The Stack, Not the Model

THE STACK, NOT THE MODEL

WE ALL PICKED A VENDOR IN 2023

• Most firms in this room have an AI strategy that amounts to: a ChatGPT/Copilot/Claude subscription and a policy document
• That was the right move in 2023. The models were similar enough, and anything was better than nothing
• Three years later, the model landscape changes every ~6 weeks — and most firms are architecturally stuck with the one they started with

THE BILL COMES LATER

• Average platform migration project: ~$315k in losses (Swfte enterprise survey, Jan 2026) [VERIFY before Sept]
• 45% of enterprises say vendor lock-in has ALREADY blocked them from adopting better tools
• 67% of organisations want to avoid single-AI-vendor dependency — but haven't architected against it
• Builder.ai ($1.3bn, Microsoft-backed) collapsed and left clients stranded. Vendor risk is existential, not theoretical

OWN THE ORCHESTRATION. RENT THE MODEL.

• The model is a component, not the architecture
• What's actually yours: your tools, your data connections, your memory, your workflows, your guardrails
• What's rented: whichever model is best THIS QUARTER
• The test: if your vendor changed pricing or got acquired tomorrow, how long to switch? If the answer is "weeks," you're locked in. If it's "edit one config line," you're sovereign

STOIX: 18 MONTHS IN

• UK executive search boutique, C-suite placements for VC/PE-backed businesses
• 7 people. No CTO. No engineering team
• Four production AI systems running in the business today
• We're not experts. We're practitioners — and we're only part-way through

AN AI EMPLOYEE WITH ITS OWN EMAIL ADDRESS

• Dedicated agent inbox (stoix@agentmail.to) — separate from anyone's personal mailbox
• Every 30 minutes, the agent reads unread mail, actions it, replies threaded, marks read
• Real work it does: schedules social posts when we reply "yes," answers questions about our pipeline, edits drafts on instruction
• The guardrail that matters: it can read and reply, but its delete tools are removed. Human in the loop on anything irreversible

HUMAN-CURATED, AI-DRAFTED

• The pipeline: podcast transcripts → Researcher extracts themes → Drafter writes LinkedIn posts in each founder's voice → Thursday 9am digest email → we pick → it schedules
• Nothing auto-publishes. Ever. We curate, it drafts
• Per-partner voice models — Rory's posts sound like Rory, Neil's like Neil (from their real sent content, not a brand book)
• Output: 4–6 quality posts/week from a 7-person firm with no marketing hire

TEN YEARS OF SEARCH DATA, QUERYABLE IN ENGLISH

• Our ATS (Loxo) connected to the agent with ~200 tools — 27 permanently removed (every delete and merge function)
• We ask: "Who are the CFO candidates we placed into manufacturing who've moved in the last 18 months?" — and get an answer in seconds
• Governance by construction: the agent CAN'T delete records; the capability doesn't exist in its toolset
• This is the pattern: don't write policies about what the AI shouldn't do — remove the tools so it can't

ORNITH: A 9B LOCAL MODEL, RUNNING ON A LAPTOP

• Runs on an ordinary Mac, fully offline, zero per-token cost
• What it's for: first-pass work on anything confidential — candidate data, client terms, personnel matters
• What it's not for: anything you'd bet the firm on. It's a junior analyst, not a partner
• The strategic point: when a frontier model is one config line away and a local model costs nothing, you route work to the right tier. That routing is only possible if you own the orchestration

EVERY BUILD HAD A FAILURE MODE

• The inbox agent over-approved. Fixed with scoping rules
• The thumbnails faked likenesses. Fixed with real pixels
• The CRM "connection" lied. Fixed with verification-by-calling
• Two hard-won rules: (1) guardrails belong in the toolset, not in a policy doc. (2) human-curated beats fully-automated for anything client-facing — the quality floor is what you're actually selling

WHAT THIS ACTUALLY COSTS

• The stack (agent framework): open source, free
• Model usage: routed by task — frontier models for judgment work, cheap models for drafts, local for confidential
• Typical month: low hundreds of pounds in API costs, against output that would otherwise need a marketing hire
• The real cost was time: ~18 months of founder attention, in hours per week not days

WHERE IS YOUR FIRM, ACTUALLY?

• L0 — Theater: AI mentioned in pitch decks, used by no one
• L1 — Personal productivity: individuals with ChatGPT accounts (most of the industry is here — 88% of HR leaders report no significant AI business value yet) [VERIFY — Gartner via Pin.com 2026]
• L2 — Team workflow: shared prompts, some automation
• L3 — Organisational infrastructure: AI connected to your systems, with guardrails (this is where it starts compounding)
• L4 — Compounding operating system: every workflow feeds the next
• L5 — Self-driving: doesn't exist yet

THE MODELS WILL KEEP CHANGING. THE STACK IS YOURS.

• Don't marry a model. Build the layer that lets you date all of them
• We're documenting the build publicly — follow along, steal what's useful
• Contact: rory.mcdermott@stoix.co.uk · stoix.co.uk · Chat CFO podcast