October 8, 2026AI & Automation
The most useful way to think about AI in analytics is not as a replacement for the analyst. It is a change in where the analyst spends time. Routine translation and drafting get faster; judgement, validation and communication become the differentiator.
What changes
1. The distance between a question and a first query shrinks
Natural-language interfaces and code assistants can turn "show me monthly churn by plan tier for the last 18 months" into a working SQL draft in seconds. That is genuinely useful. It is also only a draft. The analyst still has to confirm how churn is defined, which plans count, how cancellations and downgrades are treated, and whether the join duplicates customers.
2. First-pass exploration gets cheaper
AI can summarise column profiles, suggest plausible distributions, generate boilerplate charts and highlight missing values. This compresses the mechanical part of exploratory analysis. It does not tell you which anomaly is a data pipeline bug and which is a real change in customer behaviour.
3. Documentation and communication accelerate
Drafting a methodology note, turning bullet points into a stakeholder summary or rewriting a technical finding for an executive audience are tasks where AI saves real time. The analyst remains responsible for whether the summary is accurate and whether it reflects the decision the business actually faces.
4. Repetitive workflows become automatable
Recurring reports, data-quality checks and simple classification or routing tasks can be wrapped in automated pipelines. This is often where the largest return sits — not in a flashy model, but in removing a weekly manual step that nobody enjoys and everybody mis-executes eventually.
What stays the same
1. Defining the decision
Every analysis should start with a decision, not a dataset. Who will act on the result, what action is available, and what would change their mind? AI cannot supply that context. A fast answer to the wrong question is still the wrong answer.
2. Data quality and lineage
Models, including language models, are downstream of your data. If a CRM field is populated inconsistently or a nightly job silently drops records, AI will reproduce the problem with more confidence and better prose. Data contracts, ownership and lineage remain foundational.
3. Validation
AI-generated code can be subtly wrong in ways that still run. A join that looks reasonable may fan out rows; a filter may exclude the exact cohort that matters; a chart may aggregate at the wrong grain. The analyst has to test assumptions, reconcile totals and reproduce results independently.
4. Causal thinking
Prediction is not explanation. A model can rank customers by churn risk without telling you whether a retention offer will work. Distinguishing correlation from causation, and knowing when an experiment is required, is still human work.
5. Trust and communication
Stakeholders do not adopt a number because it is statistically elegant. They adopt it when they understand where it came from, what it assumes and what it does not cover. That translation is a relationship skill, not a generation task.
A practical workflow
- Frame the decision and agree the metric, population and time window with the business.
- Inspect the data before asking AI to analyse it. Confirm grain, keys, missingness and known pipeline issues.
- Use AI for drafts — queries, code, charts, summaries — and mark every output as unverified.
- Validate independently with row counts, totals, spot checks and a second method where possible.
- Explain the result in the language of the decision, including limitations and uncertainty.
- Monitor after delivery so the analysis does not quietly become stale.
Risks worth naming
- Hallucinated numbers. Language models can produce plausible figures that were never computed. Require every number to trace back to a query or a file.
- Data leakage. AI-assisted feature engineering can accidentally include post-outcome information. Temporal validation is non-negotiable.
- Privacy and confidentiality. Pasting customer data into public tools may breach policy or law. Use approved environments and minimise data.
- Over-trust. Fluent output feels correct. Teams need review habits and, increasingly, evaluation sets.
- Reproducibility. If the prompt is the pipeline, version it. Otherwise nobody can explain last quarter's numbers.
What to build now
The durable skills are unchanged: SQL, statistics, data modelling, domain knowledge and clear writing. What is newly valuable is the ability to evaluate AI output, design context and retrieval, and govern automated workflows. Analysts who can do both — ask a rigorous question and use AI to answer it faster — will be the ones teams rely on.
The technology changes the pace. It does not change the standard.
