Before You Celebrate AI Adoption, Write Down Which Workflows Get the Expensive Model

I've watched a lot of founders get excited about the same number this year: per-employee AI usage is up. Agents are running. People are leaning in. Anthropic's 2026 report says 81% of businesses plan to move past simple task automation into more complex projects, and over half are already deploying agents for multi-step work. That feels like momentum.

Here's what that number doesn't tell you. It doesn't say which model handled which task, what it cost, or whether the result was worth the spend. Adoption is a usage stat. It is not an outcomes stat.

Boston Consulting Group's maturity index found only 5% of enterprises are genuinely "future built." Forty-six percent are still "emerging" — using AI, struggling to get value from it. The gap between those two groups isn't how much AI they use. It's whether anyone decided, in writing, what gets routed where.

That decision has a name. It's a model-routing policy, and most operators celebrating adoption don't have one.

The old question was "which model is best?" That question is dead. The explosion of specialised models and agentic systems means the real question is the one Anna Maria Knapton frames cleanly: which intelligence should perform which task?

Not every workflow needs your most powerful, most expensive model. Some need deep reasoning. Some need a cheap classifier. Plenty need a human to sign off before anything moves. When you don't draw those lines, the default is to throw the premium model at everything — which is exactly how AI stops being leverage and becomes a line item nobody can defend.

IDC frames model routing as dynamically directing each request to the most context-appropriate model: good-enough requests to cheaper models, high-stakes ones to premium models, sensitive data only to approved or private models. They expect 70% of top AI-driven enterprises to be running multi-tool architectures for autonomous routing by 2028. The operators who win aren't the ones who adopted fastest. They're the ones who decided where each task goes.

Let me make this concrete, because "workflow criticality" sounds like a slide until it's your Tuesday.

An agent triaging your Gmail and flagging which threads need a reply today — that's a classification task. Cheap model, all day. You don't need frontier reasoning to spot that the investor follow-up matters more than the SaaS renewal notice.

Drafting the actual reply to that investor, pulling context from the last six emails and a HubSpot deal record? That's higher stakes. Get it wrong and you've sent something off-tone to someone who matters. That one earns the premium model — and arguably a human read before it sends.

Now the messy middle. An agent updating a Pipedrive stage, creating a Linear ticket from a Slack thread, pulling a number off a Stripe dashboard into a Notion doc. Mechanical, structured, low-judgment. Default those to the cheaper model and stop paying frontier prices for data entry.

The pattern is simple once you see it: classification and mechanical moves go cheap, judgment and outbound communication go premium, and anything irreversible gets a human in the loop. Most operators have never written that down, so the agent guesses — and it tends to guess expensive.

The most common reason AI initiatives die is that they start with use cases instead of workflow analysis. You get a fragile pilot that demos beautifully and never turns into anything durable.

So before you route anything, map how work actually moves through your week. Which steps can be fully automated. Which can be accelerated but still need your eyes. Where your judgment is genuinely non-negotiable. That map is the spine of the policy — the model assignment hangs off it.

I'll be honest about where this is still hard. AI agents are very good at the structured, repeatable parts of your day — triage, summarising, drafting against a known template, keeping records in sync. They're far less reliable when a task needs taste, relationship context, or a call about something that can't be undone. A good Chief of Staff knows the difference between a task you hand off and a decision you keep. A routing policy is just that instinct, written down so your agents inherit it.

This is the part of Moments I care most about — making the routing decision legible, so the system isn't quietly burning your budget on work that never needed the expensive brain.

Routing isn't only about cost. It's about accountability, and this is where most policies are silent.

A real policy answers uncomfortable questions. Who owns each deployed workflow? Who decides what data the agent can touch — your Gmail, your Stripe records, your contacts? Who reviews and signs off on what the agent produces? And critically: when a cheap model returns a weak answer, who approves spending more on a premium retry?

That last one is the lever nobody installs. Without it, agents either give up too early or silently escalate to the most expensive option every time something looks uncertain. Both are failure modes. The fix is a defined escalation rule — a human approves the costlier retry on high-stakes work, and low-stakes work just accepts the cheaper answer and moves on.

This matters beyond economics now. States including California, New York, and Colorado have enacted AI laws requiring demonstrable responsible use. Shadow AI — staff quietly using whatever tool they like — already creates compliance gaps and security holes. A routing policy with audit trails is how you answer "who decided this, and why" before someone outside the building asks.

The point of all this isn't a clever architecture. It's compressed cycle time and better outcomes. Nextgov's framing applies just as well to a founder as to a federal agency: if your real metrics aren't improving, your AI adoption isn't aligned with intent.

For an operator, those metrics are tangible. Time from inbound lead to first reply. Time to get a contract out. Time to close the loop on a thread that used to sit for three days. If your agents are busy and those numbers haven't moved, you've bought activity, not leverage.

So invest in observability. IDC's recommendation is blunt: monitor performance, quality, and cost across every route, over time. You want to see that your cheap model is handling 80% of volume at a fraction of the cost, that your premium routes are reserved for what actually matters, and that quality held. Without that view you're flying without instruments — impressive moments, no durable impact.

Review the policy every six months at least. The models change, the prices change, your workflows change. A routing map that was right for last quarter's tools will quietly drift into waste.

You don't need a governance committee to start. Take your five most frequent AI-assisted workflows — the inbox triage, the deal updates, the drafting, the meeting prep, the reporting. For each one, write three lines: which model it should use, whether a human approves before it acts, and what a costly retry requires.

That's a model-routing policy in its first honest form. It'll be incomplete and you'll revise it. That's fine. What matters is that the decision now lives somewhere other than inside an agent's guess.

Adoption was never the finish line. It was the moment you started spending real money on intelligence. The question is whether you've decided where that intelligence goes — or whether you're about to find out at the end of the quarter.