The myth of the one model
Teams still argue about a single “best” model the way people once argued about a single best laptop. AI model routing starts from a boring truth: writing a friendly customer email, reviewing a spreadsheet, and planning a multi-week project are different jobs. Different jobs reward different strengths.
If everyone is locked to one personal subscription, they force every job through one strength profile. Routing is how a workspace escapes that trap without asking people to become model critics.
What routing means in practice
Routing means the system selects a model for a request based on signals—task type, desired effort, capabilities like vision or long context—then runs the request there. Good routing is quiet. Bad routing is a mystery box that burns money on simple tasks.
In Iymer Chat, Automatic mode is described as picking the best model for each question at your chosen effort. You can still open a searchable catalog when you want a specific family. That combination—default autopilot, manual override—is what makes routing usable for mixed teams. More on the AI Models Hub.
Effort is part of routing
Not every question deserves Max. Effort controls (for example Low, Efficient, Balanced, High, Max) tell the system how hard to work—and, practically, how much cost and latency you will tolerate. Teaching the team when to raise effort prevents both laziness and waste.
- Low / Efficient — quick drafts, rewrites, simple lookups.
- Balanced — default for most knowledge work.
- High / Max — ambiguous strategy, careful analysis, high-stakes writing.
When to pick manually
Override Automatic when you already know the fit: a model that handles your file type well, a style you prefer for exec prose, or a capability badge you need (vision, tools, long context). Manual pick is a power tool, not the everyday path for every teammate.
Routing and files
Attachments change the routing decision. A spreadsheet review and a screenshot critique need different capabilities. Good catalogs expose badges like vision, files, tools, or long context so manual overrides are informed. Automatic mode should respect those needs silently when it can.
Teach people to attach the artifact instead of pasting huge blobs when the product supports files. Cleaner inputs produce cleaner routing and clearer provenance for later readers of the chat.
Organizational defaults
Decide whether new teammates start on Automatic or on a named model. Most organizations should start on Automatic with balanced effort. Publish exceptions only when truly required. Exceptions should be short and rare.
Revisit defaults when major releases land. A model that was weak at structured planning last quarter might be fine now. Routing is not a set-and-forget religion; it is an operations habit you retune lightly.
Share a monthly note on what the catalog added and whether defaults changed. Transparency keeps power users from spinning up shadow stacks because they assume IT is behind.
Cost visibility without fear
Show teams roughly how effort tiers behave. People make better choices when Max is explained as a deliberate spend, not a mysterious premium button. Routing plus honest cost education beats silent caps that train people to circumvent the system.
A worked example
Suppose a PM asks for a competitive briefing with sources. Automatic with balanced effort should pick a model strong at research synthesis. If the same PM later pastes a screenshot of a broken UI, routing should prefer vision-capable options. If they ask for a tiny subject-line rewrite, a cheaper fast tier is enough.
Now suppose an engineer wants a careful threat model. They raise effort to High or Max and perhaps pick a preferred reasoning model manually. That combination—Automatic for everyday, deliberate overrides for stakes—is the practical shape of AI model routing on a real team.
Failure modes to watch
- Always-Max culture that treats every Slack question like a board presentation.
- Always-manual culture where newcomers freeze in the catalog.
- Stale defaults after a major release nobody re-tested.
- Hidden personal seats because power users distrust routing without evidence.
Each failure mode has a social fix and a product fix. Publish examples. Retest quarterly. Invite skeptics into the evaluation guild. Routing works when people believe it is on their side.
Field notes from teams who stick with it
The teams that get leverage from model routing share a few habits. They pick one painful loop before they expand. They write norms in plain language. They measure shared artifacts, not chat volume. They keep a human accountable for quality when automation is new.
They also refuse hype as a management tool. Short sentences about what changed Monday beat slogans about the future of work. When something fails, they narrow scope instead of buying another overlapping product.
A checklist you can copy
- Write the job to be done in one sentence a skeptic would accept.
- Name the system of record where outputs must land.
- Run a two-week pilot with a mixed-role group and one operator.
- Review ten real outputs for trust before you scale.
- Cancel or pause overlapping tools only after the new path is smoother.
- Schedule a thirty-day review with leadership using business outcomes, not vanity metrics.
If you cannot complete that checklist, you are not blocked by missing AI features. You are blocked by missing clarity. Fix clarity first. The software will amplify whatever operating system you already have, disciplined or not.
Use this article as a working doc. Highlight the section that matches your next bottleneck. Ignore the rest until that bottleneck moves. Focus compounds faster than feature tourism.
Prompt hygiene helps routing
Routing works better when prompts state the job clearly. Say summarize for executives in 200 words or review this spreadsheet for anomalies. Vague asks produce vague model choices and vague answers.
A tiny template library inside the team reduces both prompt flailing and unnecessary Max runs. AI model routing amplifies clear intent; it cannot invent intent that was never written.
What routing is not
- Not a guarantee of perfect answers—garbage briefs still produce messy outputs.
- Not a substitute for permissions and data policies.
- Not an excuse to hardcode a public “complete model list” that goes stale weekly.
- Not the same as running every request on the most expensive model “just in case.”
Teaching your team
Publish three example prompts with recommended effort. Show one override. Explain that new models may appear in the catalog over time without a new SaaS shopping trip. Make success about outcomes—clear briefs, fewer private seats—not about winning model debates in Slack.
How to evaluate routing
Run the same ten tasks with Automatic on balanced effort. Spot-check where you disagree. If you disagree often on the same class of task, document a manual default for that class. If you rarely disagree, stop micromanaging chips and let the system work.
One AI model is never best at everything. Routing is how a team admits that without turning model shopping into a second job.