If you want to know whether your monthly AI spend is normal, the baseline numbers present a puzzle. Across businesses tracked by Ramp, median monthly token spend is US$2,246, while the average reaches US$140,842. Yet according to Flexera, only 31% of organisations have visibility into their AI software spend.
A vendor invoice settles an account balance. It is a billing artefact, not a consumption record. Evaluating spend requires cost attribution: connecting token usage to workloads. If contracts are in USD and you report in AUD, understanding the measurement gap behind the board line starts with the curve. The invoice tells you what you owe, but never what you bought.
What does a normal AI token bill look like for a 50 to 500 person company, and is yours actually out of line?
A normal monthly AI token bill for a 50 to 500 person company spans from a few hundred to tens of thousands of US dollars. In Ramp’s April 2026 benchmark, median monthly spend is US$2,246 and the average is US$140,842. Normalising to per-employee-per-month (PEPM) spend lets you position spend against the median and upper percentiles. PEPM tracks expenditure rather than value: a bill near the median usually signals under-instrumentation, not discipline, which is the pattern the wider case for treating AI spend as a board line sets out in full.
Ramp recorded a median of US$46 PEPM, scaling from US$28 (4 to 10 models) to US$130 (11 to 25 models) and US$442 (26 or more models), with the middle 50% between US$3 and US$352. At the tail, Ramp’s August 2026 index showed the top 1% spending about US$8,000 PEPM. Workloads scale quickly: Flexera’s AI spending trends noted Uber exhausted its 2026 coding budget in four months before capping Claude Code at US$1,500 per engineer monthly, compared to Anthropic’s guidance of US$150 to US$250 reported by SemiAnalysis. Salesforce committed roughly US$300 million to Anthropic tokens.
Because Ramp’s US$2,246 median reflects smaller accounts, derived figures help size mid-market teams. At US$46 median PEPM, a 200-person company produces derived spend of roughly US$9,200 monthly, with the middle 50% spanning US$600 to US$70,400. That band is ordinary; upper percentiles signal scaled agentic loops. Benchmarking against the average makes normal bills look excessive, even when rising totals reflect expanded capability, as detailed in a separate guide on why AI bills climb while unit token costs fall.
Trustworthy benchmarks disclose sample sizes, spend definitions, and dates. Avoid list-price baselines, because effective rates drop once caching applies. Foreign exchange swings on USD contracts also distort local reporting. Because the range is vast and the middle undocumented, you still cannot tell where your team sits on the curve.
Why is median AI spend ($2,246 a month) so far below the average ($140,842 a month)?
Median AI spend sits far below the average because token costs follow a power-law distribution. A small group of companies running intensive agentic workloads and enterprise commitments pulls the arithmetic mean upward, creating a roughly 63-fold gap between the two figures.
As CloudZero explains, AI consumption behaves like a power law where the top few percent out-consume the rest combined. The tail is driven by agentic workloads. According to research cited by Flexera, multi-step agentic systems consume 5 to 30 times more tokens than interactive chat by planning, calling tools, and retrying failed steps while billing reasoning tokens at output rates.
Enterprise commitments stretch the tail further. In Ramp’s benchmark data, the 99th percentile reaches US$831,338 a month, and Ramp compares the spread to household income: a dense cluster near the floor, a thin tail, and no meaningful central average. Backing this up, Apollo Global Management observed that roughly 80% of spend on leading model providers comes from just 1% of customer accounts.
Valid comparisons require segmenting workloads: chatbot tools with predictable seat pricing, copilot tools with moderate developer variance, and agentic loops displaying heavy-tailed consumption. Comparing conversational tools against an average shaped by autonomous agents is a category error. Because tokens represent variable costs per code path rather than fixed seats, the shift from seat licences to consumption makes percentile benchmarks the sensible reference points. Yet both figures remain aggregate totals that you cannot decompose from the invoice alone.
Why does the invoice misstate what you consumed, and why is all your spend on one API key?
The vendor invoice misstates consumption because it reconciles accounts rather than tracking operational usage. Invoices aggregate charges across accounts or single API keys, masking retries, cached reads, and repeated context. Running all spend through one key is an operational default, an early prototype shortcut that persisted as systems scaled.
As the FinOps Foundation notes, provider bills show aggregate tokens by account or key, with no concept of business units, applications, or workloads. The invoice is a billing artefact designed to prove a balance owed.
Behind that single total, technical mechanics distort the bill:
- Retries: Failed calls retry automatically, billing as fresh usage and producing two charges for one result, as documented in Finout’s tokenomics analysis.
- Cached reads: Prompt caching discounts input by up to 90%, so a smaller bill can conceal higher traffic, while a lower hit rate doubles costs on identical volume.
- Reasoning tokens: Thinking tokens bill at output rates and are often omitted from client responses, producing short visible answers with large billed counts.
- Context re-sends: Stateless models require the client to re-send full conversation history and schemas on every turn.
These factors cause your effective rate, total spend divided by tokens processed, to diverge from published rate cards. Unsanctioned activity compounds the problem. As reported by Openlayer, UpGuard found that 81% of employees use unapproved AI tools. Shadow token consumption lands on shared accounts where nobody tracks ownership.
Before the next planning cycle closes, ask engineering about key ownership, retry rates, and cache performance. Most teams cannot answer those questions. That inability highlights an instrumentation gap: the invoice cannot deliver operational visibility on its own. When teams realise this, they naturally turn to cloud FinOps for answers, and then to the ownership question the gap leaves open — why an unmeasured line still needs a named owner.
How is AI FinOps different from cloud FinOps, and where exactly does the chargeback model break?
AI FinOps adopts tagging, showback, anomaly detection, and commitment management from cloud FinOps, but the underlying cost drivers differ. Cloud spend comes from deliberate provisioning with predictable instance pricing, whereas token spend comes from dynamic code executions. Chargeback breaks because token platforms lack a default allocation key to tie requests to business units.
Cloud environments rely on subscriptions and tags to allocate spend. With tokens, an API key aggregates everything. Making chargeback work requires three conditions: workspaces or keys mapped to cost centres as outlined by Tropic, metadata normalised across providers, and retained request telemetry. As Forrester notes regarding the FinOps Foundation’s FOCUS specification, standardising billing schemas across cloud hosts and AI providers prevents bespoke mapping scripts from collapsing.
Showback, reporting spend without debiting internal budgets, is the practical first step. Spreadsheets work for one provider and a single team, but collapse across multiple providers and projects. Commercial platforms like Datadog Cloud Cost Management for AI automate cross-provider mapping, though they require adopting vendor schemas.
Attribution tooling must reconcile directly to your invoice while separating cached reads and retries. Telemetry from an AI gateway provides the request logging needed to transform undifferentiated bills into verifiable cost centres.
Ultimately, benchmarking places your organisation on an industry curve, while cost attribution reveals what put you there. The vendor invoice delivers financial reconciliation, nothing more. As the 63-fold gap between median and average spend shows, external averages cannot tell you whether your bill is reasonable. Moving from an aggregated credential to workload attribution transforms an opaque total into an accountable engineering investment — and that move is the first step in the full picture across the cluster, where the line acquires an owner, a variance story and a control layer.
Frequently Asked Questions
Which independent AI spend benchmarks should you actually trust?
Trust the ones that disclose their sample, define what counts as spend, and date the figure. Ramp’s AI Index is the most citable because it states all three: April 2026, US businesses, monthly USD. Treat any benchmark that omits sample size, currency, or whether seats and credits are included as marketing. Nearly every figure in circulation is US-shaped, so your AUD number is derived, not sourced.
Should PEPM include seat licences and prepaid credits?
Include every recurring AI cost that scales with the work: token charges, seat licences for AI-native tools, and prepaid credits as they are drawn down. Exclude credits still sitting unused, because they are prepaid rather than consumed. Then add the AI features bundled inside SaaS you already pay for, or PEPM understates your real exposure.
What is a normal AI spend per developer, rather than per employee?
Anthropic publishes US$150 to US$250 per developer per month for Claude Code, and Uber’s 2026 coding budget ran at US$500 to US$2,000 per engineer before being capped at US$1,500. Developer-level figures sit far above company-wide PEPM because only a fraction of headcount touches the expensive workloads. Use them to size coding tools, not to benchmark the whole company.
Why did my AI bill double when my usage didn’t change?
Check the effective rate before you check usage. Spend divided by tokens actually processed moves with retries, reasoning tokens billed at output rates, cache hit rates and model mix, so a rising invoice can appear with no price change and no traffic change. Pull request-level logs and compare retry volume and model distribution across two representative weeks.
Do prompt caching discounts actually show up on the invoice?
Sometimes, and not in the way the dashboard implies. Prompt caching bills cached reads at roughly 0.1 times the input rate, and some schemes exclude them entirely, so a line that looks like less work may represent more traffic. Compare spend against tokens processed, not against line-item totals, or a cache win can read as a volume drop.
Is it true that a low AI bill means your team is being efficient?
No. A median bill usually signals under-instrumentation rather than discipline. Spend that runs through agents, notebooks and scripts outside IT bills to keys nobody governs, and roughly 81 per cent of organisations report limited oversight of that shadow consumption. A low PEPM should prompt a question about what is missing from the number, not applause.
Do prepaid credits make my monthly bill look smaller than it really is?
Often, yes. Committed and prepaid credits are invoiced once and then drawn down, so the monthly statement shows a flat figure while consumption moves underneath it. Report the burn rate alongside the invoice and amortise the commitment across the months it covers, otherwise your PEPM compares a smoothed number against other companies’ actual usage.
What should I ask engineering before the planning cycle closes?
Ask which API keys exist and who owns each one, which workloads are agentic rather than interactive, what the retry rate is, whether caching is enabled and what the hit rate is, and which projects have no named owner. Most teams cannot answer three of the five. That inability is the finding, and it is an instrumentation problem rather than a spending one.
What should I look for in an AI cost visibility tool?
Test whether it reconciles to the invoice to the cent, retains request-level records long enough to investigate a spike, lets you configure the tag model rather than imposing one, separates cached from uncached reads, and distinguishes a retry from genuine consumption. A tool that reports a total you cannot trace back to a request is a prettier invoice, not attribution.
Is a spreadsheet still enough to track AI spend under 500 employees?
A spreadsheet is adequate while you run one or two providers and a single team, because the allocation key is small enough to maintain by hand. It stops working when the same key must be normalised across three providers and a dozen projects, which is exactly when the number starts reaching a board. At that point you are paying for mapping, not software.
How do I report USD AI spend to an Australian board?
Report in both currencies and state the conversion date on the page. FX movement creates line-item variance with no relationship to consumption, and a board that cannot separate the two will read a currency swing as a usage event. Budget in USD, because the contracts are denominated there, and present the AUD figure as a translation with its date attached.
How often should I re-run the benchmark?
Re-run it quarterly, or whenever the composition of your stack changes: a new provider, a new agentic workload, or a shift from seats to consumption. Token pricing moves as well, so a benchmark more than two quarters old describes a market that no longer exists. Track your own percentile against your own history in between.
