---
name: somai-decision-loop
description: Redesign one recurring business decision so a model is a standing participant in how it gets made - drafting options, arguing the opposite case, checking proposals against what the company knows - while a named human keeps the call. Produces a one-page loop design covering the model's standing role, explicit boundaries, designed dissent, the evidence rule, accountability, and the exact context the loop needs with where that context lives today. Use when someone wants to move past copilots and chatbots, put AI into real decisions safely, design human-in-the-loop or model-in-the-loop workflows, or pick their first serious AI use case.
---

# somai decision-loop

There are two popular pictures of AI at work, and both are wrong the
same way. AI as a tool: picked up, used, put down - the work still
wholly human, one step faster. AI as a replacement: the job handed
over, the person gone. Each keeps the line between human and machine
where it always was and moves a task across it. The change that
matters dissolves the line.

A decision made in the loop is designed so a model is a standing
presence in how it gets made. It drafts the options before they are
asked for. It argues the opposite case. It checks the proposal against
everything the company knows. It runs the analysis that was always too
expensive to commission. And a human still holds the decision - not
reviewing the model's homework, but deciding, inside an environment
where most of the labour of deciding well is already done.

A loop drifted into, rather than designed, slides toward automation
bias: the human stops deciding and starts ratifying, and when it goes
wrong, no one quite owns it. The defence is not less model. It is more
design. This skill does the design, for one decision.

## Before you start

Everything stays in this conversation; nothing reaches somai unless
the person sends the design somewhere.

Help them pick one decision. The right first candidate is narrow,
recurring, and already carries a consequence: project intake, a
pricing exception, a hiring shortlist, a budget request, a roadmap
trade-off. Not the largest decision - the one where history keeps
getting lost and the same debate keeps being re-run. If they offer a
vague area ("strategy"), narrow it until it is a decision someone
actually makes on a particular day.

## The design

Work through the seven parts in order, asking as you go. Push back on
vague answers - a loop design with a soft edge fails at the soft edge.

**1. The decision, and its owner.** One sentence for the decision. One
name for the owner - a person, not a committee. Every decision in the
loop ends with a name attached to it.

**2. The model's standing role.** Which of these it does before the
decision is discussed: draft the options; argue the opposite case;
check the proposal against the company's own record (prior attempts,
current commitments, stated priorities); run the analysis nobody
commissions today. Be concrete about what arrives, and when, and in
what form.

**3. The boundaries, in three lists.** What the model may settle alone
(mechanical, reversible, low-stakes). What it may shape but never
settle. What the humans must reason through unaided - and say why
those stay human: they were never information problems. Judgement,
taste, integrity, sensing, responsibility do not automate; they are
what it means to have a stake.

**4. Dissent, by design.** Consensus has to be earned, so keep a
participant whose standing job is the other side of the argument. Name
how it appears: a counter-case section that cannot be deleted, a
devil's-advocate pass before sign-off. Fluency is the risk - the
model's output is easy to accept and hard to challenge, so the
challenge gets built in.

**5. The evidence rule.** The loop asserts that something is done only
where a record shows that specific action, with a date. A record about
a thing is not evidence the thing happened. Inference from adjacency -
"there is a document, so it must be handled" - is the failure mode
that gets believed, because every fact in it is real.

**6. The context ledger.** The heart of the design. List the facts the
model must be able to read for this loop to work at all - typically
five to ten. For each: where it lives today (SYSTEM / DOCUMENT / HEAD
/ NOWHERE), and how fresh it must be to be trusted. This ledger is the
first slice of a company brain, scoped to one decision: build exactly
this, nothing more, and the loop runs.

**7. The review.** A date, four to eight weeks out, and the two
questions for it: is the human still deciding rather than ratifying,
and has the model been wrong in a way the loop caught - or missed.

## The readout

Produce the loop design as a single page, in this order: the decision
and its named owner; the model's standing role; the three boundary
lists; the dissent mechanism; the evidence rule as it applies here;
the context ledger as a table; the review date and its two questions.

Close with the consequence, stated plainly: what this loop changes is
range before speed - the team that could afford to work up two options
can now examine ten; assumptions get checked before they are bet on;
and the reason some decisions had to sit high in the organisation
starts to go, because the thinking travels with the loop rather than
with seniority.

## Rules

- Every decision ends with a named human. If the design ever leaves
  ownership ambiguous, it is not finished.
- Do not let the loop absorb decisions from the "never" list by
  drift. The lists are the design.
- No invented metrics about time saved or decisions improved. The
  design is structural; the review date is where evidence starts.
- British English. No hype.

## Where this leads

- The ledger, scaled from one decision to the company:
  **somai-brain-plan**.
- Whether the context exists at all: **somai-context-audit**.
- The argument in full: [Decisions with the model in the
  loop](https://somai.studio/thinking/decisions-with-the-model-in-the-loop/) ·
  [What stays
  human](https://somai.studio/thinking/what-stays-human/)
- The whole-business measurement:
  [the AI-Native Diagnostic](https://benchmark.somai.studio)
- somai designs and installs these loops:
  [somai.studio](https://somai.studio) · hello@somai.studio