Investment vs. waste: classifying AI spend instead of just cutting it
A budget review under pressure tends to ask one question: where can we cut? That question treats every line item the same way — as a cost to reduce — which quietly ignores the fact that some AI spend is generating real, measurable business value and some of it is an internal playground nobody has touched since a demo six months ago. Cutting both by the same percentage is not a strategy, it's an averaging error.
The better question is: for each dollar of AI spend, is this an investment or is it waste? Answering that requires evidence, not intuition — a feature's name and its cost rank tell you almost nothing about whether it's actually working.
Why cost rank is the wrong sorting key
The highest-cost feature on an account is frequently also the one generating the most business value — a contract-generation tool accelerating deal velocity, or a support bot measurably reducing per-ticket cost, both tend to be expensive precisely because they're used heavily in production. Cutting spend by starting at the top of a cost-sorted list risks cutting exactly the feature that's working best, purely because it's also the most visible line on the invoice.
Meanwhile, a genuinely wasteful feature — an internal playground environment, a deprecated summarizer nobody migrated off of, a proof-of-concept that never got a real production home — can sit quietly mid-list, under the radar of a review that only looks at the top of the sort. Waste doesn't announce itself by being expensive. It announces itself by producing nothing while still costing something.
The three-tier classification system
Cognocient classifies every feature as Investment, Waste, or Unknown using a strict priority order, so a human decision always wins over an automatic guess, and an automatic guess is always better than nothing:
Tier 1 — Manual classification
A team member explicitly classifies a feature via the dashboard or the API, with a reason attached. This always wins over any automatic logic and persists indefinitely — the honest answer when someone knows the truth beats any heuristic.
Tier 2 — Keyword heuristic
Without a manual classification, feature names matching production-oriented patterns (chat, search, extractor) lean Investment; names matching non-production patterns (test, dev, playground, debug) lean Waste. A reasonable default when nobody has weighed in yet.
Tier 3 — Token-count heuristic
As the last resort: features with average completion length above roughly 800 tokens lean Investment (a sign of substantive reasoning work), while low-output features on an untagged or conversational workload lean Waste. Anything that doesn't clearly match either pattern stays Unknown rather than being forced into a guess.
The Unknown bucket is deliberate, not a gap — a feature that doesn't clearly match either pattern genuinely needs a human to look at it rather than being silently miscategorized by a heuristic reaching past its confidence. See the full classification reference for the manual classification API and audit log structure.
What this produces on the dashboard
A three-column panel — Investment, Waste, Unknown — each ranked by spend within category, feeding a board-summary sentence generated directly from the data: “$8,200 of AI spend is driving measurable value. $4,200 (34%) is recoverable waste — eliminating it improves AI efficiency by 34 percentage points.” That sentence is legible to a CFO in ten seconds, with the full evidence one click away for anyone who wants to check it.
The same classification data feeds directly into the AI Efficiency Score and into per-feature recommendations, so a feature marked Waste gets a higher-priority recommendation than one marked Investment — the classification changes what action gets suggested, not just how the feature is labeled.
What this looks like in practice
Consider a hypothetical account with a support-chat feature at $5,100/month generating measured CSAT improvement, and an internal-playground feature at $2,800/month with no production users at all. Sorted by cost alone, both would appear as candidates for review with the playground actually cheaper and less visibly urgent. Classified by evidence, the playground surfaces clearly as Waste — zero production impact for real spend — while the support-chat feature, despite costing nearly twice as much, is confirmed as Investment and protected from an across-the-board cut that would have reduced something that was working.
A team can also integrate classification directly into its deployment pipeline — automatically marking a new feature Investment the moment it passes a production readiness review, so the classification stays current without a recurring manual pass. This is illustrative of how the classification behaves, not a specific customer's reported outcome.
Classification answers whether spend is worth it in aggregate. It works alongside, not instead of, cost-per-outcome tracking for features where a precise per-result cost figure is available — classification is the broader net, outcome tracking is the sharper instrument for features that produce a countable result.