By KYN AI Advisory Team — AI implementation specialists, Singapore
A sales agent should gain autonomy one rung at a time: suggest, draft, send after approval, then send within limits. Move a task up only when its error rate on real work is measured and low, and keep money, deletions and anything new-to-the-recipient behind a named person. The ladder is per task, not per agent. A single agent can act freely on logging and cautiously on sending.
What is the autonomy ladder?
It is a way to decide how much an agent does without a person for each task. Several vendors describe autonomy in stages. The shared idea is that trust is earned by task and by evidence. Our version has five rungs.
| Rung | Name | What the agent does | Who acts | |---|---|---|---| | 0 | Observe | Reads and reports | Person does everything | | 1 | Suggest | Proposes the next step | Person decides and acts | | 2 | Draft | Prepares the message or record | Person reviews and sends | | 3 | Approve-to-send | Prepares and queues the action | Person approves with one click | | 4 | Send within limits | Acts alone inside written rules | Person audits after the fact |
Rung 4 never covers everything. It covers specific actions with specific limits.
Which tasks sit on which rung?
Sort each task by two questions. What happens if the agent is wrong? Can it be undone?
| Task | Typical rung | Reason | |---|---|---| | Write to CRM log | 4 | Reversible, internal | | Label a reply | 4, with sampling | Internal, errors found by audit | | Draft first-touch email | 3 | Carries your name outward | | Send first-touch to a new contact | 3 | Outbound, hard to recall | | Send a reminder inside an agreed sequence | 3, maybe 4 later | Lower risk, still outbound | | Quote a price from an approved list | 3 | Money | | Change a discount | 1 | Commercial judgement | | Mark deal won or lost | 3 | Affects forecasts and commission | | Delete or merge a contact | 3 | Irreversible | | Reply to a complaint | 2 | Relationship risk |
Our standing rule: anything outbound, anything about money, anything irreversible waits for a named person. In ladder terms, those tasks stop at rung 3. A discount limit has its own reasoning in why sales agents need hard limits on discounts.
When can a task move up a rung?
Set the test before you start, in writing. A task moves up when all of these hold:
- Enough volume. A sample large enough that you trust the rate. For a task with a binary right answer, review at least 100 cases.
- Error rate under your limit. You set the limit by the cost of an error. A mislabelled newsletter is cheap. A missed unsubscribe is not.
- No costly error in the sample. Count the high-cost mistakes separately. One is often enough to hold the task.
- A working rollback. You can stop the agent in one step and undo or contain what it did.
- A named owner. One person is accountable for the task.
Document the decision with the numbers and the date. If the error rate rises later, the task steps back down automatically.
What does approve-to-send need to be fast?
Approval fails when it is slow, because people then approve without reading. Design the screen for speed and for actual review:
- show the draft, the recipient and the source of the opening line side by side;
- one click to approve, one to edit, one to reject with a reason;
- group similar drafts so a batch is reviewed together;
- show what is unusual, such as a first contact with this company or a number in the text;
- log who approved, what, and when.
A reviewer who sees a draft and its cited source can clear a batch of 30 in minutes. If review time grows, improve the drafts and the screen before you remove the gate. Our guide on setting approval gates for high-risk agent actions covers the design.
What limits belong on rung 4?
When a task reaches rung 4, write the limits as code, not as a prompt:
- Scope. Which contacts, which templates, which channels.
- Volume. A cap per inbox per day, and a cap per contact per week.
- Content. Approved templates only. No prices, dates or terms unless they come from an approved field.
- Stop rules. Pause on any reply, complaint, bounce spike or unsubscribe.
- Kill switch. One control that stops all sending and holds the queue.
- Audit. A weekly sample of sent items checked by a person.
For email, Google asks senders to keep the spam rate in Postmaster Tools below 0.10% (sender guidelines). Treat a breach as an automatic reason to drop the task one rung.
What are the common mistakes?
- Granting autonomy by agent rather than by task. One agent holds many tasks with different risks.
- Skipping rung 2. Teams go from suggest to send and never learn how the drafts read.
- Measuring volume instead of errors. Sent counts say nothing about quality.
- Ignoring hostile input. Text from a lead or a web page can try to redirect the agent. Limits in code hold when the prompt does not. See prompt injection in business AI agents.
- No way back. A rung without a rollback is a one-way door.
What does a worked example look like?
Take an illustrative reply sorter that starts at rung 2. The team reviews its labels for four weeks before moving it.
| Week | Replies sorted | Mislabelled | Costly misses | Decision | |---|---|---|---|---| | 1 | 40 | 6 | 1 unsubscribe missed | Stay at rung 2; add a code rule for stop words | | 2 | 50 | 4 | 0 | Stay; sample again | | 3 | 60 | 3 | 0 | Move to rung 3 for routine labels | | 4 | 60 | 2 | 0 | Move logging to rung 4; labels stay sampled |
These numbers are illustrative. What carries over is the order of decisions: the code rule came first, the costly miss reset the clock, and the move up was written down with its figures and its date.
What does a step-down look like?
Define the trigger before launch. A reasonable set of triggers:
- one missed unsubscribe or complaint in a week;
- an error rate above the limit for two weekly samples in a row;
- a change of model, prompt or template, which resets the task to the rung it held before the change;
- a spam-rate warning from the mailbox provider.
On a trigger, the task drops one rung, the owner is told, and the cause goes into the log. Moving back up uses the same test as the first time. A task that has no step-down rule is not on a ladder. It is on a ledge.
Where KYN fits
KYN builds agents with their own restricted accounts, row-level security in the database and every action logged with the actor. Outbound messages, money, filings and deletions wait for a named person. For an insurance brokerage, sorting replies runs without review at about S$0.01 per reply, while outbound emails at about S$0.02 each wait for approval.
KYN starts with a 14-day review: we inventory your tools and interview the people who send, approve and log sales work, then set the rung for each task. The full build ships in under 90 days. See sales systems and security.
Frequently asked questions
Should a sales agent ever send without approval?
For narrow, low-risk, reversible actions inside written limits, yes. For first contact, money and deletions, no. Decide per task and keep the evidence.
How long should a task stay on each rung?
Until the sample passes your test. That may be two weeks for logging and several months for outbound sending. Time is not the criterion. Measured error is.
What if the error rate rises after a task moves up?
Step the task down a rung at once and investigate. Write that rule before you move it up.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
