Building AI Agents Into a Marketing Product: Agent Types, Layers, and Real Build Numbers

Building AI Agents Into a Marketing Product: Agent Types, Layers, and Real Build Numbers

●2 ●7
calendar_today • schedule6 min read

TL;DR

  • An agent differs from a chatbot in one way: it acts. It takes a goal, picks the steps, uses tools, and hands off to a person when it should.
  • For agencies and martech platforms, the agents that matter live inside the product they sell, not in a subscription bolted on next to it.
  • Seven agent "shapes" cover almost everything marketing products need.
  • The model call is the easy part. Integrations, approval queues, escalation routing, and tenant isolation are where the time goes.
  • Don't build if nobody can describe the decision the agent should make, or if the volume won't pay for it.

Agent vs. chatbot, in one example

A chatbot answers a question. An agent works toward a goal.

Take a review that lands on a restaurant's listing. An agent reads it, checks the brand guidelines, drafts a reply, decides whether the review is critical, and sends it to the right manager if it is. That's four decisions and at least two tool calls before anything is published.

The same distinction applies to ChatGPT: on its own, it's a product built on a language model. An agent wraps a model with tools, memory, decision logic, and guardrails so it can finish a task.

Why marketing businesses build instead of subscribe

Agencies and martech platforms already sell marketing or customer engagement. When their clients start asking about AI, there are two options: resell someone else's agent, or ship one as a feature of their own product.

The four businesses in this article chose the second option:

  • Street Digital sells multifamily digital marketing to property owners, with its own "Apartment Intelligence" software.
  • Respage runs a multifamily lead generation and leasing platform for property managers.
  • Localyser offers reputation and guest-experience software for multi-location restaurant chains.
  • Adway runs a social recruitment advertising platform used by 300+ global enterprises.

The reasoning is commercial as much as technical. A subscription agent sits beside your product, runs on the vendor's model and your exported data, and your competitors can buy the same thing tomorrow. An agent built into your product runs on your data in your stack, and it's a feature you can price, improve, and defend.

The flip side: if a single off-the-shelf tool already does the job and you're not reselling the capability, subscribe to it.

The 7 shapes

Generic taxonomies (reactive, deliberative, hybrid) don't help much when you're scoping a marketing product. These seven shapes are the ones that actually show up, each with a live example.

1. Response agent
Drafts on-brand replies to reviews, comments, and inquiries, and escalates what it shouldn't answer. The writing is the easy part. The hard part is recognizing a refund demand, a safety complaint, or a legal threat and sending it to a person instead of a template. Example: Localyser's review auto-reply, running across 1,200 locations, cut average response time from 18 hours to 4 minutes.

2. Qualification agent
Scores and routes inbound leads and answers first-touch questions when nobody is at a desk. Respage found that most prospects reach out after 5 PM with simple questions such as pet policy or tour times. Example: Respage's leasing assistant handles 62% of first-touch inquiries without a human, so the team starts each morning with qualified conversations instead of a backlog.

3. Research agent
Collects and structures context about a prospect or market before a person acts. The salesperson still makes the call. The agent makes sure they're informed. Example: Street Digital's acquisition engine cut manual prospect research by 73% and lifted outbound capacity 3.2x without new sales hires.

4. Creative adaptation agent
Turns one asset into versions for each placement and audience. Example: Adway adapts a single open role or employer-brand story to different social placements, so nobody rebuilds it by hand for every format.

5. Assembly agent
Builds a campaign from a brief, an audience, and a set of assets, then sends it to a person for approval. Example: a BI and automation deployment for a digital marketing agency, built with Anthropic's Claude, where campaign creation dropped from 45 minutes to 10.

6. Insight agent
Turns performance and feedback data into answers you can ask for in plain language. Examples: Localyser uses it to find why ratings dropped at a specific location. In the BI deployment, dashboards return in under three seconds, so people can query campaign data instead of waiting for a report.

7. Orchestration layer
Decides which agent runs, in what order, and when a human must step in. One agent does one job well. The orchestration layer turns several of them into a product. Example: Street Digital's multi-agent system. Designing the routing logic can take more work than the model call itself, which is why AI agent orchestration is usually its own workstream.

Combined, these shapes cover the common use cases: multi-channel inbox triage (response + orchestration), renewal and retention outreach (qualification + response, with drafts reviewed by a person), and cross-channel root-cause analysis (insight).

The parts that aren't the model

A marketing agent has five layers. Most of the difficulty sits outside the AI model:

  • Channel integrations: bring messages in and push replies out, one connection per channel.
  • Agent layer with guardrails: reads context, decides, drafts, and checks the draft against brand and policy rules.
  • Human approval queue: holds anything public-facing until a person approves it.
  • Escalation routing: sends the cases the agent shouldn't handle to the right human.
  • Client data store: keeps each tenant's data separate and under that client's control.

Three things deserve extra attention.

Integrations come first. Localyser pulls Google, Yelp, TripAdvisor, UberEats, DoorDash, Instagram, Facebook, OpenTable, Talabat, Deliveroo, and more into one inbox. Each channel has its own rate limits and approval rules, and each one is a separate piece of work.

The human gate is a design element, not a workaround. Public output needs an approval path, at least at first. That means queues, roles, and an audit trail. It's also the evidence you'll need later to loosen supervision safely.

Multi-tenancy changes the build. A marketing business serves many clients, so one agent has to behave differently for each while keeping their data apart. That's a different project from a single internal agent.

What drives the cost

The model choice rarely moves the estimate. These do:

  • Number of channel integrations: usually the biggest factor
  • Brand voice fidelity: a correct but off-brand reply counts as a defect, and evaluating voice takes real engineering time
  • Multi-tenancy: per-client behavior and data isolation
  • Human-in-the-loop approval: queues, roles, audit trail
  • Escalation routing: often more work than the model call
  • Output ownership: if the agent goes into a product you sell, IP and data ownership need to be agreed in writing before the build

For calibration, from published builds:

  • Localyser: 6 weeks, a team of two and a half, one channel family, one decision
  • Respage: 14 weeks, five people, a multi-tenant assistant connected to property management systems
  • Budget bands: roughly $30K–60K for a 6-week build up to $140K–275K for a 14-week one
  • Returns: Localyser paid back its full cost within 3 months. Street Digital generated more than $320K in new enterprise pipeline in 5 months, an annualized ROI above 3.5x.

How a build usually runs

Engagements like these take 6 to 14 weeks and move through four stages. Each stage needs something specific from the client:

  1. Discovery and feasibility. Break the workflow into tasks, test whether an agent can do them, and agree on one metric. Client provides: a written workflow description, a sample of real inbound data, and the number that should move.
  2. Architecture and data access. Design integrations, tenancy, and the data model. Client provides: access to each channel and system, brand guidelines, decisions on IP and data ownership.
  3. Build with guardrails. Build the agent, brand-voice checks, approval queue, and escalation routing. Client provides: a named reviewer for outputs and an owner for each escalation path.
  4. Validation and production. Test on real cases, run a supervised pilot, then scale. Client provides: pilot accounts, sign-off, and a decision on when the human gate can loosen.

Most weak projects fail at stage one, because nobody can describe the decision the agent is supposed to make.

When not to build

An agent automates a decision you can already describe. It's not a strategy. Skip or postpone the build if:

  • Nobody can write down the rule. The agent can't learn it from nothing.
  • The volume is too low. Localyser paid back because it deflected about 1,100 hours of monthly triage across 1,200 locations. At 20 locations, the same build wouldn't.
  • You plan to publish unreviewed output. Public-facing content needs a human gate, at least initially.
  • One tool already covers it and you're not reselling the capability.
  • The workflow changes every quarter. Stabilize the process first.
  • You can't name the metric. Then you're not ready for scoping.

Wrapping up

For a marketing business, the decision comes down to renting a capability your competitors can rent too, or building one into your own product on your own data. If you go the build route, plan most of the effort for integrations, approvals, routing, and tenancy rather than the model. That's the core of martech software development with AI agents built in.

🔥 Join developers growing publicly
Share your knowledge, build in public, and grow your developer presence with a global community.

More Posts

AAIF Hits 10 Projects. Its CTO on What to Adopt Now and the Security Gaps Left Open

Tom Smithverified - Oct 23

Dozens of email accounts: learning AI coding tools as a student

rahuldk1105 - Oct 27

I’m a Senior Dev and I’ve Forgotten How to Think Without a Prompt

Karol Modelski - Mar 19

AI Agents Don't Have Identities. That's Everyone's Problem.

Tom Smithverified - Mar 13

The Sovereign Vault — A Comprehensive Guide to Protocol-Driven AI

Ken W. Algerverified - Jun 4
chevron_left
202 Points • 9 Badges
Tallinn, Estonia • intercode.com
4Posts
0Comments
1Connections
InterCode is an AI-first development hub for high-growth digital companies

We are a Claude Certifie... Show more

Related Jobs

Commenters (This Week)

3 comments
1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!