Lab · AI cost case

Same results, a fraction of the AI bill

An example of how I review and redesign a company's AI tooling. Brightline Home is a fictional 60-person online store that wired AI agents between every tool it uses. Switch between before and after to see where the tokens went.

Fictional company · modelled numbers you can change below

Select any box to see what it does.

Sync agent · Data sync

Every run, the largest model reads hundreds of recent records from four tools to find what changed, maps the fields and writes the updates. Copying data this way is slow, costly and now and then wrong.

Large model · 140M tokens · $842 a month

324M

tokens a month

$2,333

AI cost a month

26 min

per weekly report

16 min

longest sync wait

01The company

Brightline Home sells home goods online. Like most companies its size, it runs on a handful of tools, and leadership wants one weekly report: revenue, campaign return, sales pipeline and support trends.

CRM
CMS
Marketing suite
Store & billing
Help desk

A year ago they “added AI” by putting a large-model agent between every pair of tools. It worked, until the bill, the delays and the weekly revenue figure that never quite matched started to hurt.

02Where the money went

Most of the bill came from work that never needed a model: copying data and re-summarizing records that hadn't changed.

Contact summaries$1,342 · 58%
Data sync$842 · 36%
Product copy$105 · 5%
Weekly report$44 · 2%

03The redesign

Move data with code, not models. Each tool sends a webhook when something changes and plain mapping code copies it across: zero tokens, seconds instead of minutes, the same result every time.
One source of truth. Everything lands in one warehouse with a defined structure. Metrics are calculated once in SQL, so revenue is the same number in every report.
Right-size the AI. A small model for triage and summaries, a large one only for the report's story. Cached instructions, overnight batches for anything not urgent, and only changed records are sent.
Add guardrails. A monthly budget per workflow, tokens and cost logged for every run, a small test set so a model swap can't quietly make things worse, and a person approving anything customers see.

04What changed, box by box

Each workflow before and after the redesign, with tokens per month
Workflow	Before	After	Tokens / month
Data sync	Large-model agent scans four tools every 15 minutes	Webhooks + plain mapping code; no AI	140M to 0
Lead & ticket triage	Done inside the sync agent	Small model, fixed output format, cached instructions	— to 7.1M
Contact summaries	Every contact re-summarized weekly	Only changed records, small model, overnight batch	165M to 17.3M
Product copy	Product pages rewritten on every stock change	Drafts for new products only, approved by a person	11.4M to 52k
Weekly report	Raw exports in, model does the maths	SQL does the maths; AI writes the story, checked	8.6M to 16.9k

05A sample weekly report

The numbers come from SQL. The AI only writes the story around them, and a check rejects any number it makes up.

SQL · 3 s Revenue $412,380 (+6.2% on last week) · Orders 5,318 · Returns 3.1%

AI · 900 tokens“Revenue grew 6.2% to $412,380, mostly from the spring email campaign. Returns held at 3.1%; the top support topic was delivery times…”

✓ Every number in the story matches the data

06Try your numbers

Change the company's size and habits. Everything is calculated in your browser; nothing is sent anywhere.

Customer records40,000

Records that change each day2%

Old sync agent runs per day96 (every 15 min)

Reports per week1

AI bill per month

Before: $2,333
After: $18
Saved a year: $27,782

92% fewer tokens (324M → 24.4M). The diagram and tables above use these numbers too.

Assumptions behind the numbers

Illustrative prices, not any vendor's real price list. The point is the ratio, which holds across providers.

Large model: $5 in / $25 out per million tokens
Small model: $1 in / $5 out per million tokens
Cached instructions, batch jobs: 10% and 50% of the normal price
Old agents' instructions + tools: 6,000 tokens per run
Old sync agent: scans 500 records per run, 1,200 tokens per changed record
Old enrichment: every record weekly, 800 tokens in / 150 out
Old content agent: 100 product rewrites a day
Old report: 5 steps carrying the exports forward (12 tokens per record), ×1.3 for retries
New triage: 15% of changes are new leads or tickets
New report story: 3,000 tokens in / 900 out

Is your AI stack costing more than it should?

In a free 30-minute call we can look at your tools and AI workflows and find where plain code, a smaller model or caching would do the job.

Book a Free Consultation Or scope a project first →