Your competitive advantage shouldn't live on someone else's model
Orchestrate the right models, verify every answer, and own the outcome. Annie is the verified AI orchestration platform for mid-market companies that run consequential AI work and want to keep their edge.
The February-April 2026 timeline
February 2026: Figma and Anthropic are partners — an integration underway, Anthropic's CPO on Figma's board. The relationship is productive, the press is friendly, and the strategic alignment looks real.
14 April 2026: he resigns from the board. Three days later, on 17 April, Anthropic launches Claude Design — a product aimed at Figma's core market. The stock drops approximately 7% that day. The integration roadmap is now an awkward footnote.
Nothing improper occurred. No data was taken. No contract was breached. And that is the lesson: the contract protected Figma's data. Nothing protected Figma's market.
The risk isn't that they break their terms. It's that everything can go exactly according to the terms — and you're still captured.
Three ways a frontier API dependency erodes your moat
None of these require bad faith. All are visible in 2026. And none is fixable with a better contract — only with ownership.
They can become your competitor
Model providers are moving up the stack into their customers' markets — partners included (Figma), developer-tool customers included. Your contract governs your data, not their roadmap. The partnership you signed in February is not the partnership you have in April.
They control your access and economics
Your unit costs are their pricing decision. Your uptime is their uptime (June 2026 was a preview). Your product changes when they deprecate a model. Access itself has been lawfully withdrawn from companies before, and the legal reasoning will hold up in court.
Your improvements build their asset
Every prompt refined and workflow tuned deepens your dependence — while your underwriting logic, triage rules, and pricing judgment get encoded into processes that only run on infrastructure you'll never own. The compounding runs the wrong direction.
Own the weights. Own the improvement loop. Verify everything.
Three claims, mapped to the Annie architecture. Each is a structural property of how the platform works, not a marketing promise.
Own the weights
Small, domain-tuned specialist models — trained and fine-tuned on your data, running in your cloud tenancy or on your racks. No supplier decision (pricing, deprecation, competitive entry, outage) can reach you.
Annie Workforce models + bring-your-own-model pipeline roles
Own the improvement loop
The compounding value in AI isn't the base model — it's your model getting better on your data, nightly. Annie's Cognition Stream runs automated fine-tuning cycles from $5 per run, so your moat deepens every night, for you.
Cognition Stream fine-tuning cycles
Verify, don't trust
Every consequential output passes an eight-stage pipeline: routed to specialists, checked against your sources of truth, disagreement surfaced rather than averaged away, full reasoning trail retained. That's how specialist models beat a generalist on your work.
Eight-stage Cognition Stream pipeline
When does owning beat renting?
Frontier API customers pay per token. Annie customers pay an annual platform licence scoped to their deployment. The difference at scale is the difference between a number you can put in a budget and a number you have to defend every quarter.
| Frontier API (rented) | Annie (owned) | |
|---|---|---|
| Inference cost at volume | Per-token, supplier-set, rises with usage | 10-100x lower. $18K-$73K annually at 100M tokens/day vs $550K-$1.8M |
| Model improvement | Prompt labour; gains accrue to the supplier's model | $5/run automated fine-tuning; gains accrue to yours |
| Price & availability risk | Unhedgeable — the supplier sets both | Yours to manage. Pricing locked at deployment. |
| Sovereignty | Level 1-2: API dependency, foreign jurisdiction | Level 3-4: customer holds keys, runs on their infrastructure |
| Exit cost | Everything stranded. Process, prompts, fine-tunes — all gone. | You own the stack. Models, weights, pipeline, data. |
The honest crossover
Below a certain volume and criticality, renting is rational — an API key is cheaper than a platform. The crossover comes when AI runs consequential work at real volume: above approximately 10M tokens/day, self-hosted wins on cost by an order of magnitude. Add sovereignty requirements to the same calculation and the crossover shifts lower. If that's not you yet, take the briefing paper and come back when it is.
Built for companies whose process knowledge is the moat
Specialty insurers, MGAs & brokers
Underwriting judgment and triage rules are the business.
The accumulated decision logic that lets you win the right risks and decline the wrong ones.
Submission triage or underwriting support, benchmarked against your current approach.
Vertical SaaS & AI-native products
Your product runs on a frontier API today. Your supplier ships products like yours tomorrow.
Margin that's your pricing decision today becomes your supplier's pricing decision tomorrow.
Your core AI feature on owned specialists.
Firms hitting the API cost wall
AI works, which is the problem: usage is scaling faster than budget.
Every successful AI deployment makes the per-token bill more painful to sustain.
Your highest-volume inference path, re-costed.
Government, defence, or APRA-regulated? You want the sovereign track.
At a price most companies can't touch
Palantir and Nvidia have launched an air-gapped stack where customers own the weights
Palantir's CEO says out loud that customers want to "own the means of production" — a deliberate echo of older language about who controls the productive assets of an economy. The diagnosis is right. Frontier labs are vertically integrating into customer markets, and the structural exposure to capture is real.
The prescription is priced for governments and giants. Annie is that architecture at mid-market economics — ownership without the eight-figure entry ticket. Same sovereignty posture, same verification layer, same compounding improvement loop. Sized for companies whose AI work matters but whose annual IT budget is not a sovereign wealth fund.
Don't take the thesis on faith — benchmark it
Every number on this site links to a methodology. Bring your technical diligence — we'll show our working. The full evidence pack is available under briefing-paper §6 once a working relationship is established.
| Claim | Magnitude | Source |
|---|---|---|
| Hallucination reduction | 4-67% vs single-model baselines | Multi-model consensus research, 2024-2026 |
| Calibration error reduction | 49-74% across medical benchmarks | Multi-model consensus, medical-domain studies |
| Domain specialist accuracy | 80.63% vs 75.85% for a 120B generalist on financial QA | Articul8, 2026 |
| Inference cost advantage | 10-100x lower than frontier APIs at enterprise scale | Public frontier API pricing, mid-2026 |
| Fine-tuning economics | From $5 per Cognition Stream cycle | Annie Cognition Stream pricing |
Flagship case study — coming
The first Annie case study will document a real deployment against the same workload running on a frontier API. Head-to-head numbers: accuracy on domain tasks, cost at volume, deployment timeline. We will not publish a case study until we have one.
What buyers actually ask
Can a small specialist model really match GPT-class quality?
On bounded domain tasks, yes — with verification. A fine-tuned specialist running through Annie's consensus pipeline catches errors a generalist misses. The 4-18% accuracy improvement on domain tasks comes from independent cross-model evaluation, not from raw parameter count. On open-ended general tasks, no — which is why Annie's pipeline can include frontier models where policy permits, just never unverified.
Don't the API providers' terms already protect my data?
Largely, yes — and that is not the risk. The data-protection clauses work as written. The structural exposures (supplier becomes competitor, supplier controls economics, supplier captures improvement loop) survive every contract. A better contract does not change the geometry.
What does it cost compared to what I spend on APIs?
At 100M tokens/day, frontier APIs cost $550K-$1.8M annually. Annie self-hosted costs $18K-$73K annually plus one-time hardware investment. The crossover for self-hosted sits at around 10M tokens/day. Below that, frontier APIs win on simplicity. Above it, self-hosted wins on cost by an order of magnitude. Sovereignty requirements shift the crossover lower.
Do we need an ML team to run this?
No. Annie operates the orchestration layer, the verification pipeline, and the Cognition Stream fine-tuning cycles. Your team operates the specialist models on your data. The 16-week implementation program covers infrastructure setup, model integration, fine-tuning, integration testing, and handover. After deployment, the platform runs itself.
Can we keep using ChatGPT / Claude alongside it?
Yes. The open pipeline roles mean you can mix owned specialists and frontier models in the same architecture. Owned specialists handle the consequential, high-volume work. Frontier models fill roles where appropriate, always verified through the consensus pipeline before output reaches you.
How fast is a pilot?
Working session first — one hour, no commitment, one workflow that matters. Then a scoped pilot on that workflow: 4-6 weeks for the pilot deployment with measured comparison against your current approach. Then a measured decision on whether to scale. Typical full engagement runs 16 weeks from briefing to production deployment.
Where does it run?
Your cloud tenancy (AWS, Azure, GCP — including sovereign cloud regions), on your own infrastructure, on a sovereign data centre partner, or fully air-gapped for classified workloads. The weights are yours in every deployment model. No supplier decision can reach the model.
Start with the paper, or start with a workload
Two ways to engage. The briefing paper covers the argument, the architecture, and the economics in detail — for the stakeholder who needs to brief internally. The 30-minute working session is for the practitioner who has a specific workflow in mind. Either way, no sales pitch.
Working session · One workflow that matters · We'll show what owning it looks like