AI Implementation
I don't start with tools. I start with your operations, bottlenecks, and goals, including the parts where AI is the wrong answer.
Define the metric before selecting the modelA.I. won't transform your business. Understanding how to apply it will.
Your team's already using ChatGPT, Claude, or Copilot. The spend is adding up and it's getting hard to justify what the return is. The CFO wants to know what you're getting for it, and you don't have an AI engineer on staff to answer that question.
How we define value
Four pillars of a strategy that survives contact with reality
No vendor playbooks. Every engagement is built on first principles, cost discipline, and what your team actually does, not what a product demo promises.
Custom Training
Off-the-shelf models with good prompting and retrieval solve most problems. Custom training only when the capability gap justifies the investment, not as a default.
Start with existing models, train only when it countsToken Optimization
Most businesses send every request to the most expensive model. Model routing, prompt caching, and token budgets cut spend 40–60% by matching the model to the task.
40–60% cost reduction without quality lossHuman-in-the-Loop Governance
AI informs decisions; people make them. Every strategy includes defined human checkpoints for decisions that matter, from day one, not bolted on after something goes wrong.
Human review on every decision that counts
Why this matters
The consulting work and the writing feed each other
The cases on /work come directly from the same research I publish. When a company's AI spend climbs and nobody can explain why, the answer usually starts with something the system forgot.
Publication
Adding Context
A Substack on memory architecture, model routing, and the infrastructure that makes AI actually work, regardless of which model you're using. The writing is the exhaust of real engagements.
Thesis
Memory is the moat
Every AI deployment hits the same wall: context degrades, agents forget, and nobody designed the system to persist what matters. The fix isn't a better model. It's an architecture that remembers.
Services
Built on the same pattern
Strategy, routing, governance, and integration, all six disciplines below are expressions of the same principle: AI should inform, people should decide, and nothing important should get lost between sessions.
Where the work applies
Expertise, mapped to the industries that benefit most
Expertise
Six disciplines, integrated so strategy, automation, cost, and oversight stay in lockstep.
- 01AI Strategy ConsultingDirection before tooling, a strategy built around your operations, not a vendor's roadmap.
- 02Operational AutomationIdentify the repeatable workflows where AI compounds, and remove the ones it shouldn't touch.
- 03Model Routing & Cost OptimizationRoute each request to the cheapest model that meets the quality bar. Prompt caching, token budgets, and semantic routing cut spend 40–60% without capability loss.
- 04Custom Model TrainingFine-tuning and retrieval design, only when the capability gap justifies the investment, not as a default.
- 05Workflow IntegrationDeployment into the systems your team already uses, adoption is the real test.
- 06Human-in-the-Loop GovernanceDefined human checkpoints for decisions that matter. AI informs; people decide. Oversight built in from day one, not bolted on after failure.
Where this shows up most
Industry patterns where the same three conditions keep surfacing. These aren't the only places I work, they're where the problem is most visible.
- 01EducationStreamlining admin & curriculum so educators spend time on teaching, not paperwork.
- 02MarketingData-driven campaign scaling with measurable signal, not vanity metrics.
- 03Manufacturing & Supply ChainPredictive maintenance, demand forecasting, and inventory optimization grounded in real operational data.
- 04Legal & Professional ServicesDocument review, contract analysis, and knowledge retrieval that bill down overhead without losing rigor.
Is This You?
Three things are true of every engagement I take on
The same pattern shows up across industries. If all three sound familiar, the strategy has something to optimize.
01 Already bought in
You're paying for AI.
ChatGPT, Claude, Copilot, or all three. Adoption already happened. You're not evaluating whether to use AI, you're wondering what you're getting for the bill.
02 The bill is becoming a problem
Every request hits the most expensive model.
No routing logic, no cost discipline, no governance. The CFO is starting to ask questions and you don't have a clear answer.
03 No in-house AI engineer
Big enough to spend. Small enough to need help.
20 to 200 employees. Nobody on staff who can build model routing, audit token spend, or design governance around AI output.
Savings Roadmap
Reclaim Your Wasted AI Budget
If you're spending more than you budgeted on ChatGPT, Claude, Gemini, or Microsoft Copilot, you're likely sending every request to a frontier model when a smaller one would do. Get a custom savings roadmap, a breakdown of where your token spend is leaking and the specific model routing moves that bring it back under control.
Q&A
Questions operators ask before committing to an AI strategy
Real questions, plain answers, structured for easy reference by humans and by the models your customers ask.
A vendor sells a product; a consultant starts with your operations, bottlenecks, and goals. An AI strategy consultant should challenge whether AI is even the right tool for the problem, and build a plan grounded in first principles that fits how you actually operate. If the conversation starts with a product demo, you're talking to a vendor. If it starts with questions about your workflows, metrics, and constraints, you're talking to a consultant.
Every engagement defines the success metric before any model is selected, cycle time, cost per claim, campaign lift, admin hours saved, error rate reduction. The strategy is measured against that baseline, not against a generic "AI adoption score." If you can't define what "better" looks like in a number, you're not ready to invest in AI yet. The first deliverable should always be a measurable baseline, not a proof of concept.
Runaway API spend almost always comes from one source: sending every request to the most expensive model regardless of task complexity. A classification task doesn't need the same model as a complex legal analysis. The fix starts with an audit of what each request is actually doing, then routing simple tasks to smaller, cheaper models while reserving frontier models for the work that genuinely requires them. Add prompt caching, token budgets, and cost-per-outcome tracking, and you can typically cut spend 40–60% without losing quality on the tasks that matter.
Most businesses don't need custom training. Off-the-shelf models, combined with good prompting, retrieval-augmented generation (RAG), and careful task routing, solve the vast majority of real-world problems. Custom training only makes sense when you have proprietary data that creates a genuine capability gap, and when the cost of training is lower than the cost of working around the gap. The default should always be: start with existing models, prove the workflow, then train only if the capability gap justifies the investment.
The common thread isn't the industry, it's whether the same three conditions apply: the company has already bought into AI, the bill is becoming a problem, and there's no in-house AI engineer. Education, marketing, manufacturing, and legal & professional services are where the pattern shows up most frequently, in streamlining admin, scaling campaigns, optimizing supply chains, and reviewing documents. But any business with 20 to 200 employees, a repeatable workflow, a baseline metric, and a decision-maker willing to commit to change has something to optimize.
Three signals: you have a repeatable workflow with a measurable baseline, you have someone who owns the outcome and can commit to changes, and you have clean enough data that a model can actually learn from it. If any of those are missing, the first step isn't AI, it's fixing the workflow, assigning ownership, or organizing your data. AI amplifies whatever process you feed it. If the process is broken, AI just makes the brokenness faster and more expensive.
Almost certainly. The market for capable, smaller models has grown dramatically, open-source models like Llama, Mistral, and others can handle many tasks at a fraction of the per-token cost of frontier models. The key is knowing which tasks require a frontier model and which don't. Most businesses are paying premium prices for routine work that a smaller model handles just as well. A proper model audit maps each task to the cheapest model that meets the quality bar, often reducing spend by 50–80% on the tasks that don't need frontier capability. You don't need to abandon ChatGPT or Claude, you need to stop using them for everything.
Model routing is the practice of sending each request to the model best suited for its complexity. A simple classification or extraction task goes to a small, cheap model. A nuanced legal analysis or creative brief goes to a frontier model. Semantic routing takes this further by analyzing the meaning of each request and directing it automatically, no manual rules to maintain. The result is that you're paying premium rates only for the work that genuinely requires premium capability. Most businesses see 40–60% cost reduction with no perceptible quality loss, because the tasks being downgraded were never using the frontier model's full capability in the first place.
Starting with the tool instead of the problem. Businesses buy a frontier model subscription, point it at a vague workflow, and hope for transformation. What they get is unpredictable costs, unmeasurable results, and a team that doesn't trust the output. The second biggest mistake is removing human oversight too early, assuming the AI can handle edge cases it was never trained on. The pattern that works is the opposite: define the metric, map the workflow, start with a narrow use case, keep humans in the loop for decisions that matter, and expand only when the results are measurable and the team trusts the system.
AI should inform decisions, not make them autonomously, especially in high-stakes or customer-facing contexts. Human-in-the-loop (HITL) means a person reviews, validates, or overrides AI output before it becomes action. Human-on-the-loop means the AI acts, but a person monitors and can intervene. The right model depends on the risk profile of the decision: routine tasks can tolerate more autonomy; decisions involving money, customers, or compliance need a human in the loop. Businesses that skip oversight to save time end up with expensive, hard-to-debug failures, and no one who can explain why the decision was made. Every strategy I build includes a defined human checkpoint for decisions that matter, from day one, not bolted on after something goes wrong.
Submit a strategy request at the contact section below. You'll get a short intake call focused on operations, bottlenecks, and goals, no sales deck, no tool-first pitch.
Get in touch
Ready to cut the noise and start measuring impact?
Get in touch so we can better understand your business. You'll hear back with a short intake focused on operations, bottlenecks, and goals.