Every AI cost tool estimates from a price sheet. Provider invoices don’t work that way — cache multipliers, reasoning tokens, mid-month price changes, 30-day restatements. Traken reconciles every request to your provider invoice, to the cent. Then it proves what your AI actually costs, wastes, and earns.
Verified against real provider invoices — where reconciliation succeeds, the residual is $0.00; where it can’t, Traken tells you exactly why instead of hiding it.
Read-only. Token counts and hashes only — never your prompts.
Traken’s core is deterministic SQL — zero LLM anywhere in the numbers path. Every figure traces to an invoice line or is explicitly labeled MODELED with its named assumption. Never one blended ROI number.
Per-request true cost matched to provider invoices — and every gap explained, never averaged away.
Upload your AWS, GCP, or any FOCUS billing export — Traken reconciles it to the closed invoice the same way it reconciles your LLM spend, allocates shared costs with sub-0.1% error, and shows anything unallocated explicitly instead of smearing it. The ~3% silent misallocation that makes finance distrust cost tools is exactly what we refuse to do.
The gross-margin-per-feature answer investors ask for.
Invoice → general ledger → margin bridge; measured savings separated from modeled forecasts, exceptions shown, never dropped. NPV, IRR, payback — with optimism explicitly haircut, never baked in.
Cost attribution and showback per tenant, feature, and customer. Your system of record for AI spend.
One dollar-ranked list across your AI and cloud bill: model right-sizing, caching, context bloat, retries, batch — alongside unattached volumes, idle gateways, orphaned snapshots, off-hours waste. Anti-double-counted, each shipped as a ready-to-apply config diff plus an eval plan — hand it to your engineer or your coding agent, run the evals, merge. Traken never touches your production path; you approve every change. And after you ship a fix, Traken reprices your actual workload against the next invoice — so every claimed saving is measured, not asserted.
Most AI initiatives can’t demonstrate ROI — not because there is none, but because nobody defined benefits, baselines, and owners before shipping. Traken’s guided interview does exactly that: benefit classes and conversion paths, costs beyond inference, hurdle rates, ramp phasing, a skeptic pass that challenges optimistic assumptions against reference-class priors. Then actuals flow in monthly from reconciled cost, and MEASURED vs MODELED variance is tracked — so six months in, you know whether the case held, not just whether it was pretty.
Two artifacts, one engine, exportable in a day:
Verified Cost Report. What you actually spend on AI — every dollar tiered by how it’s verified: reconciled to the invoice, invoice-verified, management-declared, or uncovered. Coverage sums to 100% — we tell you what we couldn’t verify, too. Includes deterministic stress-replay: your real workload repriced under +25% provider prices, caching off, 5–10× load — same SQL, same hash, run twice.
Value Case. What it returns — NPV, IRR, payback with explicit bias adjustment, benefit targets vs measured actuals month by month, and a closed loop on every optimization: each shipped fix counterfactually repriced against the next invoice, so “we saved $X” is a reconciled number, not a slide.
A boutique cost review runs €5,000–15,000 and takes 2–4 weeks — opinions in a PDF you can’t re-run. AI-specific or enterprise FinOps work runs well past €30,000, often 4–8 weeks, with the same limitation. Traken’s report is generated from your reconciled data, reproducible bit-for-bit, from €490, in a day.
Cache writes billed at 1.25×–2×. Reasoning tokens billed as output. Prices changed mid-month. Batch and cache discounts stacking. Usage restated 30 days later. Price-sheet math silently drifts from the bill — and Traken’s effective-dated rate engine catches the drift before you see it on the invoice.
Bring the telemetry you already have — Langfuse, OpenTelemetry GenAI, or our thin SDK wrapper. Traken ingests token counters and metadata (never message content), reconciles against your usage API and invoice, and puts every number behind a SQL query you can inspect. Narrative reports are generated by an EU-resident model by default, with every number cited from SQL and hard-blocked if grounding fails.
The FinOps playbook — attribution, tag coverage, anomaly detection, rate-drift monitoring, savings ledgers, business-case governance — built into the product instead of a hire you can’t make yet. FOCUS-aligned data model, FinOps-standard KPIs, and the discipline enforced by the engine, not by a slide deck.
One engine, one methodology — from your Anthropic invoice to your AWS bill.
Dashboards estimate. Traken reconciles. If a number doesn’t tie to an invoice line, it’s labeled modeled — or it doesn’t ship.
You’d get estimates. Reconciliation requires an effective-dated rate engine, a residual-allocation waterfall, and restatement handling — and a build-it-yourself spreadsheet won’t survive diligence.
Price, speed, and reproducibility. A boutique review runs €5–15k and takes 2–4 weeks; AI-specific or enterprise work runs well past €30k and 4–8 weeks — and both deliver opinions you can’t re-run. Traken’s report is generated from your reconciled data by a deterministic engine — every number traceable to SQL and invoice lines, reproducible bit-for-bit, from €490, in a day. When a number is challenged, you re-run it instead of scheduling a call.
Yes — upload your billing export (FOCUS, CUR 2.0, BigQuery export); no cloud credentials needed. We reconcile finalized bills to the invoice, allocate to teams and features, and detect billing-visible waste. We deliberately don’t do real-time dashboards, commitment trading, or workload rightsizing — we reconcile truth, not guesses.
No. Token counts, hashes, and business metadata only. Our collector config is open for inspection.
Never. Read-only, off-path. Remediations ship as config diffs with eval plans; you apply them.
No one honestly can. We show dollar-ranked opportunities grounded in reconciled cost — and after you ship a fix, we counterfactually reprice your workload against the next invoice, so the savings ledger tracks what actually landed in your P&L.
EU infrastructure, EU-resident narration model by default. A premium model lane is available opt-in.
Upload one invoice. Watch it reconcile.