An employee, not another app to check.
Small business owners answer email at 10pm and chase unpaid bills on Sunday because nobody else can. The tools on offer either do nothing until you click, or guess and hope. We built Workbench to work the way a careful new hire does: it does the routine work, asks when it isn’t sure, and asks less as it learns.
How we build it, and why.
Decisions come from questions, not from a chatbot.
A chatbot that reads your email and “does what seems right” is hard to trust, because you can’t see why it did anything. Workbench asks a fixed set of plain questions about each email, bill or message and gets back answers with how sure it is. Code turns those answers into an action. You can read every question, every answer and every rule it applied.
Writing and deciding are different jobs.
A language model is good at a first draft and bad at knowing when to stop. So it only writes: replies, summaries, the morning brief. It never chooses whether to send, archive or forward.
You approve anything that goes out.
Sending to a customer, deleting, forwarding: by default, these wait for your yes. The same rule holds when you ask it for something in words. You can loosen it one job at a time, never all at once by accident.
Trust is earned per job.
A new hire doesn’t work alone on day one. Workbench starts by asking, counts how often you agree, and only takes over a job after 15 clean approvals in a row. A single Undo sends that job back to asking.
It should tell you what it did.
Silent automation is how things go wrong without anyone noticing. Every decision is in a log with its reason, and it reports twice a day, in the morning and at the end of the day.
Two models with separate jobs, one install per business, and a growing list of connections.
The decider
Jev, a model from TypeSafe that answers typed questions (yes or no, a place on a scale, one pick from a list) with a probability. It doesn’t write text, which is the point.
The writer
A language model from OpenAI drafts replies and summaries, in the voice and style you give it, using what it knows about the customer.
The install
Each business gets its own copy of the app and its own database. There’s no shared pool of customer data. Account tokens are stored encrypted.
A logo that shows what it’s doing.
The Workbench mark is two pieces: a notch and a box. Inside the app it moves with the work. Envelopes drop into it while it sorts, a light runs its outline while it decides, it nudges you when a question is waiting, and it turns green when a job is done. You can tell what your employee is up to from the corner of the screen.
See it on your own inbox.
On the call we look at how your business runs day to day, decide what it should handle first, and set it up. It asks a lot in the first week, which is how it learns your business.