Credits Saver watches every message and quietly hands the easy ones (greetings, quick acknowledgements, 'make that shorter') to a lighter, cheaper model, while real work still runs on the model you picked. On everyday back-and-forth that often means 40 to 50% fewer credits, with no drop in quality where it counts. It is on by default and fully in your control: turn it on or off anytime, and see tracked credits saved, its router cost, and the net benefit on its own page.
Not every message your employee receives needs your best, most expensive model. "Hey", "thanks", "make that shorter" don't require deep reasoning, they require a fast reply. Credits Saver classifies every turn the moment it arrives and routes the easy ones to a lighter model automatically, no setup, no rules to write.
Real work never gets downgraded. The same classifier that catches small talk also recognizes when a message is actually asking for work, drafting an email, researching a prospect, writing code, and keeps that on the model you selected for the employee. The saving comes entirely from the messages that never needed the expensive model in the first place.
It's on by default across every plan and fully visible: its own settings page shows exactly how many turns it classified as chat versus work, tracked credits saved, its router cost, and the net benefit, so the discount isn't a black box.
This is what makes an AI employee affordable to run all day instead of only for high-value moments: the credits your employee would have spent answering "got it, thanks" go toward the research, drafting, and analysis you actually hired it for.
The routing isn't a fixed word list. An obvious greeting or thank-you in English, French, Spanish, German, Italian, or Arabic script is recognized instantly and never even reaches the classifier, that recognition step costs nothing. Everything else is compared by meaning against what a routine chat turn looks like versus what real work looks like, so a paraphrase, a typo, or a sentence in a language the list never anticipated still lands in the right lane.
Every incoming turn is scored before your employee replies. Greetings, acknowledgements, and requests to reshape an answer already given (shorter, friendlier, bullet points) score as chat. Anything that asks the employee to produce, research, decide, or act scores as work and keeps the selected model.
When the signal is ambiguous, the system defaults to work rather than chat. A missed discount costs a few extra credits; a downgraded real task costs quality, so the fail-safe always favors the safer, more expensive path.
A bare "yes" or "go ahead" is a special case: on its own it reads as idle chat, but it's usually approving whatever your employee just proposed. When Credits Saver sees a short approval like this, it pulls in what the employee's last message was actually offering to do and classifies the two together, so approving a real task keeps the full model instead of quietly running it cheap just because the approval itself was short.
Credits Saver only changes which model answers a turn, never what the employee is allowed to do, remembers, or has access to. A cheap-model reply still has the same tools, memory, and guardrails as an expensive-model one; the only difference is which model generated the text.
It also stays out of the way of your own model choice. Credits Saver never changes which model an employee is set to use, that's a separate setting on the employee's profile, it only decides, turn by turn, whether that setting is worth paying for right now.
Credits Saver's own settings page shows the real split for your workspace, how many turns were classified as chat versus work, tracked credits saved, the router cost for those same turns, and the net benefit, pulled from your actual usage. You're not trusting a headline percentage, you're reading your own numbers.
The classification call itself is billed too, at a small internal rate, and shown on the same page rather than hidden. A greeting that never needed the classifier is logged as a zero-credit turn so it still counts toward the percentage, but nothing is charged for it.
If the model you've selected for an employee already costs the same or less than Credits Saver's own fast lane, routing would only add an extra step for zero benefit, so it switches itself off for that employee automatically rather than spend credits deciding something that can't save you anything.
This is checked fresh on every turn, not just once. Move an employee to a pricier model and Credits Saver starts routing again on the very next message; move it back down and the router steps aside again, with nothing for you to toggle by hand.
A quick read on every turn, before the employee answers
When Credits Saver is on for an employee, each incoming message is classified once, before the main reply is generated. An obvious greeting or acknowledgement is recognized instantly and costs nothing to classify. Anything else is compared against what a routine chat turn looks like versus what real work looks like, and the message is sent down whichever lane it matches.
Routine turns go to a fast, low-cost model chosen specifically because it is quick and inexpensive while still capable of handling short replies and small edits well. Everything that reads as real work, an analysis, a document, a task, a nuanced question, goes to the employee's normal model at full quality. If anything goes wrong while classifying a message, the platform always falls back to the full model rather than risk downgrading real work.
Credits Saver also turns itself off automatically for any employee whose selected model is already at or below the cost of the fast lane. In that case, routing would add a step for no savings, so the platform skips it and answers every turn on the employee's own model, and re-checks this every turn so it adjusts the moment you change the employee's model.
An employee that fields a steady stream of quick questions, greetings, and small follow-ups (a support or front-desk style employee, for example) spends a large share of its turns on things that do not need its full model. Credits Saver routes that traffic to a fast model automatically, so the credit cost of the conversation drops without you having to manage it turn by turn.
If you keep an employee on your most capable model because some of its tasks genuinely need deep reasoning, but most days are made up of routine check-ins and short replies, Credits Saver lets you keep that model selected while only paying its full rate on the turns that actually use its depth.
Because the routing reads meaning rather than a fixed list of phrases, a greeting or acknowledgement in another language is still recognized and sent to the fast lane. Teams or customers who message in a mix of languages get the same savings as an English-only conversation, without any extra setup.
The Credits Saver page under Technical settings shows how many turns went to the fast model versus the full model, and what the routing itself has cost. That gives you a direct view of how much of an employee's conversation is routine chatter versus real work, and how much Credits Saver is saving on top of it.
Employees that regularly propose a plan and wait for a quick "yes" or "go ahead" before acting benefit from the approval fix directly: the model cost of the work that follows is judged by what was approved, not by how short the approval message was, so a one-word yes to a big task doesn't get answered by the cheap lane.
No. The classifier defaults to your selected model whenever a turn isn't clearly small talk, so ambiguous cases always keep the more capable model.
Yes, it's a per-employee switch. It ships on by default because most workspaces save credits without noticing any quality difference.
Its own page shows how many turns were classified as chat versus work and the credits that saved, so the number is visible, not estimated.
No. It only decides which model answers a turn. Tools, memory, and guardrails stay identical either way.
It turns itself off automatically once that employee's selected model is already at or below the cost of Credits Saver's fast lane, since routing would no longer save anything. It re-checks this on every turn, so changing the employee's model back up switches it on again without you touching a setting.
Not when it's approving something. A bare approval like "yes" or "go ahead" is matched against what your employee just proposed, so approving real work keeps the full model even though the approval itself was only one word.
Credits Saver is part of Things that control how they behave.
Skip the settings panels. Tell your AI employee what skills to learn, what rules to follow, and what personality to use, all through natural conversation. They can even browse the skill catalog and configure themselves as work evolves.