🟢 Foundation

AI for business: the only decision matrix you need

Stop asking 'should we use AI here?'. Start asking 'is this process Reversible? What's the Blast radius? Auditable?'. Practical matrix to decide what to delegate to AI in 2026. Translated from the PT-BR original.

May 15, 2026 · 10 min · ai-agency-skills

The wrong question

The question “should we use AI here?” produces two equally bad answers: “everything” or “nothing.” When the answer is “everything,” you end up with an autonomous agent approving financial transactions without review. When it’s “nothing,” you sit on the sidelines while a competitor automates what was 80% of the work.

The right question is more surgical: “this specific process, with this specific risk, should it be delegated to AI — and at what level of supervision?”

To answer this, at SkilLab AI we use a 5-dimension matrix. We call it the Agent Trust Stack (there’s also a dedicated framework hub; this article explains how to apply it).

The 5 dimensions

1. Reversibility

Question: if the AI gets it wrong, how costly is it to undo?

High reversibility: email draft, proposal draft, formula suggestion. Error = you discard and redo. Cost: minutes.
Medium reversibility: classifying an invoice into a cost center. Error caught in monthly review = reclassify. Cost: ~30 min + stress.
Low reversibility: financial transfer, external announcement, production change. Error = permanent consequence (lost money, retracted communication, downtime).

Rule: the LOWER the reversibility, the HIGHER the level of human review before action.

2. Blast radius

Question: if the AI gets it wrong, how many people/systems are affected?

Personal blast radius: affects only the requesting user (e.g. personal email draft).
Internal blast radius: affects the team/company (e.g. internal report, scheduling automation).
External blast radius: affects end customers, suppliers, regulators, the public (e.g. invoice sent, social media post, customer support reply).

Rule: external blast radius ALWAYS goes through human review before going out.

3. Auditability

Question: after the AI acts, can we reconstruct WHY it made that decision?

Fully auditable: logs of prompt + response + tool calls + data consulted. Full reconstruction.
Partially auditable: logs of prompt + response, no tool trace. Reconstruction possible but incomplete.
Not auditable: “the AI decided,” with no trail. Reconstruction impossible.

Rule: processes with regulatory obligations (LGPD/GDPR, financial regulators like BACEN/SEC/FCA, healthcare HIPAA, etc.) require full auditability.

4. Cost (of error)

Question: if the AI gets it wrong AND we don’t catch it in review, what’s the financial / reputational cost?

Low cost: < $100 or recoverable in a feedback cycle.
Medium cost: $100 – $10k or affects an important transaction.
High cost: > $10k, affects multiple customers, triggers regulatory action or reputational crisis.

Rule: high cost requires multi-layer review (one human approval isn’t enough).

5. Time sensitivity

Question: if the AI doesn’t act now, what’s the cost of waiting?

Not time-sensitive: can wait 24h for human review.
Sensitive: minutes-to-hours window (live customer support, trade timing).
Critical: seconds window (medical alert, real-time anti-fraud).

Rule: time-critical process + high cost = don’t delegate to AI until proven reliability > 99.5%.

How to apply it — practical matrix

For each candidate process, classify on the 5 dimensions (high/medium/low) and sum.

Total	Meaning	Recommended autonomy level
5–7	Low risk	Autonomous AI ok with log + sampling
8–10	Medium risk	Assistive AI — suggests, human approves
11–13	High risk	Strictly auxiliary AI — human decides with AI input
14–15	Critical risk	Don’t delegate to AI today — automate only pre-processing

Real examples

Case 1: classifying invoices in an accounting firm (uma vertical SaaS contábil brasileira)

Reversibility: high (correct classification at month-end close)
Blast radius: internal (affects the firm’s client, not the chain)
Auditability: full (Harness Stack layer 8)
Cost: low per individual error, medium aggregate
Time sensitivity: not sensitive (monthly routine)

Total: 7. Autonomous AI with confidence gating + log. That’s where we landed.

Case 2: sending end-customer collection emails in a campaign

Reversibility: medium (email sent is gone, but retraction is possible)
Blast radius: external (affects customer, brand, possible consumer protection action)
Auditability: partial (depends on template)
Cost: high (badly-run collection campaign = liability)
Time sensitivity: sensitive (but not critical)

Total: 13. Strictly auxiliary AI — generates draft, human approves every send. Do not run in “fire and forget” mode.

Case 3: automatic quote pricing (legal-vertical SaaS example)

Reversibility: medium (can renegotiate)
Blast radius: external (sent proposal creates expectation)
Auditability: full (system should log)
Cost: medium-high (mispricing = lost margin or lost customer)
Time sensitivity: sensitive (customer expects fast response)

Total: 12. Assistive AI — proposes, human approves within minutes. As reliability proves out, can evolve to autonomous on simple cases.

Frequent anti-patterns

“Let’s automate EVERYTHING.” Total = 5 on some processes, 15 on others. Treating uniformly creates an incident.
“Since we have GDPR/LGPD, we don’t use AI.” GDPR doesn’t prohibit AI; it requires auditability. Adjust the level, don’t cancel the program.
“The model is strong, so it can act autonomously.” Strong model + low reversibility + high blast radius = still requires a human. A strong model doesn’t compensate for missing governance.
“There’s already a human review, so it can run.” Review is a bottleneck. If the human doesn’t have time to review with quality, the process stays risky even with the formal step.

Next steps

Apply the matrix to the 5 most frequent processes in your team today. Sum the totals. Processes with 5–10 are where to start.
SkilLab Workshop — Corporate Consulting & Training. We apply the matrix with your leadership in a 4h session and leave with prioritization. Details.
SkilLab AI Newsletter. Sign up below.