TL;DR: Across 5,995 complaints about AI API spend, the most common cost-tracking tool is a spreadsheet — 128 independent mentions, more than every tool-based approach combined. The most requested feature is not a dashboard but recommendations (135 mentions). The gap between those two numbers is the entire product opportunity.
What people actually do
We aggregate what people describe themselves doing, not what they say they want. For AI cost tracking, the behaviour is unusually consistent:
- 128× Manual tracking in spreadsheets
- 42× Manual spreadsheet tracking of AI costs specifically
- 26× Manual prompt engineering to reduce token usage
- 21× Caching responses to reduce API calls
- 21× Building custom caching layers
- 20× Manually monitoring provider API dashboards
- 19× Switching between providers based on cost
- 15× Building custom routing logic in-house
The first two lines are the finding. More than 170 engineering teams independently decided that the correct tool for tracking AI spend is a spreadsheet they maintain by hand. These are people who build software for a living, working at companies that pay for observability tooling, choosing manual data entry anyway.
That only happens when the alternatives do not fit. And the last three lines say what "does not fit" means — 19 teams are manually choosing providers by price and 15 have written their own routing layer. That is not a monitoring gap. That is people building half a product internally because the whole one does not exist in a form they can buy.
What they ask for
- 135× Cost optimization recommendations
- 41× Token usage analytics dashboard
- 41× Cost monitoring and optimization dashboard
- 27× Token usage monitoring and alerts
- 25× Budget alerts and forecasting
- 21× Real-time token usage dashboard
- 15× Multi-model cost comparison dashboard
- 14× Automatic routing between cheap and expensive models by task
"Recommendations" outranks every visibility request. This is the part most easily missed when building here: a dashboard that faithfully reports a large number is not the product people are describing. They can already see the number — it arrives monthly from the provider. What they cannot see is which of their forty prompts is responsible for it, and what to do about it.
The second signal is alerting before, not reporting after. Budget alerts and forecasting (25) plus real-time usage (21) both describe wanting to know on the way to a large bill rather than at the end of the month.
Why the existing tools miss
The complaints about current options are specific:
- 4× Per-token pricing becomes expensive at production scale
- 2× Generic cloud cost tools are not designed for AI-specific token consumption patterns and pricing models
That second one is the structural problem. Established cloud cost tooling models spend as a function of resources running over time. Token spend is a function of request shape — how long the prompt is, which model answered, how much context was attached, whether a retry happened. A tool built for instance-hours has nowhere to put that.
Meanwhile the provider dashboards are per-provider by construction, which is exactly wrong for a team that has deliberately spread work across three of them to control cost.
Who is running into it
The personas are narrow and consistent:
- Software developers and engineering teams using AI coding assistants (26 mentions)
- Developers and startups building AI-powered applications (16)
- AI engineers building LLM-powered applications (6)
- Engineering teams and CTOs managing AI/ML infrastructure budgets (5)
Notice the split. Most complaints come from people whose tools cost money unpredictably (coding assistants), not from people running large inference workloads. The problem is arriving in ordinary engineering teams, not only at AI-first companies — which widens the addressable group considerably.
The honest caveats
- 48% willingness to pay is elevated, not extraordinary. Against a 35% baseline it is a real signal, but this is not the strongest category we measure. Payment infrastructure runs at 68%.
- This is a crowded space. Several funded companies already sell LLM observability. The gap the data describes is narrower than "AI cost tool" — it is the recommendation layer and the multi-provider view, not the dashboard.
- Complaints skew toward the annoyed. We see people who wrote a post, not the teams for whom provider dashboards were adequate. Nothing here measures how many of those exist.
- "Manual spreadsheet" is a strong signal but a weak moat. The reason 128 teams tolerate it is partly that the work is small. A product has to be meaningfully better than fifteen minutes a week.
How this was assembled
Aggregated from PainHunt's analysed discussion set: public complaints gathered across every source we crawl, each scored for severity and willingness to pay by the same model. Counts are occurrences of a theme across independent posts, with near-duplicate phrasings merged — "manual tracking in spreadsheets" and "manual spreadsheet tracking of API costs" are counted as one practice where the phrasing matched closely. No post content or author is reproduced.
Related reading
- How to tell if a pain point is worth building for — the scoring behind the 35% baseline
- Cost control for AI coding assistants — the same cost problem seen from the coding-tool side
- Reduce LLM token costs by preprocessing documents — one concrete reduction tactic
- Check a specific version of this idea in validate, or browse signals in the dashboard