Stop 13 · Business Communication & StorytellingTopics · 2/7

Communicating Model Results to Stakeholders

In short

Your model has an AUC of 0.87. Your business head nods politely and understands nothing. Results only count when people can act on them, so talk in rupees, risks and decisions.

Presenting model performance and impact in business terms (revenue impact, error cost) rather than only ML jargon (F1 score, AUC) that a non-technical audience can't act on.

The mascot in a doctor's coat explains a report to a smiling family across a clinic desk.

Think of it like this

The doctor who explains

A good doctor doesn't read out 'haemoglobin 9.2' and leave. She says: you're low on iron, eat more greens, take these tablets, come back in a month. The numbers stay in the report. The conversation is about what to do next.

Watch it happen

Turning a metric into a decision

You

Ops head

Finance

Recall is 0.80

1/6You open with the metric. The ops head looks politely lost.

The big ideas

What you need to know

Lead with the decision

Start with what you recommend and why. Metrics come after, as evidence.

Counts beat ratios

'Catches 8 of every 10 frauds' lands better than 'recall 0.8'. Same fact, friendlier shape.

Attach a rupee value

Estimate money saved or earned, and state your assumptions so finance can check them.

Show the trade-off

Every model makes mistakes. Say who gets hurt by each kind of error and how often.

See it in code

Walk through the code

python

1tp, fp, fn = 800, 200, 200          # from the test-set confusion matrix2avg_fraud = 5000                     # rupees lost per missed fraud (assumed)3review_cost = 50                     # rupees to review one alert (assumed)4 5saved = tp * avg_fraud6cost = (tp + fp) * review_cost7missed = fn * avg_fraud8 9print(f"Caught {tp / (tp + fn):.0%} of frauds")10print(f"Net saving: Rs {saved - cost:,}  |  Still missed: Rs {missed:,}")

1/4True positives, false positives and false negatives from your test set.

A few lines turn a confusion matrix into sentences a manager can repeat in her own meeting.

Save yourself the pain

Common mistakes

Oops

Opening the meeting with F1 scores, ROC curves and model names.

Do this instead

Open with the recommendation and its impact. Keep the metrics on a backup slide for anyone who asks.

Oops

Hiding the false alarms so the model looks better.

Do this instead

Show both error types and who they affect. Stakeholders trust you more when you name the downsides first.

Your mission

Rewrite a model summary

~20 min

Quick check

Which sentence works best for a sales head deciding whether to use a churn model?

Go deeper

1Rewrite a model performance summary from ML metrics into business-impact language.
2Practice the 'answer first' structure: recommendation, then supporting evidence.
3Prepare for the common question: 'how confident are you, and what happens if it's wrong?'