How do I see ROI and cost efficiency per AI feature?
Per-feature ROI analysis. Feature Intelligence answers the question every engineering leader and CFO needs to answer: of all the AI features we're running, which ones are worth the cost — and which ones aren't?
Feature Intelligence answers the question every engineering leader and CFO needs to answer: which AI features are worth the cost? It shows cost trend, waste percentage, investment classification, and cost-per-outcome for every feature in your product.
What Feature Intelligence shows
Feature Intelligence is a per-feature breakdown that goes beyond raw cost. For each feature you have tagged with X-Cost-Feature, it shows:
| Metric | What it tells you |
|---|---|
| Monthly cost | Absolute spend — how much this feature costs to run |
| % of total AI spend | Relative weight — is this feature punching above its business value? |
| Cost per call | Efficiency signal — rising cost-per-call means prompts are getting longer |
| Cost per outcome | ROI signal — cost per ticket resolved, contract drafted, etc. |
| Waste % | Optimization opportunity — how much of this feature's cost is recoverable |
| Classification | Investment or Waste — your team's judgement on the feature's value |
| Trend | 30-day cost trajectory — growing, stable, or declining |
| AI efficiency score | Composite 0–100 score: lower waste + better model fit = higher score |
Investment vs. Waste classification
Every feature in Cognocient is classified as either Investment or Waste. Cognocient starts with an automatic classification based on call patterns and task complexity. You can override any feature with one click.
Classified as Investment
High task complexity, appropriate model, measurable business outcome. Spend here is justified — and the board should see it framed as investment, not cost.
- contract-drafting — complex reasoning, gpt-4o, measurable outcome
- code-review — high value, Sonnet, engineers use it daily
- customer-support-lvl3 — resolves complex tickets, proven ROI
Classified as Waste
Low task complexity on a premium model, high volume with no clear outcome, or low usage rate despite high cost. Recoverable with routing or removal.
- sentiment-analysis — gpt-4o on a task gpt-4o-mini handles perfectly
- welcome-email — 800 calls/day, template content, no LLM needed
- draft-preview — 90% of sessions abandoned, no outcome
If Cognocient classifies a feature incorrectly — for example, a feature your team considers strategic but that uses a simple model — click the classification badge and set it manually. Manual overrides are preserved across Cognocient's re-analysis cycles.
How to use Feature Intelligence for decisions
Planning an optimisation sprint
- Sort by "Waste %" descending
- Identify the top 3 features with >20% waste
- For each, click through to see Cognocient's specific recommendations
- Apply the routing recommendations for instant savings
- Schedule the deeper fixes (context pruning, model swap) for the sprint
Typical result: 20–35% cost reduction with one sprint of work
Board presentation on AI ROI
- Filter to Investment-classified features
- Export the feature list with cost and outcome data
- Use the "cost per outcome" column to frame spend as investment
- Include the efficiency score trend to show continuous improvement
Data-driven answer to "Is our AI spend justified?"
Deciding whether to shut down a feature
- Review the feature's cost per outcome over the past 90 days
- Compare against the cost of an engineer building a non-AI equivalent
- Check the usage rate — features with fewer than 10 calls/day rarely justify their fixed overhead
- Mark as Waste and set an archive date if the data supports it
Evidence-based feature retirement decisions
The AI Efficiency Score
Each feature has an AI Efficiency Score from 0 to 100. It is a composite metric calculated from:
| Factor | Description |
|---|---|
| Model fit (40%) | Is the model appropriately sized for the task complexity? Using gpt-4o-mini for simple tasks scores higher than using gpt-4o. |
| Waste rate (30%) | What % of spend is identified as recoverable waste? Lower waste = higher score. |
| Cache efficiency (20%) | How much of the spend could be eliminated by prompt caching or semantic caching? |
| Error rate (10%) | What % of calls fail? Higher error rates mean wasted tokens on failed attempts. |
A score above 80 means the feature is well-optimised. Below 60 means significant improvement is available. The organisation-wide efficiency score on the Executive View is the average across all Investment-classified features.
Next steps: Recommendations · AI Advisor · Outcomes
Related articles
Waste Detection
Automatically identify retry waste, model mismatches, and context bloat.
Context Tax Analyser
Identify features paying static prompt overhead on every call — prime caching candidates.
Token Maxing Detector
Detect frontier models used for short outputs — tasks a cheaper model could handle.