Every automation conversation reaches the same moment. The demo works, the drafts look right, and someone asks: so does it just send these itself? Our answer is no, not at first, and for some processes not ever. This article explains the design rule behind that answer, because it is the single most important decision in any AI system that touches your customers.
The asymmetry that decides everything
When a person makes a mistake in an email, one customer gets one wrong message and usually forgives it. When an automated system makes a mistake, it makes it at machine speed, in your name, to everyone matching the pattern, until someone notices. The cost of errors is asymmetric: automation multiplies whatever you feed it, including the failures.
That asymmetry, not technical capability, is why "can the AI do this alone" is the wrong question. Current models can draft a correct reply most of the time. The right question is what happens on the day they do not, and whether that day costs you a customer, a deadline or a professional-liability case.
The four levels
We describe every process on a four-level ladder:
Level 1: the system suggests. It finds, sorts and summarizes; a person does the work with better information. Example: incoming mail triaged by matter and urgency. Nothing leaves the house that a human did not write.
Level 2: the system drafts, a person approves. The reply, the quote, the reminder is fully prepared in your tone; a person reads it and clicks send, or edits first. Approving takes seconds. Writing took minutes. This is where almost every customer-facing process should start, and it is the level we build by default.
Level 3: the system sends, people spot-check. After weeks of level 2, the approval history tells you exactly how often drafts were edited. When a category sits at practically zero corrections, routine cases can go out automatically while samples and edge cases still land on a human desk. The system earned this with data, not with a demo.
Level 4: full automation. Justified when three things hold at once: the action is reversible, the cost of a rare error is trivial, and no rule requires a human. Filing attachments, tagging documents, internal summaries, moving data between your own systems. Plenty of valuable automation lives here precisely because nobody outside your company ever sees it.
The ladder only moves in one direction with evidence. Skipping straight to level 3 because the demo looked good is how automation projects end up as cautionary tales.
Where the human is not optional
Some steps do not climb the ladder, whatever the accuracy statistics say:
- Deadlines in a law or tax practice. Missed deadlines are the most common trigger of professional-liability claims. Our systems extract and prepare deadline entries; a person confirms every single one. The automation makes the checking fast, not optional.
- Anything that is professional advice. Legal and tax assessment of an individual case is reserved to the professional by law. A system may collect facts and prepare drafts for the professional's review; it may not answer the client's legal question.
- Money and commitments. Prices, contract terms, payment decisions. A draft can propose them; a person owns them.
The EU AI Act points in the same direction: it is built around the principle of human oversight, and from August 2026 your chatbot has to tell people it is a machine. A system designed as draft plus sign-off satisfies the spirit of those rules by construction instead of by retrofit.
What this costs you, honestly
Level 2 means a person still spends time approving. For a mailbox with 40 routine replies a day, that is perhaps 20 minutes of reviewing instead of three hours of writing. The saving is real but it is not 100 percent, and anyone who promises you 100 percent on a customer-facing process in week one is selling the part that breaks.
What you get for those 20 minutes: every message that leaves the house was seen by someone who can catch the one case the pattern missed, an approval history that tells you when level 3 is actually safe, and a team that trusts the system because it never went around them.
That is the trade. We think it is the only honest one on the market, which is why every proposal we write names the level each process starts at, and what evidence would move it up the ladder.