last can tell you which brand, channel or agency made gross profit, because that answer depends on data that already ties every sale to the spend that bought it.

The four kinds of AI reporting tool

Most lists of the best AI tools for reporting mix these together. They are different products for different jobs, and the difference that matters is not the AI model. It is what data the AI can see.

AI chat on a spreadsheet

This is where most AI data analysis starts, and for a single clean table it works well. ChatGPT can analyse uploaded spreadsheets, summarise trends and build charts, and OpenAI's own guidance asks for structured data with clear column names and one record per row (OpenAI Help Center). That guidance is the catch. A Google Ads export, a Meta export and a CRM extract use different dates, names and definitions. The assistant will answer whatever you ask, but it cannot know that two rows in two files are the same customer.

AI inside a BI tool

Copilot in Power BI can summarise report pages, answer questions about the data and help authors build new pages. It needs a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher), so a Pro licence alone is not enough. Microsoft is candid that the semantic model must be prepared first; without that, Copilot "can misinterpret the data and return generic or inaccurate results" (Microsoft Learn). Google offers conversational analytics, powered by Gemini, in Data Studio (called Looker Studio until April 2026) and Looker, and advises users to validate all output because it can look plausible and still be wrong (Google Cloud documentation).

These are good tools. If you already have a well-built Power BI model of your sales and marketing, Copilot is a sensible next step. The point is that the AI sits on the model, and the model is the work.

Insights in ad platforms and agency dashboards

Ad platforms and reporting dashboards increasingly add automated insights and AI-written summaries. They help you spot a change inside one channel quickly. But each judges performance on its own conversions, and an agency dashboard covers the channels that agency runs. Neither sees your order system or your margin, so neither can tell you whether the sales it reports were real, or claimed by another platform as well.

An assistant on joined data

The fourth kind starts from the data rather than the chat window. Ad platforms, analytics, CRM, sales or bookings and finance are joined into one model first, with duplicate sales removed and margin applied, and the assistant answers from that. Questions can then cross brands, channels and agencies, and the answers come back judged on gross profit rather than on each platform's own conversions.

What AI reporting tools answer well, and what they cannot

The quickest way to choose is to test the questions you actually ask. Here is how each kind copes with typical ones.

The pattern is plain. AI handles the first three rows on almost any data. The last three need the data underneath to join spend to sales and margin, and no amount of prompting fixes data that is not joined.

Why the answer is only as good as the joined data underneath

An AI assistant reads numbers; it does not audit them. If Meta's export says 300 leads and the CRM recorded 180, the assistant repeats whichever file it was given, fluently and with confidence. Three things go wrong most often:

• Double counting. Google and Meta can both claim the same sale, so adding platform conversions together overstates results. Our guide to cross-channel attribution explains why.
• Missing margin. Platform ROAS is built on revenue. Without each brand's or product's gross margin, an AI ranks channels on revenue and can favour the one selling low-margin lines.
• Broken tracking read as performance. When a tag or form stops sending conversions, the platform's results fall and its cost per result rises. An assistant reading the platform alone sees a channel getting worse and suggests new creative or a budget cut.

That last one is the story in Orion's demo, on sample data. One site's new booking form stopped Meta's Conversions API Lead event firing, so Meta was optimising blind and cost per booking rose. Read in Meta alone, it looked like a channel in decline. The cause showed up only by comparing what Meta reported with what the CRM recorded.

Disconnected data is the normal state, not the exception. Only 26% of marketers in Salesforce's 2026 State of Marketing survey of 4,450 marketing decision makers said they were completely satisfied with their data unification (Salesforce, 2026). Yet Gartner's 2026 CMO Spend Survey found CMOs allocate 15.3% of marketing budgets to AI initiatives (Gartner, 2026). Money spent on the AI layer goes further when the data layer comes first.

A worked example: one question, two answers

Say an ecommerce business spends £18,000 a month through two agencies: £10,000 on Google Ads and £8,000 on Meta. It asks an AI assistant which channel works best. The figures are made up for the example.

Given only the two platform exports, an AI tool will rank Google first, because its ROAS is higher. Given the joined data, the same question gets the opposite answer: Meta leaves more than four times as much gross profit after its cost. The model did not get smarter between the two answers. The data did. Our guide to what a good ROAS is covers the arithmetic in more depth.

How to choose an AI reporting tool

Whether you are choosing for a growing business or picking enterprise AI tools for advanced reporting, the same checks apply:

1. Write down the ten questions you ask most. Test every tool on those, not on its own demo questions.
2. Check what it can see. Ad platforms only, or also your CRM, order or booking system and finance? If it cannot see sales and margin, it cannot answer profit questions.
3. Check that it shows its working. Every answer should name the sources and date range behind it, so a finance director can check it.
4. Check the definitions are fixed. What counts as a sale, and how gross profit is calculated, should be set once in the data, not reinvented in each answer.
5. Check who acts. Suggestions are cheap. Changes to budgets and bids should wait for a person to approve them.
6. Check where your data goes and what it costs. Read the provider's data processing terms, and count every licence or capacity charge; Copilot in Power BI, for example, needs a paid capacity, not just a Pro licence.

Board reporting automation: let AI write the words, not the numbers

The safest use of AI reporting in board and monthly packs is narrative. The numbers come from a fixed, reconciled data model; the AI drafts the commentary around them; a person checks it and signs it off. In practice:

• Every figure in the report is calculated by the data model, never typed by the AI.
• The AI is given those tables plus last month's report, and asked to explain what changed and why.
• It is told to flag anything it cannot explain, such as a gap between platform and recorded sales, rather than smooth it over.
• A named person reviews the draft before it reaches the board.

What the report should contain in the first place, from spend to gross profit to next month's moves, is covered in our guide to marketing reporting across several agencies, brands and ad accounts.

When each kind is enough

Be honest about scale. If you run one brand with one agency, and your sales system already records where each customer came from, AI chat on a clean export or Copilot on a well-built Power BI model may be all you need. The case for an assistant on joined data grows with the number of brands, sites, agencies and ad accounts, because that is when the platforms' numbers overlap and the spreadsheets multiply.

That is the situation Orion is built for. The Intelligence build joins ads, analytics, CRM, sales or bookings and finance into one live view in your own accounts. On top of it, the platform writes reports on demand, ranks the next moves with their expected gain, and Ask Orion answers questions in plain English from the joined data. Anything it could change, such as budgets or bids, waits for a person to approve it: you approve, it runs. You can put questions to Ask Orion in Orion's demo, on sample data: a fictional group with several brands and sites and twelve sources. For what the data layer itself should look like, see our guide to business intelligence dashboards.