Given a choice between shipping automatic cost-cutting or shipping visibility into where cost actually comes from, we chose visibility first, on purpose.

Sales kept asking for one feature above all others in the cost category: an automatic switch that would cut ingest costs by X% with no customer involvement. We built the cost attribution model instead, and made the automatic cutting an opt-in layered on top rather than the headline feature. Here's the reasoning, since it went against what was actually being asked for.

Automatic cost-cutting without attribution is a blunt instrument

Any automatic system that reduces ingest without understanding what's actually valuable will, on average, cut evenly — which means it degrades the 4% of your traces that are errors just as readily as the 96% that are noise, unless it has a real model of what matters. Building that model well is most of the actual engineering work; the "cutting" part is comparatively simple once you have it.

What the cost model actually computes

from ingest_usage
| where team == "payments"
| summarize
    total_cost=sum(cost_usd),
    value_weighted_cost=sumif(cost_usd, has_error or route in top_slo_routes)
  by signal
| project signal, total_cost, waste_ratio = 1.0 - (value_weighted_cost / total_cost)

That waste_ratio is the actual number worth acting on — it tells a team what fraction of their spend is going to data that's neither an error nor on a route they've told us matters for their SLOs.

Where automatic cutting fits now

With the cost model live, automatic sampling adjustments (the hard-cap option in cost controls) now target the specific low-value traffic the model identifies, rather than sampling everything down uniformly. It shipped smaller and later than sales wanted, and it's a better feature for having the model underneath it.

What sales actually got instead

The waste-ratio number turned out to sell itself better than an automatic switch would have — a prospect evaluating Helix against an incumbent can see, from their own trial data within the first week, exactly what fraction of their current spend elsewhere is going to data nobody's using. That's a harder number for a competitor to argue with than a vague promise of "we'll cut your bill," and it's become one of the more effective parts of a technical evaluation conversation.

  • Cost attribution with a per-signal waste ratio, not just a total spend number
  • Automatic cost-cutting targets identified low-value traffic specifically
  • Sequencing chosen deliberately over a faster, blunter automatic switch
  • Model reusable across cost controls, sampling, and future retention features