You're paying too much for AI.
Not because you chose the wrong tool — because you're using the expensive one for everything.
AT&T figured this out. They route 40% of their AI work to open-source models they run themselves. Only the hard jobs go to frontier systems. The result: a 56% cost cut for roughly a 2% quality tradeoff. In some applications, smart routing cut costs 80-90%.
A security company tested this too. Aikido ran 11.7 billion tokens through security benchmarks and found that three cheap model runs beat a single expensive Opus 5 pass on coverage.
The pattern is clear: most of your AI work doesn't need the most expensive AI.
The 40% Rule for SMBs
Here's the framework. Walk through your AI tasks and sort them into three buckets:
Bucket 1: Routine Work (≈40% of tasks)
This is your "cheap model" tier. Tasks like:
- Drafting social media posts
- Summarizing meeting notes
- Formatting data into templates
- Writing follow-up emails
- Generating routine reports
These don't need frontier reasoning. A $0.10/task model handles them fine. Run them on the cheapest model that produces acceptable output.
Bucket 2: Judgment Calls (≈40% of tasks)
This is your "mid-tier" tier. Tasks like:
- Customer email responses requiring tone sensitivity
- Reviewing proposals for gaps
- Prioritizing a task list based on business context
- Writing job postings that attract the right candidates
These need more nuance. Use a mid-range model. Not the most expensive — just one that understands context.
Bucket 3: Hard Problems (≈20% of tasks)
This is your "expensive model" tier. Tasks like:
- Complex contract review
- Strategic decision analysis
- Debugging production code
- Financial modeling
Only here do you need the $200/month, frontier-tier model.
How to Set This Up Today
Step 1: Audit one week of AI usage. Write down every task you used AI for. Don't overthink it — just list them.
Step 2: Sort into buckets. Use the three buckets above. If a task is repetitive and follows a pattern, it's Bucket 1. If it needs judgment but not brilliance, it's Bucket 2. If getting it wrong would cost you real money, it's Bucket 3.
Step 3: Assign models. Bucket 1 gets your cheapest option. Bucket 2 gets a mid-tier. Bucket 3 gets the expensive one.
Step 4: Track cost per accepted result. This is the metric AT&T uses and it's the one that matters. Not cost per token — cost per output you actually use. If the cheap model produces 8 out of 10 usable results and costs 90% less, it wins even though the expensive model goes 10 for 10.
The Math That Makes the Case
Say you run 100 AI tasks per month:
| Bucket | Tasks | Model Cost | Monthly Cost |
|---|---|---|---|
| Routine | 40 | $0.10 each | $4 |
| Judgment | 40 | $0.50 each | $20 |
| Hard | 20 | $3.00 each | $60 |
| Total | 100 | $84 |
Versus running everything on the expensive model: 100 × $3.00 = $300.
That's a 72% cost reduction. And the quality difference? For the routine tasks, there isn't one. For the judgment calls, it's marginal. You only lose quality on the 20 hard tasks — and those still get the expensive model.
What to Watch Out For
Don't over-optimize. Start with the obvious routine tasks in Bucket 1. Moving too many tasks to the cheap model too fast creates quality problems.
Re-evaluate quarterly. Models improve. What needed the expensive model six months ago might be handled by the cheap one today.
Keep a human in the loop on Bucket 3. This isn't about replacing judgment — it's about not wasting expensive AI on tasks that don't need it.
The Bottom Line
The 40% rule isn't about being cheap. It's about being intentional. AT&T didn't switch to cheap AI — they switched to the right AI for each task. That's the move.
Start tonight. List your last 10 AI tasks. Sort them. You'll probably find 4 that could run on a cheaper model without you noticing the difference. That's your 40% — and your savings.
SMB Scale Up helps small businesses use AI without the enterprise price tag. Follow for practical AI frameworks, not hype.
Top comments (0)