Part 1 of 3: Reasoning in the Intelligence Era
This is the first article in a series on building AI-powered digital labour. Part 1 focuses on reasoning when Small Language Models fit, when they don’t, and how to make the decision deliberately. Parts 2 and 3 will cover perception and action.
What follows isn’t theory. It’s a framework distilled from deploying agentic AI systems for global enterprises where the tolerance for failure is zero. These patterns have been tested in production and validated against the only metric that matters: measurable business outcomes in environments that don’t forgive mistakes.
Now, onto the main article.
Every BigTech AI vendor has the same pitch: their model is the most powerful. Bigger is better. Buy the biggest.
This is vendor logic, not business logic.
SLM vendors aren’t immune either, many are selling proprietary small models at premium prices for tasks that open-weight models handle at a fraction of the cost. Same hype, different packaging.
Strategic leaders need a different frame, one that starts with how reasoning actually works.
What the Brain Teaches Us
The brain is not one massive general-purpose processor. It’s a modular architecture: distinct regions executing discrete cognitive functions, connected by integration hubs that coordinate when needed. Visual processing happens in specialised regions. Language in others. Motor control, spatial reasoning, emotional regulation each optimised for its function.
This architecture evolved because modularity enables both specialisation and flexibility. The most capable reasoning system in nature is not monolithic. It’s modular, specialised, and orchestrated.
How this translates to AI: Your invoice classification model is like the visual cortex specialised, fast, pattern-matching. Your exception reasoning model is like the prefrontal cortex judgement and decisions. Your orchestration layer is like the thalamus routing to the right specialist. You don’t need a frontier model doing all three. A 7B specialist will outperform a 175B generalist on the task it was trained for faster, cheaper, and under your control.
The Business Case for Reasoning
Where does adding reasoning capability create value? The business case exists when reasoning addresses these conditions:
1. Decision Volume Exceeds Human Capacity. Thousands of similar decisions daily. Value lever: Throughput.
2. Decision Quality Is Inconsistent. Different people apply different judgement. Value lever: Consistency.
3. Decision Speed Creates Competitive Advantage. Narrow windows to act. Value lever: Velocity.
4. Decision Context Is Scattered. Information across multiple systems. Value lever: Synthesis.
5. Decision Expertise Is Scarce. Specialised judgement is bottlenecked. Value lever: Leverage.
Example: Accounts Payable
Most large enterprises don’t run AP with expensive onshore teams. They’ve already optimised BPOs in Manila, Bangalore etc. Workflow engines, case management, SOPs, labour arbitrage. Costs compressed to $3-5 per invoice.
The value isn’t replacing cheap labour. It’s solving what cheap labour can’t.
Exception reasoning remains slow. Invoices that don’t match fall out of straight-through processing. Resolution: 3-5 days. Cost: 5-10× a clean invoice.
Early payment discounts are structurally broken. Capturing 2/10 net 30 terms requires validation within 10 days. When 35% of invoices hit exceptions, most enterprises capture only 25-35% of available discounts.
Expertise doesn’t scale. Complex exceptions require senior analysts who are scarce in low-cost locations.
Where reasoning fits: An SLM fine-tuned on your invoice patterns can analyse exceptions, synthesise context from ERP and contracts, and either resolve autonomously or recommend action. It identifies discount-eligible invoices at risk of missing the window. Senior expertise gets leveraged across more decisions.
Business case: 50,000 invoices monthly, 35% exceptions, 31% discount capture. With reasoning: 60% exception auto-resolution, 78% discount capture. Annual value: ~$1.1M against $200-300K implementation cost. Even at conservative assumptions (40% auto-resolution), value is ~$700K; still 2-3× return.
The pattern generalises: Claims adjudication, trade finance, customer escalations, regulatory reporting same structural conditions, same opportunity.
Two Kinds of Vendor Hype
Frontier vendors push maximum capability. Reality: you’re paying for breadth you don’t need and sending sensitive data to external APIs.
SLM vendors sell proprietary small models at enterprise pricing for tasks Mistral, Llama, Qwen, or Phi handle comparably. Reality: you’re paying premium for packaging.
The question isn’t big vs. small. It’s: What reasoning capability do I need, and what’s the most economical way to own it?
Open-weight models fine-tuned on your domain data often match proprietary alternatives at a fraction of the cost. Vendors on both ends benefit from you not knowing this.
Build vs. Buy
Buy when: Task requires genuine proprietary innovation. You lack ML capacity. Speed matters more than long-term cost. Compliance requires vendor contracts.
Build when: Task is bounded. Open models perform adequately. You can fine-tune. Long-term cost and control matter.
Three questions to assess build readiness:
- Engineers who’ve fine-tuned an open-weight model in the last 12 months?
- GPU infrastructure within your security perimeter?
- Labelled data from the target workflow?
Two yes: build. Zero: buy. One: build capability while buying time.
In our deployments, 60-70% of reasoning tasks are bounded, domain-specific, and repetitive exactly where fine-tuned SLMs outperform general-purpose models.
When SLMs Work
1. Privacy and Control Non-Negotiable. SLMs run on-premise. Data never leaves your environment. You own the weights.
2. Economics Must Scale. Frontier API: $15-60 per million tokens. Self-hosted SLM: $0.10-0.50. At 100M tokens/month, that’s $6,000 vs. $50.
3. Domain Expertise Beats General Capability. A fine-tuned 7B model reasons better about your work than a general-purpose giant.
When SLMs Don’t Work
Open-ended reasoning. Broad creativity and novel problem-solving favour larger models.
High stakes with limited verification. If you lack robust validation, frontier robustness may justify the cost.
True breadth required. Applications needing encyclopaedic knowledge within the model itself.
Making It Real: Governance, Risk, and Change
For your CISO: On-premise SLMs mean no data leaves your environment. Security posture improvement over API-dependent approaches.
For your board: Start advisory mode; AI recommends, humans decide. Measure for 90 days. Graduate to agentic only where the system proves it matches human judgement.
For your ops teams: Position as augmentation, not replacement. Less data gathering, more judgement work. Reframe roles before you deploy.
For your CFO: Business case works under conservative assumptions. Budget 6 months longer than vendor promises.
Monday Morning: What To Do Next
- Pick one high-volume workflow where you have mature operations
- Quantify exception cost and speed penalty
- Assess build readiness with the three questions
- Run 30-day proof of value advisory mode only
- Measure accuracy against human decisions before expanding scope
The future of intelligent work isn’t about buying the most powerful reasoning available. It’s about designing reasoning architecture you can own, control, and compound deployed where it actually earns its place.
No vendor will optimise for that. Make the decision deliberately.
Next: Part 2 – Perception: How AI Sees, Reads, and Understands.
Where are you seeing the clearest cases for adding reasoning to mature operations? I’m interested in what’s shaping those choices.