規則與目標：市場營銷自動化轉為智能體後，程式碼有何變化

TL;DR

• Marketing automation runs rules a human wrote in advance. An agent gets a goal plus guardrails and picks its own path through a perceive → reason → act → learn loop.
• In production, the model is rarely what breaks. Bad upstream data, mismatches between agents, brand-voice drift, and a team that doesn't trust the tool do.
• Guardrails, approval queues and audit logs aren't policy documents. They're features you have to build and test like any other.

The problem: decision trees don't scale to goals

Classic marketing automation is a decision tree someone drew before launch:

Prospect views the pricing page → wait 24 hours → send email variant A → if opened, set lead score X.

Every branch is a path a human anticipated and hard-coded. That works until the number of channels, audiences and creative variants outgrows what anyone can map by hand.

An agentic system flips the contract. You don't write the branches. You define an objective and hard boundaries (budget, compliance, brand safety), and the system decides which segment to target, which creative to assemble, how to allocate bids and when to change pacing as live signals move.

For engineers, that means the core of the system stops being a rules engine and becomes a loop plus a guardrail layer.

The loop

Most goal-driven marketing agents reduce to four stages:

1. Perceive. Continuously ingest behavioral signals, purchase history, inventory or catalog state, channel metrics and consent status. This is a live stream, not a nightly batch.
2. Reason. Evaluate those signals against the goal and the active guardrails, and decide which intervention has the highest expected value.
3. Act. Execute it in the right channel: launch a campaign shell, shift a bid allocation, send a personalized reply.
4. Learn. Capture the conversion or engagement feedback and update parameters before the next cycle.

The interesting engineering sits between reason and act: everything that decides whether the agent is allowed to do what it thinks is best.

Guardrails are features, not a slide

"We'll keep a human in the loop" sounds like a management decision. In practice it's a set of things engineers have to build. Any agent that takes public-facing actions or commits client media budget needs:

• Staging queues where proposed actions wait for review
• Role-based permissions that define who can approve what
• One-click overrides for when a human disagrees
• An immutable audit trail of every decision and action

If you're building for an agency or a platform that serves many clients, this gets bigger. A single brand's internal agent works with one brand voice, one catalog and one privacy perimeter. A platform serving dozens or hundreds of competing clients needs tone, bidding guardrails and conversion goals that adapt per client, with each client's data kept apart. That's a materially larger build, and it's why this kind of work sits closer to AdTech platform development than to a weekend prototype.

Where it actually breaks

Across production deployments and industry post-mortems, four failure modes explain most stalled projects. Notably, "the LLM misunderstood the task" isn't one of them.

1. Upstream data quietly lies

Agents make confident, mathematically sound and commercially wrong decisions when their inputs are broken. If an ad platform's conversion pixel double-counts purchases, a goal-driven agent will keep shifting budget into that broken channel until a guardrail stops it.

Data access is also what drives the estimate. In our builds, the number of external integrations consistently set the engineering effort. One reputation agent had to aggregate reviews and messages from 10 sources (Google, Yelp, TripAdvisor, UberEats, DoorDash, Instagram, Facebook, OpenTable, Talabat and Deliveroo) before it could draft a single reply. If campaign telemetry, CRM records and catalogs live in unlinked silos without stable APIs, the project starts with data plumbing.

2. Handoffs fail before models do

In a multi-agent workflow, say a research agent feeding a scoring agent feeding an outreach generator, failures rarely come from the model's comprehension. They happen in the handoffs: state management and schema mismatches between agents. Coordination, not individual intelligence, is the main engineering challenge.

3. Brand-voice drift is its own defect class

In marketing, a reply or ad variation that's factually right but slightly off-brand is treated by the client as a critical bug. Catching tone violations across 1,200 hospitality locations isn't something you solve by tweaking a prompt. It takes automated evaluation that scores every generated output against a client-specific brand rubric before release.

4. The team routes around the tool

This one isn't technical, but it kills projects. Staff who worry about being automated away, and who weren't involved in rollout, either bypass the agent or magnify its early mistakes to justify shutting it down. A gradual, phased rollout with hands-on training isn't optional.

Rollout: shadow mode first

The release pattern that holds up in production:

• Shadow mode. The agent runs silently against live traffic and publishes nothing. QA scores its output, for example on brand-tone accuracy.
• Parallel runs. For research-style agents, compare the agent's output with what senior analysts produce on the same inputs.
• Phased rollout. Enable live actions for a first cohort of accounts or locations, monitor, then expand.

Then assign a named AgentOps owner on the client side: one person who reviews overnight execution logs, audits edge-case escalations and refines guardrail parameters. An unmonitored agent will drift as ad platform rules and client offerings change.

What this looked like in real builds

Numbers from projects delivered for marketing businesses:

• Outdoor advertising booking (Billboards America). A multi-agent chatbot that automates client consultations and campaign setup. Campaign creation went from 45 to 10 minutes, lead conversion from 15–20% to 35–45%, and consultation automation from 0% to 70%.
• Multi-agent customer acquisition (Street Digital). 10 weeks, about 4 FTEs. Manual prospect research dropped 73%, and outbound capacity grew 3.2x with zero new hires.
• Multi-location reputation agent (Localyser). 6 weeks, 2.5 FTEs, 1,200 locations. Response time fell from 18 hours to 4 minutes.
• Multi-tenant leasing agent (Respage). 14 weeks, 5 FTEs. It handles 62% of first-touch inquiries without a human.

The pattern holds: integration count and architecture, not model choice, set the timeline. Production builds took 6–14 weeks with 3–5 specialists.

When not to build one

If a single off-the-shelf tool solves an internal task and you aren't putting the capability into something you sell, buy it. Build when the agent becomes part of your marketing product and you have enough repeatable volume to pay it back.

Before any of that, make sure someone can write down the rules the agent is supposed to follow:

• What counts as a qualified prospect
• What cost-per-lead variance triggers a budget reallocation
• Which conditions require immediate human escalation

If those rules only exist as a senior media buyer's intuition, the agent has nothing to learn from.

Wrapping up

The model is the easy part of a campaign agent. The work lives in the guardrail layer, the handoffs between agents, output evaluation and the rollout. Treat those as first-class engineering and the agent behaves. Skip them and it fails confidently.

The business side (buy vs. build, the economics for agencies, and a week-by-week delivery plan) is covered in our full write-up on agentic advertising.