There is a lot of noise about AI in analytics. Some of it is true, some of it is a demo that falls apart on real data. After 6+ years of doing analytics work, this is my honest scorecard: where AI genuinely saves time, where it needs supervision, and where I would not trust it yet.
The Scorecard
| Task | How AI helps | How much I trust it | What still needs a person |
|---|---|---|---|
| Classifying free text (tickets, reviews, survey answers) | Sorts thousands of comments into themes in minutes | High, after testing on a sample | Agree the categories, check a sample |
| Writing SQL and code | Speeds up analysts a lot on routine queries | High for drafts, not for unchecked output | Review every query before it drives a decision |
| Data cleaning | Finds duplicates, fixes formats, standardises names | Medium to high | Keep originals, review what it changed |
| Spotting anomalies | Watches many metrics at once | Medium, depends on tuning | Set thresholds, decide what is worth an alert |
| Questions in plain English | Lets non-analysts self-serve | Medium, only with defined metrics | Own the metric definitions, audit the queries |
| Forecasting | Can beat simple methods on complex patterns | Medium, must beat a simple baseline | Compare against last year or a moving average |
| Explaining why a number moved | Suggests hypotheses | Low, treat as a starting point | Everything, because it cannot see your business context |
Where It Genuinely Saves Time
Turning text into numbers
This is the most underrated use. Most companies sit on thousands of support tickets, app reviews and survey comments that nobody reads. A language model can group them into themes ("delivery delays", "billing confusion", "missing feature X"), so you can count them and track them over time. To trust the output, take 100 comments, label them yourself, and see how often the model agrees. If it matches you most of the time, you can scale it. If not, fix the category definitions first.
Writing and reviewing queries
An analyst with an AI assistant writes routine SQL and code noticeably faster, and gets a second pair of eyes on logic. The catch is the same as for any junior colleague: it is fast, it is confident, and it is sometimes wrong. Read what it wrote before you rely on it.
Cleaning up the mess
Duplicate customers with slightly different spellings, addresses in five formats, product names typed a dozen ways. AI is good at spotting these. The rule I follow: never let it overwrite the original. Write the cleaned version to a new column, keep the raw one, and review a sample of what changed, especially anything it merged.
Watching many metrics at once
No one can watch fifty metrics daily, but a system can. Start simple: compare each day with the average of the same weekday over the last four weeks and flag large gaps. Only move to more advanced methods when the simple version misses things that matter. Most teams get more value from five well-tuned alerts than from an advanced model that fires every morning and gets ignored.
Where I Would Be Careful
Arithmetic and totals
Language models are not calculators. If you paste a table into a chat tool and ask for a total or a growth rate, you can get a plausible wrong number. Let code and your database do the maths and use AI for interpretation and wording. Any decent setup works this way.
"Why did it happen?"
AI can list possible reasons for a drop in sales. It cannot know that you changed pricing last Tuesday, that a competitor ran a sale, or that the payment gateway had an outage, unless someone tells it. Treat its explanations as hypotheses to test, not conclusions.
Small data and edge cases
Forecasting models need enough history, and a month of data for a new product will not give you a reliable prediction. If a model has not beaten a simple baseline on data it has not seen, use the baseline.
Confident, tidy, wrong stories
The most dangerous output is a well-written summary that sounds right. Numbers wrapped in a smooth narrative get believed. That is why I insist on being able to trace any figure back to a query.
Five Checks for Any AI-Generated Analysis
- Can you reproduce the number? Run the query yourself, or recompute one figure by hand.
- Check the denominator. Percentages are where most mistakes hide. Percent of what?
- Compare with a figure you already trust. Does total revenue match finance?
- Ask what would change the conclusion. If nothing could, be suspicious.
- Look at ten raw rows. It is astonishing how often that reveals the problem.
Where to Start
Pick one job that is repetitive and checkable. Good first choices: categorising customer comments, automating one weekly report, or setting up a few sensible alerts. Time the manual version first so you can show what improved, and use the method in process automation consulting: what to automate first to check it will pay off. When you are ready to put dashboards, alerts and plain-English questions on one foundation, read how to build a business intelligence engine using AI.
AI is not going to replace your analysts. It takes over the repetitive parts so they can spend more time on decisions. If you want help setting that up, see my data analytics and business intelligence service or get in touch.
Frequently Asked Questions
How does AI help in data analytics?
Mainly by classifying text data at scale, speeding up query writing, cleaning data, watching many metrics for anomalies, letting non-analysts ask questions in plain English, and improving forecasts. It works best with human review and agreed metric definitions.
Can AI replace a data analyst?
No. It automates repetitive work such as first-draft queries, data cleanup and report writing. Judgement about what matters, why numbers moved and what to do next still needs a person with business context.
Is AI-generated analysis accurate?
It can be, but it needs checking. Language models can make arithmetic errors, misread ambiguous terms and present wrong numbers confidently. Have calculations done by code or your database, and verify results against figures you already trust.
What is the best first AI use case in analytics?
Something repetitive and easy to check, such as categorising customer feedback, automating a weekly report, or setting up a few well-tuned alerts. Measure the manual version first so you can prove the improvement.
About the Author: Utkarsh Gupta is an AI, Analytics & Automation consultant with 6+ years of experience helping companies build data-driven capabilities. He works with B2C, D2C, and B2B companies across India to implement AI-powered analytics solutions. See his AI consulting services in India.