SellerbridgeSB-01 / NIGHT SHIFT
Decision Audit
[ back to dispatches ]
19 Aug 2026Guardrails desk4 min read

Watch, Recommend, Act: how much rope to give it

The question that decides whether any of this is usable is not how good the decisions are. It is what happens to a decision you have not read yet. That is a permissions question, and it should be yours to answer per account, per agent, and per dollar. The guardrails page lets you switch between the three modes and watch the same night print differently under each.

Watch

Agents observe and report. Nothing is ever changed. Use it for the first two weeks — not because the software needs a probation period, but because you need to see a fortnight of decisions before you can judge which ones you would have wanted applied automatically. One such fortnight, printed out, is a reasonable thing to read first.

03:42
Watch mode
Margin flagged a negative-contribution ASIN. No action taken.

Recommend

Every decision arrives staged and costed, with the reasoning chain attached. One click applies it. This is the default on a new account, and it is where most agencies stay, because the bottleneck was never the clicking — it was knowing which thing to click.

05:05
Recommend mode
Spend staged a bid cap at $1.42. Awaiting approval.

Act

Inside caps you set. A per-target step limit, a per-account exposure cap, a per-agent scope. Everything is logged and reversible, and anything above your threshold escalates to a person instead of executing.

The useful pattern is not picking one mode for everything. It is Act for the reversible, bounded, boring decisions — a bid cap inside a $0.50 step limit — and Recommend for anything a client would notice. Client-facing sends never leave Recommend, in any configuration. Which agent owns which decision is set out on the roster, and what the mode is worth in analyst hours is on the agency math page.

The part that matters when it is wrong

It will be wrong sometimes. What you need then is not an apology but the reasoning chain: what was observed, what was checked, what was ruled out, and which guardrail applied. That is attached to every decision in all three modes, which is what makes a wrong decision a correctable one rather than a reason to switch the whole thing off. Every decision in a Decision Audit arrives in Watch mode, with that chain attached, before you have configured anything at all.

Read one of these on your accountStart a Decision Audit
What this piece is about

Read next

The other two dispatches, picked for subject rather than for date.

Connect one account, read-only. See seven days of what the agents would have caught.

That is the Decision Audit. No configuration, no write access, no call required to start.

Manifest — what lands in your inbox
01Every decision the seven agents would have surfaced, timestamped.
02The money attached to the ones you missed.
03The reasoning chain for each, so you can check our work.
04A read-only audit log of everything we touched.