The narrative section – the commentary that interprets the numbers – is where AI saves the most time in report production and where it can do the most damage. Drafting commentary is exactly the kind of text-heavy, checkable work AI is good at, and exactly the kind of prose where a plausible but wrong sentence is most dangerous.

This guide covers how to use AI to draft the narrative section safely. It is the companion to Using AI to Draft Reports and The Human-Verified Reporting Workflow.

AI can draft the commentary. It cannot confirm that the commentary is true.

What the narrative section does

The narrative section turns data into meaning. It answers, for each movement in the numbers, what happened and why it matters. A report with data and no narrative leaves the reader to do the analysis; a report with narrative unsupported by data misleads.

That dual nature is why AI helps and why it needs a leash: the data side must be exactly right, and the meaning side must be genuinely true.

What AI does well here

Given a clean data table and a brief, AI can:

  • Describe movements accurately from supplied figures.
  • Draft a first pass of the commentary structure.
  • Maintain a consistent voice across sections.
  • Produce multiple versions for different lengths.
  • Flag where a number looks unusual or needs explanation.

This is genuine leverage. A narrative section that once took hours can be drafted in minutes, leaving human time for the part that matters: confirming that what it says is true.

Where it goes wrong

The failure modes in narrative drafting are specific.

  • Invented causes. “Revenue rose because of new customers” when the data does not say that.
  • False precision. Confident explanation of a movement too small to interpret.
  • Dropped caveats. A qualification removed in the interest of cleaner prose.
  • Over-claiming. Language that asserts more certainty than the data supports.
  • Generic phrasing. Commentary that could apply to any period and says nothing.

The first four are accuracy failures; the last is a relevance failure. Both are caught by a human who understands the business.

How to brief AI for narrative

The quality of the draft depends on the brief. Give AI:

  • The data, structured, with periods and comparisons.
  • The argument – what the report is saying this period, set by a human.
  • The constraints – do not invent figures or causes; preserve caveats; flag uncertainty.
  • The voice – the report’s register, and any style guide rules.

A brief that states the argument and forbids invention produces a far safer draft than an open “write the commentary” request.

Verifying the narrative

Narrative verification is a specific task: check claims, not prose.

  • Check every figure against the frozen data.
  • Check every cause with someone who knows – the reason a number moved is a fact, not a draft.
  • Check every caveat survived editing.
  • Check the certainty matches the evidence.
  • Cut anything unsupported rather than softening it.

Separate the verifier from the drafter where the report matters. See [Verifying AI-Assisted Report Content] through The Human-Verified Reporting Workflow.

A worked narrative draft

A concrete example shows where the human line sits. For a variance of “revenue down 3% against plan”:

  • AI draft (description): “Revenue finished 3% below plan, with the shortfall concentrated in the enterprise segment.”
  • AI draft (unsupported cause): “The shortfall was driven by increased competition.” – the data does not say this.
  • Human correction: replace the cause with the known reason – “The shortfall reflects two enterprise deals slipping to the next quarter,” confirmed with sales.

The description is safe to automate; the cause must be confirmed. A brief that forbids invention turns the second line into a flag rather than a fabricated explanation.

Keeping the voice consistent

Consistent voice across sections is where AI genuinely helps.

  • Give it the style guide so terminology and register stay constant.
  • Draft all sections in one pass so the voice does not shift.
  • Check terminology against the metric dictionary.
  • Edit the human-added context to match the generated prose, not the other way around.

Consistency is a real quality gain, not just a speed one. A report that reads as one document is more credible than one that reads as several.

Common narrative patterns to watch

Some AI narrative habits recur, and knowing them speeds up verification.

  • Cause inflation. Every movement gets a reason, even small ones that are simply noise.
  • Confident generalities. “Driven by strong demand” when the data does not distinguish demand.
  • Caveat erosion. A qualification present in the numbers disappears from the prose.
  • Segment drift. The commentary describes a different segment from the one that moved.
  • Rounding optimism. A decline described in a way that reads as growth.

A verifier who knows these patterns checks them specifically, rather than reading the commentary for general sense.

Common mistakes

  • Drafting with no brief. AI invents the argument as well as the prose.
  • Letting AI explain causes. Causes are facts; they must be confirmed, not generated.
  • Trusting clean prose. Fluent commentary is not true commentary.
  • Skipping the caveat check. Lost qualifications are a quiet accuracy failure.
  • Never updating the brief. The same draft every cycle, regardless of what changed.

Frequently asked questions

Can AI draft the commentary in a report?

Yes – it can describe movements from supplied data and produce a first-pass narrative, provided a human sets the argument and verifies every figure and cause.

How do you stop AI inventing reasons for a change?

Forbid it in the brief, and require a human to confirm every cause before publication. The reason a number moved is a fact to be verified, not a sentence to be generated.

What should a narrative drafting brief include?

The structured data, the human-set argument, the constraints (no invented figures or causes, preserve caveats) and the report’s voice.

How do you verify AI-drafted commentary?

Check every figure against source, confirm every cause with someone who knows, check that caveats survived, and ensure the certainty matches the evidence. Cut anything unsupported.

Does AI make commentary better or just faster?

Faster and more consistent, if verified. It does not make the argument better – that remains a human judgment – but it removes the drafting time around it.

How much of the narrative can AI draft?

Most of the description and a first-pass structure. The causes, the materiality judgments and any recommendation remain human, because they are facts and judgments the data does not contain.

Can AI draft commentary without the argument being set?

It can, but the result is generic or, worse, invents an argument. Always brief the model with the human-set argument before drafting.

What are the tell-tale signs of unreliable AI commentary?

Cause inflation, confident generalities, eroded caveats, segment drift and optimistic rounding. A verifier who knows these patterns checks them directly rather than assuming the prose is sound.

How much editing does an AI narrative draft need?

More than most sections, because the narrative is claim-dense. Plan a real verification pass over every figure and cause, not a light edit.

Can AI produce short and long versions of the same commentary?

Yes – it is good at adjusting depth for different audiences. Keep the facts identical across versions and vary only the level of detail.

What should the human rewrite focus on?

The causes, the caveats and the recommendation – the parts the model cannot know. Leave the mechanical description AI produced; spend the human time on the meaning.

Should AI draft the executive summary first or last?

Last, as a human does. The summary reflects the settled results, so drafting it after the data and the argument are fixed produces a safer, sharper draft.

Can AI draft commentary for a report it has not seen before?

It needs the data and a brief for each report; it does not carry context between them. Treat each report as a fresh brief, and never assume the model remembers a previous cycle or the reasons behind last period’s numbers. Carry forward the template and the definitions, never the model’s memory. A fresh brief each cycle is not inefficiency; it is the control that keeps the commentary tied to this period’s data and this period’s facts. Fresh context each cycle is a feature, not a failure of the tool or of the process.

Next step

Use AI for the drafting, not for the truth. Give it a brief, then verify every figure and cause before the narrative is approved. See The Human-Verified Reporting Workflow for the operating model, and book a reporting pilot to run a cycle with us.

Sources

  • Stanford HAI, 2026 AI Index and Artificial Analysis AA-Omniscience benchmark: hallucination rates of 22-94% across 26 models.
  • Loopio, 2026 RFP Response Trends & Benchmarks Report (1,500+ teams): generative-AI drafting and answer generation.

Numbers are cited from their sources and dated. Where a source is a vendor benchmark, the sample size is stated.