Most of a business process runs on rules. If the total is over $500, ask a manager. Every Monday at 7:00, send the report. Rules are easy to automate because a computer can work them out from the numbers in front of it, the same way every time.
But some decisions don't fit a rule. Is this email a sales lead, or a newsletter? Does this customer sound upset? Is this receipt within our expense policy? A person answers questions like these in a second. A formula can't answer them at all, which is usually where automation stops and someone's inbox begins.
Lodol workflows can now ask those questions directly. Set a condition to Use judgment and finish the sentence "This is true when…" in plain words, like "the sender wants to buy something." Lodol decides it by reading rather than by computing. Only that one decision involves AI. Every other step in the workflow still runs exactly the same way on every run.
"Not sure" is an answer
This is the part we care about most. Every judgment comes back with how likely the answer is to be yes, and you decide where the line goes. Turn on Let this condition answer Not sure and you get two lines instead of one, starting at yes above 70 percent and no below 30, which you can move. Anything in between takes its own path: send it to a person, label it for review, or post it in Slack and ask someone to look.
Most AI features have to pretend every answer is certain, so a borderline call gets made with the same confidence as an obvious one, and you find out which was which when something goes wrong. Here, "I can't tell" is a place your workflow can go, and it's the place you'd want a careful colleague to go too.
What it looks like at work
A shared inbox that sorts itself. Leads, vendor invoices and newsletters all land in the same place. A workflow asks of each new email, "is the sender asking to buy something?" Yes, and the email goes to #sales in Slack with a link to reply. No, and it's left alone. Not sure, and it's labeled for someone to check.
Replies to your outreach. Each reply to a sales email gets one question: "do they want to talk?" Interested people get your booking link. People who said no come off the list, so nobody gets a follow-up they asked not to receive. And the replies in between, the "maybe next quarter" and the "who are you, again?", come to you.
Expenses against your own policy. Write the policy in plain English and hold each receipt against it. Clearly within policy, and it's approved. Clearly not, and it's declined with the reason. Anything unclear goes to finance with the receipt attached, rather than being waved through.
A list, filtered by meaning. A judged condition can also filter a list, one item at a time: keep only the reviews that mention a damaged delivery, or the support tickets that are really feature requests. Each item is read on its own, up to a thousand in one step.
What it can and can't do
A judged condition answers one question with one kind of answer: yes, and how sure it is. It doesn't write messages, call other apps or decide what the workflow does next. Your workflow does that, with the paths you drew. That narrowness is deliberate. The judgment sits at one point in a process that otherwise runs the same way every time, which is the design we think AI belongs in. (Our CEO wrote about why in Agents Are Having a Moment. Reliability Is Having a Reckoning.)
It only sees what the workflow has in hand at that moment: the run's own values, like the email it just read, and the settings your workspace keeps for conditions, like a spending limit. It never sees your secrets, and if a secret's value has found its way into an ordinary value, it's blanked out before the question is asked.
Every run records the answer and how likely it was, such as yes at 87 percent. A branch taken at 51 percent and one taken at 99 percent go the same way, but they're very different facts, and when you can see which one you got, you can move the lines if they're in the wrong place. And if a question can't be answered at all, the step fails rather than guessing, so a Try / Catch step can decide what happens next, like any other failure.
What it costs
Nothing extra. A judgment call is part of the run it happens in, and there's no AI account or key to set up for it, because it runs on ours. That's different from AI steps that write something, like drafting a reply, which run on your own Anthropic, OpenAI or Gemini account. A judged condition only ever answers yes or no, so it's cheap enough for us to include.
Try it on your busiest inbox.
In the editor, switch the condition on any If / Else, Repeat while or Filter list step to Use judgment and write when it's true. Or start from one of our ready-made workflows and add a judgment call where it needs one. Start free if you don't have a workspace yet.
