AI Token Spend Diagnostic
Your AI Spend Report.
Now eliminate the waste.
A ranked analysis of token spend, workflow inefficiency, and identified savings. Designed to be read by your engineering team and presented to your CFO.
Client
Meridian Advisory Group
Analysis Period
1 Jul — 31 Aug 2026
Report Generated
9 September 2026
Identified Monthly Saving
$47,600
$571,200 annualised across 5 ranked changes
Recoverable Spend
42.5%
Of current monthly AI outlay
Efficiency Score
5.8/10
Moderate: optimise selectively
Executive Summary
Page 2 of 6
Meridian: $47,600 recoverable every month.
Based on 61 days of Meridian's usage data across 4 providers and 6 workflow categories. Every recommendation can be implemented without changing your AI stack. That is $571,200 per year from 5 structural changes.
Monthly AI Spend
$112,000
4 providers via OpenRouter · 2-month average
Identified Monthly Saving
$47,600
More than covers the cost of this diagnostic, every month
Recoverable Spend
42.5%
More than two-fifths of your monthly AI bill is recoverable
Efficiency Score: where Meridian sits today
Score 5.8: Moderate
Critical
Significant Waste
Moderate
Optimised
0.0 – 2.5: Critical, immediate action
2.5 – 5.0: Significant waste, high ROI on changes
5.0 – 7.5: Moderate, optimise selectively
7.5 – 10.0: Optimised
Top three findings
$18,700/mo
System prompt bloat in Meridian's Client Briefing Pipeline. Prompts are averaging 11,400 tokens (roughly 8,500 words) before each request. Only about 2,000 of those tokens are doing useful work; the rest is static boilerplate the AI reads and discards every time. Compressing to a structured template removes the waste; your engineers should check output quality against sample requests before deploying. This is the single highest-return action in the report: 39% of total identified savings, implementable in under one working day.
$13,800/mo
GPT-4o used for document classification where GPT-4o mini costs 94% less. GPT-4o is a frontier reasoning model, well-suited to complex tasks. Document classification (assigning incoming files to a category) does not require that level of capability. Your team is using a precision instrument for a routine sorting job. Switching 4 identified workflows to a cheaper model costs one day of engineering work. Test a sample of outputs before switching to confirm quality holds.
$8,200/mo
Agent loop redundancy in the Research Automation workflow. 31% of API calls within this workflow ask the AI an identical question to one already asked in the last 60 seconds. Each call is billed as a fresh request. The fix is a lightweight cache that checks recent answers before making a new call. If the question is the same, the cached answer is returned and no charge is incurred. Engineering effort: approximately 3 working days.
Verdict: Meridian's highest-impact action is compressing the Client Briefing Pipeline system prompt from 11,400 tokens to approximately 2,000 tokens. This single change saves $18,700 per month (39% of total identified savings) and can be implemented in under one working day. It requires no model changes, no infrastructure changes, and no retraining.
Spend trajectory
Projected annual AI spend at current patterns, at 3x and 5x usage growth, and at 1.2x and 1.5x provider pricing increases, with and without the identified fixes applied.
| Scenario |
Annual Spend (current) |
Annual Spend (fixes applied) |
Annual Saving |
| Volume growth scenarios |
| Current run rate |
$1,344,000 |
$772,800 |
$571,200 |
| 3x usage growth |
$4,032,000 |
$2,318,400 |
$1,713,600 |
| 5x usage growth |
$6,720,000 |
$3,864,000 |
$2,856,000 |
| Provider pricing stress (current volume) |
1.2x provider pricing Same workload, rates up 20% |
$1,612,800 |
$927,360 |
$685,440 |
1.5x provider pricing Same workload, rates up 50% |
$2,016,000 |
$1,159,200 |
$856,800 |
Volume growth: structural inefficiencies scale proportionally with usage. Fixing them now costs one to three days of engineering time at any volume. Provider pricing stress: eliminating the structural waste reduces the absolute bill you are defending against rate changes, because the baseline starts lower.
AI Layer Gross Margin
One additional input unlocks this calculation: your AI-attributable annual revenue. This is the value of work delivered with AI embedded in the workflow, such as client fees, contracts, or output volume. On the live diagnostic, you provide this figure and TokenomicsIQ calculates your AI Layer Gross Margin before and after the identified fixes. The illustrative example below uses $2,880,000 AI-attributable annual revenue.
| Metric |
Before fixes |
After fixes |
Change |
| AI-attributable revenue (illustrative) |
$2,880,000 |
$2,880,000 |
— |
| AI token spend (annual) |
($1,344,000) |
($772,800) |
$571,200 |
| AI Layer Gross Margin |
53.3% |
73.2% |
+19.9pp |
AI Layer Gross Margin = (AI-attributable revenue − AI token spend) / AI-attributable revenue. The $47,600/month in identified savings flows directly into margin. That is the number an investor reads.
Meridian Advisory Group · AI Token Spend Diagnostic · September 2026
TokenomicsIQ
2 of 6
Spend Landscape
Page 3 of 6
60% of Meridian's spend is in one model. Much of it on the wrong tasks.
GPT-4o accounts for $67,200 of Meridian's monthly bill: the right choice for complex reasoning, but over-deployed for classification and templated drafting.
Spend by model
Monthly average, all providers. GPT-4o dominates at 60% of total spend. The model is appropriate for some workflows but significantly over-used for simpler tasks where cheaper models perform equally well.
Spend by task type
Monthly average. Document analysis and classification combined account for 40% of spend, and both categories show high substitution potential. Code generation is the one category where current model selection is appropriate.
Cost efficiency by task type
Efficiency Score for each task type = 10 × (1 − recoverable spend ÷ monthly spend for that task type). A score below 5.0 means more than half of that task type's spend is recoverable. P&L classification reflects how each workflow maps to your income statement: COGS is AI spend tied directly to client service delivery, R&D is product and capability development, OpEx is internal operations.
| Task Type |
Model |
Avg Tokens / Request |
Monthly Spend |
Efficiency Score |
P&L Class. |
Action |
| Document Analysis |
GPT-4o |
4,200 |
$33,600 |
5.6 |
COGS |
Optimise prompts |
| Code Generation |
GPT-4o |
3,100 |
$22,400 |
7.8 |
R&D |
Keep current model |
| Client Report Drafting |
GPT-4o + Claude |
8,600 |
$19,600 |
4.5 |
COGS |
Compress system prompt |
| Data Extraction |
Claude Sonnet |
2,800 |
$16,800 |
6.9 |
OpEx |
Keep current model |
| Classification |
GPT-4o |
1,900 |
$11,200 |
2.0 |
OpEx |
Substitute model |
| Other / Uncategorised |
Various |
2,200 |
$8,400 |
6.5 |
? |
Review & categorise |
Note on Other / Uncategorised: These calls could not be reliably mapped to a named workflow: typically ad hoc queries, developer test calls, and multi-purpose scripts without a consistent prompt structure. Adding a workflow_id tag to your API call metadata enables full categorisation in subsequent diagnostics.
Meridian Advisory Group · AI Token Spend Diagnostic · September 2026
TokenomicsIQ
3 of 6
Inefficiency Analysis
Page 4 of 6
Five inefficiencies. Total monthly impact: $47,600. That is $571,200 per year.
Ranked by monthly saving potential. Difficulty rating reflects estimated engineering hours to implement the fix.
| # |
Workflow / Pattern |
Monthly Cost (current) |
Monthly Saving |
Saving % |
Difficulty |
1 |
System Prompt Bloat Client Briefing Pipeline |
$22,000 |
$18,700 |
85% |
Low |
2 |
Model-Task Mismatch Document Classification (4 workflows) |
$14,600 |
$13,800 |
94% |
Low |
3 |
Agent Loop Redundancy Research Automation |
$16,400 |
$8,200 |
50% |
Medium |
4 |
Batch Eligibility (additional observation) Daily Report Generation |
$7,200 |
$4,300 |
60% |
Low |
5 |
Prompt Cache Miss Content Review Pipeline |
$5,800 |
$2,600 |
45% |
Medium |
| Total |
$66,000 |
$47,600 |
72% |
|
Current cost vs. optimised cost: top 5 inefficiencies
The amber segment on each bar is money leaving your P&L unnecessarily each month. The indigo segment is the baseline cost of running the same workflows efficiently. Fixing just the top two items recovers $32,500 per month: more than the cost of this diagnostic, at roughly two days of engineering work between them.
Meridian Advisory Group · AI Token Spend Diagnostic · September 2026
TokenomicsIQ
4 of 6
Recommendations
Page 5 of 6
Structural findings in detail. Two flagged, one minor.
TokenomicsIQ runs five structural detectors on every engagement. Below are Meridian's two flagged structural findings and one minor timing observation, followed by a ranked action list a developer can start from tomorrow morning.
System Prompt
Bloat
Critical
System prompts average 11,400 tokens in the Client Briefing Pipeline, accounting for 38% of total token spend in that workflow. The excess is static boilerplate sent with every request: content that never changes between calls, billed each time as if it does.
In plain terms
Your team is sending roughly 8,500 words of instructions to the AI before every client briefing request. Only about 1,500 of those words vary between requests. The rest is the same standing brief read aloud at the start of every meeting. Cutting this to a lean 2,000-token template is the highest-return action in this report: $18,700/month saved, less than one day of work, with output quality checked against sample requests before deployment.
$18,700/mo
identified saving if resolved
Agent Loop
Redundancy
Flagged
31% of Research Automation API calls repeat identical or near-identical context within 60 seconds of a prior call. This is a classic agent loop signature: intermediate outputs are not being cached, so the AI is asked the same question multiple times in quick succession and charged each time.
In plain terms
Nearly one in three calls in this workflow is a duplicate the AI has already answered moments earlier. A lightweight cache (essentially a short-term memory layer) checks whether the same question was asked in the last 60 seconds before making a new API call. If it was, the cached answer is returned at no cost. Engineering effort: approximately 3 working days.
$8,200/mo
identified saving if resolved
An additional observation, outside the five structural detectors. One instance found in the Daily Report pipeline, included in Priority Action 4 as a low-effort saving opportunity.
In plain terms
Some AI tasks are time-sensitive: a chatbot response needs to return immediately. Others are not: a daily report can run overnight. When time-insensitive jobs are processed in real-time (synchronous), they pay full rate. When batched, providers typically charge about 50% less for the same work. Where jobs pay real-time prices without needing real-time delivery, the report notes them. Meridian's other core workflows are already well timed.
No critical issue
minor optimisation opportunity noted below
Model substitution recommendations
Each recommended substitution carries a quality-impact note based on the task type. These are expectations to confirm, not test results: check a sample of outputs before switching.
| Task Type |
Current Model |
Recommended Model |
Monthly Saving |
Quality Impact |
| Document Classification |
GPT-4o |
GPT-4o mini |
$13,800 |
No change |
| Internal Routing Logic |
Claude Sonnet 3.7 |
Claude Haiku 3.5 |
$2,400 |
Minimal (<3%) |
| Simple Q&A Queries |
GPT-4o |
GPT-4o-mini |
$1,100 |
No change |
Priority action list
Ranked by monthly saving. Effort in working days for a mid-level engineer familiar with Meridian's codebase.
1
Compress Meridian's Client Briefing Pipeline system prompt from 11,400 to ~2,000 tokens
Refactor static boilerplate into a shared context template. Test against 20 sample outputs before deploying.
2
Switch Document Classification workflows (4 identified) from GPT-4o to GPT-4o mini
Update model parameter in each workflow configuration. No prompt changes required. Test a sample of outputs before switching.
3
Add 60-second deduplication cache to Research Automation agent loop
Hash context window content; return cached response if identical call made within 60s. Redis or in-memory cache both work here.
$8,200/mo
Effort: 2 – 3 days
4
Convert Daily Report generation to async batch API calls
Daily report jobs are non-time-sensitive and currently running synchronous. Batch API pricing on OpenAI and Anthropic is 50% below standard rates, and running these jobs as a batch also removes the duplicate retry calls they currently make on timeout.
$4,300/mo
Effort: 1 – 2 days
5
Enable prompt caching for Content Review workflow repeated system prompts
Anthropic and OpenAI both support prompt caching for repeated prompt prefixes. Single configuration change per workflow.
$2,600/mo
Effort: 0.5 day
Meridian Advisory Group · AI Token Spend Diagnostic · September 2026
TokenomicsIQ
5 of 6
How TokenomicsIQ classifies and scores your usage.
TokenomicsIQ runs one classification pass over your request-level usage export, then five structural detectors. Here is how the export is handled, classified and scored.
01
Usage Ingestion
Your request-level usage export is normalised into one schema: OpenRouter (CSV) and TypingMind (JSON) today, with other gateways such as Portkey on request. Token counts, model IDs, timestamps, and endpoint identifiers are standardised into a unified schema. No API connection required.
02
Workflow Classification
Each request is assigned to one of eight task types. A sample is classified by an LLM, using usage metadata and, where present, the first 500 characters of the prompt; the rest are classified by pattern inference from usage metadata. Each report states how many records were classified each way, and flags any record below a confidence threshold as estimated.
03
Benchmark Comparison
Actual cost per task type is compared against a benchmark cost built from published provider pricing: a capability matrix maps each task type and complexity band to an appropriate model. No client data is used to build benchmarks.
04
Five Structural Detectors
Model-task mismatch, agent loop redundancy, system prompt bloat, prompt cache miss and vendor concentration run on the full dataset. Findings inferred from patterns are flagged for your engineers to confirm. Additional observations, such as batch eligibility, are noted where found.
05
Savings Calculation
Identified savings are calculated conservatively, using current provider pricing at the client's actual volume tier, with no assumption of future pricing changes. Model substitution savings use published prices through the capability matrix, and each substitution carries a quality-impact note.
06
Efficiency Score
Efficiency Score = 10 × (1 − recoverable spend ÷ monthly spend), rounded to one decimal place and held between 0 and 10. Meridian's identified saving of $47,600 on $112,000 of monthly spend (42.5%) gives a score of 5.8, in the Moderate band.
Data handling
Your usage export is used only to produce this report and is deleted within 24 hours of delivery; email copies are removed on receipt. TokenomicsIQ works from token counts, model IDs, timestamps and cost. Where the export includes prompt text, only the first 500 characters of each prompt are read, and only to classify the task type; AI outputs are not analysed. A sample of records, including any such excerpts, is sent to Anthropic's Claude API for that classification only, and Anthropic does not train on it. Your data is never used for benchmarks or shared for any other purpose. See our privacy policy for details.
About TokenomicsIQ
TokenomicsIQ is a fixed-fee AI token spend diagnostic for engineering teams and the CFOs they report to. It works from a usage export from your AI gateway, such as OpenRouter or TypingMind. Start with a free assessment within 24 hours; the full ranked report (HTML and PDF), with specific, implementable savings recommendations, follows within 2 working days of commissioning. No integration. No ongoing subscription. No consultant retainer. $3,500: the kind of clarity that used to cost ten times as much.
Get your report at tokenomicsiq.ai
Meridian Advisory Group · AI Token Spend Diagnostic · September 2026
TokenomicsIQ
6 of 6