Skip to content
Conhecimento
Academy Insights Workshops
TheNeil
Sobre Vagas Contato
Falar com a TheNeil → Every leap begins with a first step
Executive Track · Course 04 of 05

Who answers when AI gets it wrong

The question always arrives after the first visible error, and by then it is too late to design the answer. Accountability is defined beforehand, per operation, not per tool.

Course 04 of 0513 min readSignup requiredEnds in an instrument
13 minute read · By the end you can

Build the inventory of AI-assisted operations in your company with an owner, authority per action and a spend limit, and identify which operations currently have nobody answering for them.

Access to course 04

Sign up once to unlock this course

Course 01 is open. For courses 02 to 05 we ask for six fields. It is a single signup: done once, valid for the whole track. It is not a free diagnosis and it does not trigger automatic sales contact.

The starting point

A tool has no owner, an operation does

Companies try to govern AI through the tool: which model is permitted, which subscription was approved, which team may use it. That is a useful and insufficient layer, because the risk does not sit in the tool, it sits in what the tool does with nobody watching.

The unit of governance is the operation: recurring work, with an expected outcome, that today happens with AI support. Qualifying a lead. Checking a document. Answering first-line support. Each of those has a possible owner, a possible budget and a possible limit.

Inventory

What is not listed cannot be governed

The first step is uncomfortable: most companies do not know how many AI-assisted operations already exist. Areas bought tools, teams built automations, someone connected an agent to a database. None of it passed through a common register.

A minimal inventory has five columns: which operation, which area, what it decides or produces, who answers, and how much it consumed last month. Assembling that table usually reveals two things: operations without an owner, and spend nobody had added up.

Authority

Authority per action, not per system

Saying an agent "has access to the CRM" does not describe authority. Authority is described per action: it may read the history, it may draft the reply, it may send without review up to a certain value, it may not grant a discount, it may not close the case.

Three bands that cover most cases

Free action, for what is reversible and low impact. Action with approval, for what commits a client, a contract or cash. Prohibited action, for what carries accountability that cannot be delegated.

Defining the three bands before deployment avoids the hardest conversation, which is removing authority after an incident.

Budget

An agent without a budget is spend without an owner

People have a forecast cost. Agents, in most companies, still do not. Consumption shows up on a corporate card, with no cost centre, no ceiling and nobody tracking the curve.

Each operation needs a monthly ceiling and an alert before the ceiling. This is not bureaucratic control: it is what lets the operation grow predictably and lets you defend the investment when finance asks.

Evidence

Without a trail, the answer is always "we do not know"

When an error happens, the question is what the operation used as its basis, which action it took and who approved it. If that information is not recorded at execution time, it does not exist afterwards.

Minimum evidence: which source was consulted, which output was produced, who approved when approval applied, and what was changed by human intervention.

The honest limit

Governance does not eliminate error

No design avoids every error, and promising that destroys trust at the first failure. What governance delivers is something else: error within a known band, detected quickly, with a defined owner and a trail to explain what happened.

Exercise

Operations inventory and authority matrix

List the AI-assisted operations that already exist in your company, including the informal ones. Record who answers and the maximum authority granted today.

OperationWho answers todayMaximum authority (free, with approval, none)

Every row where the owner column came out empty or ambiguous is an operation running without an owner. Those are the first to address, before expanding any use of AI.

Basis

Basis

This course treats governance as operational design, not as a policy document. Policy sets principle; an operation needs an owner, an authority band, a spend ceiling and an evidence trail.

Next step

When the inventory becomes an investment decision

If the inventory shows several operations with no owner and no ceiling, the next step is AI Operations Governance, and deployment on Arden.AS when volume justifies it.

Explore Prumo Discovery Course 05 →