Accountability and major decisions
System: Check policy, documents, thresholds and consequences.
Human: Make and own a justified exception decision.
Automated preparation → Human decision → Automated execution
People should not be involved out of habit. They should step in where their participation genuinely adds value or is necessary because of the nature of the task.
A system can collect data, verify documents, calculate figures, retrieve information, prepare decision options and update business software. Yet some decisions require accountability, trust, expert judgement or an understanding of circumstances that cannot be captured reliably by fixed rules.
We do not try to remove people at any cost. Instead, we design workflows so employees participate only at meaningful decision points. The preparatory and follow-up steps remain automated wherever this is safe, technically feasible and authorized.
Standard operations can usually run without routine approval. However, when facts are missing, consequences are significant, interests conflict or something genuinely unusual occurs, an automated action may create more risk than value.
A reliable process must be able to stop, explain why it stopped and send the right person a prepared case file. The employee decides within their authority; the system then resumes the approved workflow.
System: Check policy, documents, thresholds and consequences.
Human: Make and own a justified exception decision.
System: Prepare client history, commercial terms and options.
Human: Build trust and negotiate an acceptable agreement.
System: Gather evidence, surface contradictions.
Human: Interpret circumstances outside standard rules.
System: Pause unsafe steps and create an actionable case.
Human: Choose an authorized, safe continuation.
Choose a scenario. The system first completes everything it can automate, then hands over a specific question, documents and permitted options when needed. You can choose a course of action and see how the automated workflow continues.
A request, enquiry or order.
Collect documents and verify facts.
Are the facts and authority sufficient?
A specific question for the responsible person.
Only the authorized next action.
A customer requests a return for order #1284, 47 days after delivery. The company uses a 30-day internal window for standard returns. The customer reports intermittent equipment faults, but a diagnostic report has not yet been provided.
The system identifies an exception, not an automatic rejection. Internal standard-return rules do not override any other contractual or legal rights that may apply.
| Human choice | System continuation |
|---|---|
| Request diagnostic evidence | Prepare a request and a follow-up task; status: Awaiting Information. |
| Approve an exception within authority | Record the grounds and start the authorized inspection/return workflow, without faking a completed refund. |
| Decline with a reason | Draft the response, record the basis and the applicable escalation path, if any. |
A long-standing B2B customer wants to renew a 12-month contract and requests a 12% discount plus 45-day payment terms. The company authorizes standard offers of up to 5% discount and 14-day payment terms without additional approval.
CRM and calculation modules prepare transaction history, commercial figures and possible options. An AI assistant may draft proposals, but it must not promise terms beyond the employee’s authority.
| Human choice | System continuation |
|---|---|
| Offer standard authorized terms | Draft a proposal; do not assume the customer accepted it. |
| Request approval for an exception | Create an approval task; do not promise unapproved terms. |
| Prepare a different proposal | Generate a draft alternative for the manager to discuss personally. |
An important customer is expecting 80 units of equipment. ERP lists product B-120, but scanning reveals labels reading B-102 on 12 packages. The other 68 units pass the automated comparison. The shipment is scheduled for the following day.
The system cannot conclude whether the issue is incorrect labeling or wrong physical products. Rather than ship potentially incorrect goods, it holds the disputed items and prepares a precise exception report.
| Human choice | System continuation |
|---|---|
| Request a physical inspection | Create inspection task; keep disputed units on hold. |
| Propose a partial shipment | Prepare a customer approval request for the 68 verified units; do not ship without authorization. |
| Pause the shipment | Keep the hold and prepare an incident notification and follow-up task. |
A request, enquiry or order.
Collect documents and verify facts.
Are the facts and authority sufficient?
A specific question for the responsible person.
Only the authorized next action.
The question for the responsible person will appear after the automated checks.
What is missing:
Start the automated preparation. A human decision is needed only at the checkpoint.
Human decisions are explicit, permission-controlled transitions within an otherwise automated workflow.
Standard, well-supported events can proceed automatically. When information is contradictory, an authority limit is reached or material exceptions arise, the process moves into a controlled waiting state. The responsible employee receives the question, evidence, source links and relevant choices. Once the decision is recorded, the system executes only the authorized next step.
Verified facts and sufficient authority → automated action.
Prepared case file → human decision → authorized continuation.
Await information or approval; the unsafe action remains blocked.
Typical components include Python business rules, PostgreSQL, CRM/ERP integrations, APIs, task queues and role-based access controls. LLMs or machine-learning models may assist with extracting facts, retrieving documents and preparing drafts, but they do not replace an authorized human decision where one is required.
| Component | Purpose |
|---|---|
| Python / Business Rules | Validation, calculations, routing and state changes. |
| CRM / ERP / API | Source-of-truth data, documents and permitted business actions. |
| PostgreSQL | Case states, decision reasons and audit history. |
| LLM / Semantic Search | Extract context, retrieve evidence and draft communications. |
| Role-based Access | Verify the decision-maker’s authority. |
| Task Manager / Notifications | Assignments, deadlines, escalation and follow-up. |
Each human decision point needs a designated owner, an explicit reason for escalation, an authorization boundary and an audit trail. A missing response or integration failure must never be treated as implicit approval. When safe continuation is impossible, the process remains on hold or stops cleanly.
Employees should not have to re-enter existing data, repeat automated checks or manually distribute the approved decision. The surrounding workflow remains automated.
Organizations do not have to choose between complete automation and universal manual control. Reliable, repeatable tasks run in software while qualified employees handle the exceptions and judgements where they add real value. This reduces unnecessary approvals, frees expert time and keeps important decisions accountable.
Meaningful measures include straight-through processing rate, justified handoffs to people, time to resolve exceptions, preventable errors and repeated manual work. The objective is not simply fewer human touches, but better process economics without losing quality or responsibility.
First, eliminate unnecessary steps. Next, automate the necessary repeatable operations. Finally, preserve human participation where it truly adds value.
Do your teams spend time on reviews and approvals even though most cases are routine? Share the process and we can identify what to eliminate, what to automate reliably, and where an accountable person is genuinely needed.