AI costs are bleeding budgets dry. A single customer support workflow that routes tickets, extracts intent, and drafts responses can consume $0.50 to $1.00 per execution when built entirely on large language models. For a mid-sized Indian enterprise processing 10,000 such workflows daily, that is ₹4.1 lakh per day. ₹1.5 crore per year. For one workflow.
The Token-Light, Code-Intensive (TLCI) approach, developed by Carlos Gorricho and documented on GitHub, offers an escape. His benchmark workflows with 6 code steps and 2 AI steps cost $0.02 to $0.05 per execution. Same outcome. Same reliability. Eighty percent less cost.
This is not about using worse AI. It is about using intelligence only where intelligence is needed.
Decompose Before You Decide
Most AI cost overruns begin with a single mistake: treating an entire workflow as one indivisible task that "needs AI." The result is a prompt that tries to do everything, calls a frontier model for every execution, and burns tokens on work that never required reasoning in the first place.
The TLCI method demands decomposition first. Break every workflow into discrete, observable steps. For a typical customer service automation, this might yield:
- Step 1: Validate input format (code)
- Step 2: Check cache for duplicate ticket (code)
- Step 3: Classify urgency and category (small model)
- Step 4: Route to appropriate queue (code)
- Step 5: Draft personalized response (frontier LLM)
- Step 6: Apply compliance filters (code)
- Step 7: Log and notify (code)
- Step 8: Update CRM record (code)
Only steps 3 and 5 touch any model. The rest execute in milliseconds at zero token cost. This decomposition discipline, what BASH Consultants call the deterministic-first doctrine, is the foundation of controlled AI spending.
Deterministic work is cheaper, faster, and more reliable. Code does not hallucinate. Code does not drift. Code executes in microseconds and costs nothing per execution beyond infrastructure. The goal is not to eliminate AI. It is to reserve AI for the specific sub-problems where probabilistic reasoning adds value.
Route to the Right Execution Tier
Once decomposed, each step gets assigned to one of three execution tiers. This routing decision is where cost control happens.
| Tier | Technology | Cost per 1K executions | Best for |
|---|---|---|---|
| Code | Python, SQL, rules engine | $0 | Validation, transformation, routing, lookups |
| Small Model | DistilBERT, fine-tuned classifiers | $0.10-2.00 | Classification, sentiment, entity extraction |
| Frontier LLM | GPT-4, Claude 3 Opus, Gemini Ultra | $10-100 | Complex reasoning, generation, synthesis |
A Mumbai-based logistics company applied this routing to their shipment exception handling. Previously, every exception triggered a GPT-4 call to classify the issue and suggest resolution. Cost: $0.78 per exception. After decomposition, 94% of exceptions were handled by code (invalid address format, known carrier delay patterns, customer preference lookups). Only 6% reached a small classification model. Complex edge cases, roughly 0.3%, escalated to GPT-4. New cost: $0.04 per exception. Same resolution quality. Faster average response time.
The Octahedroid framework for choosing the right layer emphasizes this: automation handles the known, AI handles the uncertain. Most business workflows contain far more known than uncertain.
Implement Three-Level Metering
Routing intelligence to the right tier prevents waste. Hard budget controls prevent surprises. The TLCI approach implements metering at three levels, with every check occurring before any token spend.
Per-execution budget: Each individual workflow run carries a maximum token allowance. Exceed it, and the call fails closed, not open. No partial responses, no truncated outputs that create downstream errors. Clean failure with logged context.
Per-workflow budget: Aggregate limits by workflow type. Customer support automation gets ₹50,000 monthly. Invoice processing gets ₹15,000. Marketing content generation gets ₹30,000. When a workflow hits its cap, it halts and alerts. No cross-subsidy between workflows.
Global budget: A hard ceiling on total AI spend across all systems. This is the circuit breaker. If metering systems disagree or a bug causes runaway calls, the global budget stops the bleeding.
Each level is checked in sequence. The per-execution check takes microseconds. The per-workflow check queries a cached counter. The global check hits a separate, minimal service. If any check fails, no API call occurs. Cost prevented, not merely observed after the fact.
Pessimistic controls fail closed. If your budget system is unavailable, the workflow must not proceed. This creates apparent downtime, but that downtime is cheaper than uncontrolled spend. Many teams build "optimistic" controls that assume the budget system will catch up. They learn the hard way when a configuration error generates a ₹12 lakh weekend bill.
Build Observable Cost Records
Estimates create false confidence. Every TLCI workflow records actual costs per execution, not projections. Each log entry contains:
- Workflow ID and version
- Step-by-step tier assignments
- Actual token consumption by model
- Total execution cost in local currency
- Timestamp and correlation ID
This granularity serves two purposes. First, it enables continuous optimization. A Bangalore SaaS company discovered through per-execution records that their "small model" classification step was actually routing 15% of calls to the frontier tier due to confidence thresholds set too low. A threshold adjustment moved 12% of volume down-tier, saving ₹3.2 lakh monthly.
Second, it creates accountability. When the CFO asks what AI spend delivered, the answer is specific: this workflow, these executions, this outcome, this cost. Not a blended average that hides waste.
These records also feed back into decomposition. When a code step consistently fails and escalates to AI, that is signal. Either the rules need refinement, or the problem genuinely requires intelligence. Both learnings improve the system.
Secure Your Foundation
Cost control without standardization is temporary. Teams that build ad-hoc AI integrations without governance find their savings eroded by shadow implementations, model drift, and duplicated effort across business units.
The TLCI approach embeds standardization into its architecture. Decomposition templates, tier routing rules, and metering schemas become shared infrastructure. New workflows inherit the controls. Teams do not reinvent cost discipline for each project.
This standardization also enables security. When AI usage is routed through known chokepoints, data classification and access controls apply consistently. Sensitive data never reaches models inappropriately. Audit logs capture every interaction. The same infrastructure that meters cost also enforces policy.
The ElevAIte perspective: We have implemented TLCI architectures for Indian enterprises across manufacturing, financial services, and healthcare. The pattern holds: decomposition discipline and hard budget controls deliver predictable costs without capability sacrifice. The 80% reduction is not theoretical. It is measured, documented, and repeatable. The constraint is organizational will, not technical possibility.
Start With One Workflow
The complete TLCI transformation is significant. The first step is not. Select one high-volume workflow with unpredictable AI costs. Decompose it into discrete steps. Identify which steps truly require reasoning. Build code replacements for the rest. Implement per-execution metering. Measure the result.
Most organizations find their first TLCI workflow within two weeks. The cost reduction is immediate and visible. That success creates organizational appetite for systematic application. Eighty percent savings on one workflow is a proof of concept. Eighty percent savings on twenty workflows is a competitive advantage.
Intelligence is expensive. Use it precisely.
Sources
Share this article