VDF AI Networks

Smart model routing

Let VDF AI pick the right model for each step automatically — by quality, by sustainability, or by what your organization and regulators allow.

3 min read
On this page

Why this exists

Inside every network, every step that talks to an AI has to use a model. You could pick the model yourself for every step. For some steps, you may want to. But for most steps, the right answer is: let VDF AI pick the model for you.

The reason is simple. Different steps benefit from different models. A quick summarization step doesn’t need the most powerful model on the planet. A nuanced legal analysis probably does. A step you run a hundred times an hour should optimize for cost. A step you run for a regulated workflow should stay inside what compliance allows.

Smart routing makes those choices automatically — and tells you what it picked so you can override when needed.

You can think of smart routing as a teammate who knows the model catalog cold. You describe what the step is trying to do. They pick the model that fits.

The three routing modes

Smart routing offers three styles. You pick one per network — or one per step if you want finer control.

Auto

Pick the best model for the job. Optimizes for quality first, with a sensible eye on cost and speed.

Sustainable

Favor models with lower energy and carbon impact. Quality stays high; the model choice tilts toward more efficient options.

Regulated

Stay inside the models your organization or regulator permits. Useful for compliance-bound workflows.

Auto

The default. Auto looks at the step — what it’s doing, what kind of input it has, what kind of output is expected — and picks the model that fits.

A few patterns Auto tends to follow:

  • A quick summarization step gets a fast, cost-efficient model. You wouldn’t notice the difference in quality, and the run finishes sooner.
  • A nuanced reasoning step — a critique, a strategic analysis — gets a high-quality model. Worth the extra time and cost.
  • A formatting or transformation step gets the lightest model that can do the job correctly.

Auto is the right choice for most networks. You’ll often hear it described as “let the platform do the boring optimization while you focus on the workflow.”

Sustainable

Same job. Greener model choice.

Sustainable mode applies a second filter on top of Auto’s quality decision: among the models that would do the job well, prefer the ones with lower energy and carbon impact.

What this means in practice:

  • The platform leans toward smaller, more efficient models where they’re capable enough.
  • For steps where only a large model can produce quality results, sustainable mode still picks that model — it doesn’t sacrifice the outcome.
  • You see the routing decision and the energy estimate per run, so you know what you saved.

Sustainable mode is a great default for:

  • High-volume networks. A scheduled morning briefing that runs every day across hundreds of customers has a real carbon footprint. Sustainable mode makes that footprint smaller.
  • Background work. Networks that run unattended on a schedule don’t need the absolute highest quality — they need consistently good results at the lowest reasonable cost.
  • Organizations with sustainability targets. When your team has a public commitment to reduce AI’s environmental impact, sustainable mode is the easiest lever to pull.

Sustainable isn't slower. A common misconception is that greener models are slower. The opposite is often true — smaller models run faster. Sustainable mode usually produces faster runs and lower costs at the same time.

Regulated

For workflows that operate under organizational or regulatory constraints, regulated mode keeps every model choice inside a permitted list.

This matters in three common situations:

  • Data residency. Some organizations require models that run inside a specific region or on specific infrastructure.
  • Regulatory frameworks. The EU AI Act and similar regulations limit how certain models can be used for certain decisions. Regulated mode enforces those limits at the routing layer.
  • Internal policy. Your organization may simply have decided that certain workflows should use certain models — or never use others.

In regulated mode, the platform won’t silently fall back to a non-compliant model if a permitted one isn’t available. Instead, it tells you a permitted model is needed and lets your admin handle the situation explicitly.

For more on how regulated routing connects to policy and audit, see Policies and budgets and Governance and admin.

How a routing decision actually looks

Every run shows you which model each step used and why. A typical step might show:

  • Step: “Critique the draft for clarity and tone.”
  • Mode: Auto
  • Picked model: (name)
  • Why: High reasoning demand, no constraint on latency, ample budget.

For sustainable runs, you’ll also see an energy estimate. For regulated runs, you’ll see which policy the model satisfies.

You don’t have to look at this every time. But when a result is unexpected, the routing detail is often the first thing to check — sometimes the issue is that a particular step needed a heavier model than Auto picked.

Overriding the routing

For any individual step, you can override the routing decision and pin a specific model.

Two common reasons to pin:

  • You’ve tested both options and one is clearly better for this step. Pin the winner.
  • You’re optimizing for cost at a step and you know exactly which cheaper model is good enough. Pin it.

Pinning a model takes that step out of the smart-routing pool. The other steps in the network continue to route as you’ve configured them.

A useful rule: pin sparingly. Each pinned model is a small commitment that won’t update as new, better, cheaper models arrive. Auto and Sustainable modes get those improvements automatically.

Mixing modes within a network

Most networks pick one routing mode. A few find it useful to mix.

  • Sustainable across the network, with one step on Auto. Cost-effective for the bulk of the work, with a single critical step getting the best available model.
  • Auto everywhere except one regulated step. The bulk of the work is unconstrained; one specific decision is governed by policy.
  • Regulated across the network with a few Auto steps for non-decision-making work. Useful when the workflow has a regulated core and a few descriptive or formatting steps that don’t need the same constraints.

The platform makes the mix explicit per step so a teammate can read the network and understand the choices.

A short FAQ on smart routing

Will the model my network uses change over time?

In Auto and Sustainable modes, yes — and that’s part of the point. As better models arrive, the platform routes to them automatically. You benefit without having to update every network.

Can I see what changed?

Yes. The run history shows the model used for each step on each run. If you notice a quality shift between runs, the routing log is usually the first place to look.

What if Sustainable picks a model that produces worse output?

Tell us. Sustainable mode is calibrated to keep quality high while reducing impact. A clear case where quality dropped is feedback we use to retune.

What about cost?

Cost is one of the inputs to routing. None of the modes will ignore cost completely — but Auto and Sustainable both keep cost reasonable, and you can layer a hard budget on top via Policies and budgets.

Smart routing is opt-in for change. If your workflow needs to produce identical results across runs for compliance reasons, pin the models. The benefit of smart routing — automatic improvement — is also a kind of change. Pin where stability matters most.

Where to go next