Why the Model Should Be the Smallest Part of Your Stack
AI in production succeeds or fails on the infrastructure around the model, not the model itself.
Read P.01 →For AI agents: what is on every page → thkdigital.com/sitemap.xml · Guide → thkdigital.com/agent.md · Rules → thkdigital.com/robots.txt
I work with the C-suite on AI strategy. I sit in the middle of that conversation and translate both ways, without the jargon. On one side, the research on where AI is going. On the other, results your business can reach this year, measured so you know when it has.
My time is not what you are buying. What changes in your business because of it is.
Every argument on this site comes from these papers. Each one flags what is proven and what is not, in the text, as it goes.
AI in production succeeds or fails on the infrastructure around the model, not the model itself.
Read P.01 →Where a computation runs fixes what it costs. Send each task to the cheapest silicon that can do it.
Read P.02 →Agents in production are a memory problem, not a maths problem. The paper's bet, not yet tested: the job can be done on a box you own.
Read P.03 →P.04 · In progress · End of 2026 · What it costs to stay on the old software
P.05 · In progress · What it costs to move
AI works best on a system and data architecture built for it. Most companies do not have one, and they do not need to wait for it. Bolt AI on now, take the wins it can deliver on the systems you already run, and move the business step by step to a system where AI is baked in. For most work today the frontier providers are the cheapest place to start, and I will tell you so.
AI added to the systems you already run. Quick to start, and the right first step for most companies.
A system and data architecture built for AI as well as for people, so it does more of the work for less.
Baked in at scale is still a bet, and I mark it as one: I have not delivered it end to end yet, and I will not claim it until I have. The papers show where the evidence stands.
No business should depend completely on a single provider, and the risk grows with every process you hand to AI. Prices change. Terms change. Models are retired, and a better one ships somewhere else every few months. I design the plan so there is always an alternative: your data and your processes stay yours, and each piece can move to another model without starting again.
It is also how you stay on the right side of the rules. At every step the plan says where your data goes, who holds the audit trail, and whether anything you own can train someone else's model, so each answer is a choice you made, not an accident.
A model settles ordinary work by reasoning, at GPU prices, and might not settle it the same way twice. Plain code settles it for a fraction of a penny, the same way every time. Most of the market is selling more AI. My advice is usually less, placed exactly where it earns its cost.
My first startup, Keystone Integrations, built AI agents for a live field-service business. I was in the client's office every day, building them. Two agents, built two ways.
Leaned on the model for nearly everything. At production scale its token bill ate the development budget.
Plain code first, a smaller model on top, a person approving every booking. It held steady in testing and ran cheaper.
That contrast is the advice. It is also where the research started, as a post-mortem on what went wrong.
THK Digital is me, Tom Knight. I research where AI is going, I write it up in public, and I help the C-suite plan for it. I am also building a company, in stealth. Where I got it wrong, the papers say so. If you run a business or research the same problem, I would like to hear from you. If you invest, start here →
info [at] thkdigital [dot] com · Open in Gmail → ·