Achilles — AI Foundation: Cost accounting

The platform meters every model call — who spent what, on what, and how much — to show spend and hold the budget. Accounting is end-to-end: one monetary aggregate by model, plus token attribution to people from the journals. Money is counted by model and by company; by person — for now, in tokens.

✦ AI calls indexing · search
chat · agents

──▶ every call is metered

▦ Accounting — two paths model_usage · aggregate
journals · attribution

──▶ the screen reads it

▤ AI usage total · by person
by model · limits

model_usage — the platform’s full spend, in money journals — token attribution to a person

What we count and from where → AI usage

The screen shows four views; each has its own source, unit, and scope. One accounting circuit, four projections.

View	Source	Unit	Scope
Company total	`model_usage` · `SUM` over the period	tokens + $	all functions
By employee	journals: `messages` · `agent_runs`	tokens	chat + agents
By model	`model_usage` · by model · function	$ (input / output)	all functions
Limits	`platform_settings`	tokens · $	caps + alert

Why two sources

Spend is written to two places not out of duplication, but because they answer different questions.

model_usage aggregate

accrues requests · input · output · cost by model · function · day
catches impersonal spend too — indexing and RAG search
the basis for “company total” and “by model”, in money

journals attribution

chat — messages.tokens_used; agents — agent_runs.tokens_used
per-person spend — SUM from the journals, no counter
the basis for “by employee”, in tokens

They overlap on chat and agents, but the aggregate is broader: it catches indexing and search, which can’t be attributed to a person. So “company total” ≠ the sum over people. Money per person requires splitting the input/output price across each message and run — that’s v2.

Periods and budget

The week, month, and year panorama is one and the same table under a different window.

Total for the period — a derived SUM over model_usage for the calendar window (week · month · year), not a separate counter table. The same technique as the agent budget — “a derived budget”.
Budget alert — a monthly $ threshold for the whole organization (ai_monthly_budget): the monthly “total” tile is checked against it, and on reaching it — a Budget notification; requests aren’t blocked.
Weekly token caps per employee (agents · chat) are checked against journal spend, the window is calendar-based, reset Sunday 00:00 in the organization’s timezone.

v1 / v2 boundary

v1 Tokens everywhere; money by model and by company; the monthly budget alert; weekly token caps (agents — blocking, chat — informational).

v2 Budgets by org structure — caps and alerts not as two poles (whole company + person), but a full-fledged system of allocating tokens to teams and departments, by hierarchy. Money per person; trends, charts, and export; blocking chat over the cap; a daily user_usage aggregate as load grows.