Marketing leadership / AI / Decision Debt

Agentic AI in Marketing: What Leaders Must Decide Before They Automate

By Ângelo Valente · 12 November 2026

An agent that can build an audience, draft an offer and trigger a campaign looks like a productivity breakthrough. But what happens when Marketing and Sales disagree on the definition of a qualified lead? Or when the cheapest acquisition brings in customers who never become profitable?

My concern is the unresolved judgement behind the task. Give an agent an ambiguous objective and the ambiguity becomes an execution problem. Give it permission to act and that problem can reach customers before anyone has reviewed it.

I use Decision Debt to describe the accumulated cost of unresolved organisational choices. Agentic AI makes those choices more consequential. The leadership task is to make delegation explicit.

First, decide whether you need an agent

Anthropic distinguishes workflows, where models and tools follow predefined paths, from agents, where the model directs its process and tool use. It recommends starting with the simplest solution that works.1 The useful question is how much judgement a task actually requires.

Fixed rules
Route a lead using agreed country and account criteria.
Assisted judgement
Draft a campaign brief for a human to review.
Bounded agency
Investigate a performance change using approved data and propose the next action.

A predictable routing task may need ordinary automation. An uncertain investigation may benefit from an agent. Neither example implies permission to contact customers or change budgets.

Before automating the work, reconsider the business behind it

Agentic AI should prompt a strategic question before a technical one: are we improving the way we deliver value, or does the value proposition itself need to change? Automating an existing process may be useful. It may also make a weak strategy run faster.

I would separate three leadership decisions. They require different evidence, and progress at one level does not prove that the next is justified.

01 / Process redesign

How can we deliver the same value more effectively?

Example: an agent investigates campaign performance and prepares recommendations for review.
02 / Strategy redesign

Where will we compete, and why will customers choose us?

Example: prioritise customer segments where better diagnosis and service create lasting value, rather than optimising acquisition volume alone.
03 / Business-model redesign

Should we change what we offer, how we deliver it and how we earn revenue?

Hypothesis: move from periodic advisory reports to a subscription for continuous monitoring, with human escalation for consequential decisions.
The evidence gate

Does the proposed model solve a customer problem worth paying for?

Test demand, service quality, delivery cost and contribution margin before committing to scale.

The subscription example is a possibility to test, not a prediction. Faster analysis alone does not establish demand for a new service. Customers might still prefer expert advice at specific moments. A redesigned offer must earn its place through customer evidence and viable economics.

This is where marketing leadership should shape innovation: bringing customer insight into choices about the offer, pricing, distribution and growth, alongside product, finance and operations. The future deserves a place in the strategy discussion without becoming an excuse to abandon what already creates value.

Strategy Before Automation · A leadership canvas

Use these five questions before the four operational decisions below. This is a proposed discussion tool, not a validated assessment.

DecisionQuestion to resolveEvidence to seek
Customer valueWhich unmet need could we address differently, beyond doing our current work faster?Customer interviews and a test of the proposed offer.
Competitive advantageWhy would customers choose us if competitors could access similar AI tools?A specific advantage in expertise, trusted relationships, service or legitimately usable data.
EconomicsShould we change the pricing model, and can we deliver profitably after human review and exceptions?Willingness to pay and contribution margin including the full cost of delivery.
Routes to marketIf customers delegate part of buying to agents, how would our offer be discovered, compared and selected?A scenario test of product information, differentiation and purchasing friction; do not assume adoption.
ExperimentWhat is the smallest test that could disprove our business-model hypothesis?A named owner, a baseline, a time limit and predefined continue-or-stop criteria.

Once that strategic choice is explicit, the next task is to decide what authority an agent should have within it.

Tool 1 · The four decisions

This is my proposed application of Decision Debt to agentic marketing, not a validated industry standard.

01 / Decision rights

Who is accountable? Specify one owner, the actions the agent may take and the decisions requiring approval.

Deliverable: a signed decision-rights matrix.
02 / Definitions

What counts as success? Agree qualified lead, conversion, consent, exclusion and customer value before connecting systems.

Deliverable: a shared metric and data dictionary.
03 / Incentives

What must not be sacrificed? Optimising lead volume alone can reward poor fit. Balance the primary outcome with quality and customer constraints.

Deliverable: an outcome metric with guardrails.
04 / Reversibility

What can be undone? A draft can be discarded. A sent message cannot be unsent. Define suspension, recovery and escalation separately.

Deliverable: a tested stop-and-recovery procedure.

Tool 2 · A decision-rights matrix

Illustrative starting policy for a B2B campaign pilot. Adapt permissions to the organisation and use case.

ActionAgent permissionHuman accountability
Analyse approved aggregate campaign dataExecute within read-only accessAnalytics owner checks evidence and quality
Recommend audience or creativeDraft onlyCampaign owner approves the proposal
Change budget or targetingApproval required; no write access in pilotBudget owner authorises each change
Send messages or export personal dataBlocked in initial pilotAccountable owner completes the applicable reviews before access changes
Change consent or suppression rulesProhibitedDesignated control owner manages these rules

Approval must be enforced through permissions and the execution layer. A sentence in a prompt is not an access control. Log the recommendation, evidence, approval and resulting action so an incident can be reconstructed.

Tool 3 · Is this pilot ready?

Tick only what you can demonstrate. This checklist is a discussion aid, not certification. Answers remain in this page and are not saved or sent.

0 / 8

Tool 4 · The pilot scorecard

Start with one workflow and a baseline. For a campaign-analysis agent, compare the time required to produce a usable recommendation, the proportion accepted by reviewers and the rate of unsupported claims. Keep customer-facing execution under human control initially.

MeasureDefinition to agree before launch
Business outcomeIncremental contribution where measurable; otherwise a clearly labelled leading indicator.
QualityAccepted recommendations ÷ reviewed recommendations. Track corrections separately.
Total costModel and tool costs + integration + review + monitoring + rework.
Customer guardrailsExisting consent and suppression rules; agreed complaint and error limits.
Stop conditionAny unauthorised action or data exposure triggers suspension and investigation. Set other limits before the pilot.

Hours saved are capacity released, not automatically cash saved. If you cannot attribute revenue or margin to the intervention, do not label the result revenue ROI. A pilot can still be valuable if it produces reliable evidence about where agency helps.

A worked scenario

Imagine an agent asked to improve B2B lead generation. The original brief says “increase leads”. A more useful brief says “identify opportunities to increase accepted leads from agreed customer segments, using approved data; propose changes without publishing them”.

The agent now has a defined output, evidence boundary and permission boundary. The campaign owner reviews its recommendations. Sales checks whether accepted leads meet the shared definition. A comparison group, where feasible, helps distinguish the effect of changes from seasonality or demand shifts. This is an illustrative scenario, not an employer case study.

From pilot to delegated authority

Observe
Run against historical cases. Check evidence and failure modes.
Recommend
Work on live inputs with human review. Record rejected recommendations.
Act within limits
Expand permissions only after agreed evidence thresholds and recovery tests are met.

The NIST AI Risk Management Framework provides a broader structure through Govern, Map, Measure and Manage.2 The tools here are a narrower operational contribution: a way for marketing leaders to make delegation discussable before implementation.

The future of marketing will require leaders who can design good decisions into the work. Before asking how many tasks an agent can complete, ask what kind of business that delegation should help you build—and which decisions you are prepared to stand behind when it acts.

Continue with Growing Without Losing Control.

Also read: From Marketing Automation to Marketing Orchestration →

Watch: Growing fast is not the same as growing well →