AI + Analytics
How to check an AI answer about your numbers before you act on it — and why DataLens always shows its work alongside the answer.
Articles in this collection
21 articlesYour AI agent can now edit your ad account
MCP connectors let AI agents pause ads and move budgets. Why the risk is actions, not answers — and the approval gates and caps that keep an agent safe.
Prompt injection is hiding in your marketing data
Campaign tags, comments and search terms are written by strangers. How text in your data steers an AI summary, and defences that hold when the model is fooled.
ChatGPT is sending you customers, and GA4 calls them Direct
AI assistants send real visitors, many with no referrer. Measure a floor in GA4, estimate a ceiling from Direct, and narrow the range with customer answers.
Ask the same question twice, get two numbers
The same question can get different AI answers. Why ambiguity, not randomness, is usually the cause — and how agreement across runs becomes a confidence check.
Why your AI report always has a reason
AI reports give a cause for every wobble. Why models narrate noise, and the normal ranges, maturity flags and evidence rules that let them say “no change”.
Reasoning models vs fast models: which questions deserve “thinking”?
Reasoning models help on multi-step diagnosis and add nothing to simple lookups. How to route analytics questions by type to cut model cost and wait time.
Why Fine-Tuning LLMs for Internal SQL Is Almost Always a Mistake
Teams spend months fine-tuning a model on internal SQL, then it breaks the day a table changes. Why dynamic schema retrieval beats fine-tuning.
15 prompts marketers actually reuse (copy-paste library)
Fifteen copy-paste prompts for three jobs: pace the money, diagnose the move, decide the budget. Each makes the answer checkable in 60 seconds.
Text-to-SQL Isn't the Hard Part — Schema Context Is
Correct SQL syntax is the easy 20%. Knowing which of five tables holds "revenue" and how they join is what decides whether an AI answer is right.
How to Audit an AI-Generated Query in 60 Seconds
A fast checklist for scanning any AI-written SQL query before you trust its output — the joins, the filters, and the aggregation, in that order.
Why Every DataLens Answer Ships Its SQL
A product philosophy note: why we decided a number without its query attached isn't a trustworthy number, and what that decision costs us.
Bring Your Own API Key: What That Actually Means for Your Data
DataLens lets you connect your own model provider key instead of routing through ours. Here is what that changes about where your data goes.
Benchmarking an AI Analyst on Your Own 20 Golden Questions
Fluency is not accuracy and demos are not decisions. Building a verified question-SQL-answer suite that scores models, prompts and schema changes.
Anomaly Alerts That Do Not Spam: Static Thresholds vs Dynamic Baselines
Every false alarm trains readers to ignore the real one. Respecting data maturity, expecting seasonality, and the severity ladder that keeps channels trusted.
Least-Privilege Data Access for AI Tools: Views, Redaction, Row Limits
Exposing every table to an AI tool is a privacy risk and an accuracy problem. The views, tokenisation and caps that narrow access and improve answers.
What an AI Analyst Can't Do Yet — an Honest List
Eight things an AI analyst genuinely cannot do, written by someone building one. Where the ceiling is, why it is there, and which limits are shrinking.
Semantic Layers vs. Prompting: Where Business Definitions Should Live
You can put "active customer means ordered in 90 days" in a prompt, or in a semantic layer. One of those is a definition; the other is a suggestion.
How We Keep an LLM From Inventing a Column Name
A model that has never seen your schema will guess order_total and be wrong. Four layers that stop it: retrieval, context, dry-run validation, repair.
The Real Cost of an AI Analytics Question
Two costs stack on every AI-generated answer, and the smaller one gets all the attention. A line-by-line breakdown of where the money actually goes.
Scheduled AI Reports vs. Dashboards: When a Daily Narrative Beats a Chart
A dashboard waits to be opened. A scheduled narrative arrives whether or not anyone was looking. Which failure mode each one has, and how to pick.
Seven Questions to Ask Before You Trust an AI's Answer About Your Data
A syntactically valid query can still answer the wrong question. Seven concrete checks to run before you act on a number an AI gave you.