Tesseract SA

“AI prepares, a person approves”: where to place the approval

Approving 100% of AI outputs rebuilds the very bottleneck you wanted gone. Here is where to place human approval, with a risk-tier grid and the cases where to skip it.

Matte 3D illustration: white cubes travel along a grey rail and pass through a single red arch, the point where a person approves before the next step.

Where should human approval sit in an automated process?

Short answer: right before the first action that cannot be undone or that leaves the company (a message to a client, a payment, a final accounting entry), and nowhere else by default. The AI prepares, a person approves the small number of decisions that commit you. Anything reversible, internal and repetitive runs on its own, with a sampled check.

Saying “the AI prepares, a person approves” does not design a process. The real question is where exactly the person steps in, on what, and what they see at that moment. Placed badly, approval costs time and protects nothing. This article gives a placement rule, a grid by risk level, and the cases where you are better off with no approval step at all.

Why approving 100% of outputs is a bad idea

The cautious reflex is to have every result reviewed. It is the surest way to rebuild the bottleneck automation was meant to remove: the task is no longer done by hand, but it still waits in a queue for a person. Tensoria, an automation provider, describes the same effect in audited small businesses: approval nodes ticked within seconds, without opening the attachment.

An approval ticked without reading is worse than none, because it suggests a control that does not exist. Flowable, a process tooling vendor, points to automation bias: the more reliable a system proves, the less people check it, and the rare errors slip through more easily. The practical conclusion is counter-intuitive: fewer checkpoints, but better chosen and actually read.

The placement rule: before the first irreversible action

Several specialist sources agree on the same principle: the approval gate goes immediately before the first action that enters critical territory, not at the end of the chain. A final review comes too late. By then the AI has read, calculated, drafted and sometimes sent; the person approves something already done, with little context to judge the steps that produced it.

A concrete example for an unpaid-invoice reminder: the AI spots the late invoice, pulls the client history, computes the balance and drafts the message. None of that needs approval. The checkpoint goes just before sending, because an email that has left cannot be recalled. If the balance is wrong, the person sees it before, not after.

Three questions are enough to place a step:

  • Is it reversible? A draft, an editable accounting category or a CRM field can be corrected. A bank transfer, an external message or a deletion cannot.
  • Does it leave the company? Anything a client, supplier or authority can read puts your reputation at stake.
  • Is there a legal commitment or a significant amount? A contract, a quote or a closing entry is not handled like sorting messages.

The grid: which level of control for which step

The table follows a three-tier logic found in several authors’ work and adapts it to the small-business processes we build. The examples are reference points, not a rule for every company.

TierCharacteristicsSmall-business examplesRecommended control
SafeInternal, reversible, no commitmentSummarise a call, sort an inbox, suggest an expense categoryNo approval, an occasional sampled check
SensitiveReversible, but visible or sometimes ambiguousUpdating a CRM record, bank reconciliation, extracting an invoiceApproval only below a confidence threshold or on a detected discrepancy
CriticalIrreversible, external or committingSending a quote, a payment reminder, a year-end close, anything touching sensitive client dataMandatory approval before execution

The confidence threshold: a setting, not a truth

For the sensitive tier, the AI acts alone when it is sure and triggers approval when it doubts. The threshold is set by the cost of an error at that step. Tensoria cites values often between 70 and 85%. Treat that figure as a starting point to test, not a standard: a threshold that suits sorting invoices does not suit sending a payment reminder to a large client.

The sample: control for the safe steps

Even when everything goes well, a person reviews a sample of routine outputs. The aim is not to fix every error but to spot drift: a supplier changing format, a new kind of client that does not fit the expected cases. Tighten the sample at launch, relax it once reliability has been measured.

Block before, or allow undo after?

Two models exist. In the first (“human in the loop”), the process stops until a person decides. In the second (“human on the loop”), the system acts and a person supervises, able to step in. The first suits irreversible or high-stakes actions. The second suits frequent, low-risk actions that can be undone.

Our rule: block before for anything leaving the company, allow an undo window for anything that stays internal. A single process often mixes both, step by step. For quotes and contracts, this split is detailed in our article on safeguards for AI-generated quotes and contracts.

What the person must see to decide quickly

An effective approval takes a few seconds if the screen is well designed. Three elements are enough:

  1. The source: the document, message or data row the conclusion comes from, one click away.
  2. The precise discrepancy: what triggered the request (an amount that differs from the quote, a client missing from the CRM), rather than a raw score.
  3. The possible decision: approve, correct or reject, with no extra step.

Also log who decided, when and why. That trail helps in a dispute and lets you adjust thresholds from facts.

When not to add an approval step

A needless approval is a cost. Skip it when the step is internal, reversible and an error can be fixed in seconds, or when the person asked to approve has neither the time nor the context to judge. In that second case, adding approval does not protect anything: it shifts responsibility onto someone who cannot carry it. Redesign the step instead, for example by reducing the AI’s autonomy, rather than adding a checkbox.

Conversely, do not automate the decision itself when it calls for human judgement: granting an exceptional discount, answering an unhappy client, ruling on an edge case. The AI can prepare the file, not decide for you.

The nature of human oversight depends on the applicable text. The GDPR gives a right to request human intervention for certain automated decisions (Article 22). The EU AI Act requires human oversight for systems classed as high risk. Most small-business automations (reminders, document extraction, internal sorting) fall outside that category, and operational common sense then guides the design. In Switzerland, the revised FADP sets its own rules on personal data: see our article on the FADP and data sent to a model. If in doubt about your situation, have a lawyer check it.

Where to start

Take one process, list its steps, place each one in the grid, then keep approval only on the critical tier. If you are unsure which process to tackle, the express diagnostic in five questions helps you choose. You can also browse the processes the studio automates to see where the approval gate sits in each.

Frequently asked questions

Where should human approval sit in an automated process?

Right before the first irreversible or external action: a message to a client, a payment, a final entry. Internal, reversible steps run on their own, with a sampled check.

Should 100% of AI outputs be approved?

No. It rebuilds the bottleneck and leads to approvals ticked without reading. Keep approval for high-stakes cases, cases where the AI doubts, and a control sample.

How do you set a confidence threshold?

By the cost of an error at that step. Sources often cite 70 to 85% as a starting point. Test it on your own data and tighten it for sensitive steps.

Approve before, or allow undo after?

Approve before for anything leaving the company or that cannot be undone. Allow undo after for internal, frequent, reversible actions.

Is human oversight a legal requirement?

It depends. The AI Act requires human oversight for high-risk systems and the GDPR gives a right to human intervention for some automated decisions. For most small-business automations, it is mostly a matter of operational common sense.

Sources
Read next