Lesson 6.4
Human Accountability and Decision Rights
Why a person, not Claude, is always accountable for a business outcome, how to define which decisions Claude may inform and which it may never make, and how to design oversight that is real rather than ceremonial.
Claude can draft, analyse, recommend and, with connectors and agentic features, act. What it cannot do is be accountable. When a customer is wrongly refused, a report misleads a board or a message goes out with an error, the responsibility lies with the person who used the tool and the organisation that deployed it. This lesson is about making that accountability explicit before anything goes wrong: deciding who owns which decisions, where Claude's involvement stops, and what "a human reviewed it" has to mean to count.
The accountability principle
Using Claude does not transfer responsibility. "The AI wrote it" is not a defence a client, regulator or court will accept, any more than "the spreadsheet calculated it" is. The person who sends, signs, publishes or acts is accountable for the content, and the organisation is accountable for how it deployed the tool. The practical corollary is that anyone who uses Claude for consequential work must be willing and able to stand behind the output as their own, which means reviewing it to the appropriate tier (Domain 2) and understanding it well enough to defend it.
Decision rights: inform, recommend, act
A useful framework classifies each decision by the role Claude is allowed to play.
| Level | Claude may | A person must | Examples |
|---|---|---|---|
| Inform | Gather, summarise, analyse, present options | Decide, with or without Claude's input | Market analysis, risk summary |
| Recommend | Propose a course of action with reasoning | Accept, modify or reject before anything happens | Draft reply, suggested classification |
| Act with checkpoint | Carry out defined steps | Confirm at each consequential step | File a routine form, update a record |
| Act autonomously | Complete the task end to end | Set the rules in advance and audit afterwards | Low-stakes, reversible, well-understood tasks only |
| Never | Nothing | Everything | Decisions reserved by law, policy or ethics |
The "never" row matters. Some decisions must be made by a named human with no AI input, because policy or law says so, or because the organisation has judged it inappropriate: terminating employment, final clinical or legal determinations, certain regulatory sign-offs. Knowing which decisions in your area fall here is part of using Claude responsibly.
Making review meaningful
Oversight fails when it becomes ceremonial: the reviewer skims because the output looks fine and the queue is long. Signs of ceremonial review include approval rates near 100%, review time that shrinks as volume grows, and reviewers who cannot explain a figure in the document they approved. Design against it:
- Make the reviewer do something. Reconcile one figure; answer one question about the content; choose between options rather than approve a single one.
- Surface uncertainty. Instruct Claude to flag assumptions and low-confidence items so the reviewer's attention goes where it is needed.
- Limit batch size. A person can meaningfully review ten drafts; fifty in a sitting invites rubber-stamping.
- Rotate and sample. A second reviewer checks a random sample; the sample grows if errors appear.
- Measure review quality, not just throughput: errors caught, errors missed, time per item.
- Give reviewers authority to stop the workflow when something looks wrong.
Audit trails
Accountability requires being able to show what happened. For consequential workflows, keep:
- the instructions and knowledge in force at the time (Project history helps);
- the inputs Claude received;
- the output as produced and as finally sent, so changes are visible;
- who reviewed and approved, and when;
- any flags Claude raised and how they were resolved.
Enterprise plans provide audit logs and a compliance API for this purpose; on other plans, conversation history plus your own records will do.
Scenario: the grant assessments
A foundation uses Claude to help assess two hundred grant applications. The programme director sets decision rights: Claude may inform (summarise each application against the criteria) and recommend (a provisional score with reasoning); it may never decide. Each assessor reviews twenty applications, must write one sentence per application in their own words about the strongest and weakest point, and may override any score. A second assessor samples ten per cent. Scores Claude flagged as uncertain go to a panel. Results: assessor time halves, override rate is eleven per cent (showing review is real), and the trustees receive a record of every decision with human reasoning attached. When an applicant appeals, the foundation can show exactly who decided and why.
Checklist for accountable use
- Have I classified this decision as inform, recommend, act with checkpoint, autonomous or never?
- Is a named person accountable for the outcome and aware of it?
- Does the review step require the reviewer to engage with the content?
- Are uncertainty and assumptions surfaced for the reviewer?
- Can we reconstruct inputs, outputs, changes and approvals afterwards?
- Do reviewers have the authority and time to stop the process?
Exam focus
- Accountability for outcomes always rests with people and the organisation; "the AI produced it" transfers nothing.
- Classify decisions by Claude's permitted role: inform, recommend, act with checkpoint, act autonomously, or never.
- Some decisions are reserved for named humans by law, policy or ethics and must have no AI input.
- Meaningful review requires engagement (reconcile, answer, choose), surfaced uncertainty, limited batch sizes, sampling and measured review quality.
- Audit trails capture instructions, inputs, outputs, changes, reviewers and flags; Enterprise adds audit logs and a compliance API.
Check yourself
5 questions
1.A manager uses Claude's analysis to decide which vendor to select. Who is accountable for the decision?
2.Claude's analysis of supplier performance recommends terminating a long-standing supplier. The procurement manager issues the termination the same afternoon without checking the underlying data. How should this be judged?
3.A department is writing its guidance for using Claude. What should the section on decision rights contain?
4.An internal auditor asks how a team can show, months later, how an AI-assisted recommendation was produced. Which practice supports this?
5.A Claude-drafted customer email contained an incorrect refund amount and was sent without review. The team's first reaction is to blame the model. What is the more useful response?