# Build or Buy a Marketing AI Agent? Compare HubSpot, Agentforce and Custom Workflows

Canonical URL: https://www.martechdevs.com/blog/build-vs-buy-martech-ai-agent

By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo)

Data Engineer and Co-founder, martechdevs

Published: 2025-10-22

Last updated: 2026-09-08

A marketing AI agent is easy to justify in a demo. The harder decision is whether your team should own the software behind it, configure a platform you already pay for, or combine the two. Make that choice around a specific workflow, its operating cost and the person who will keep it working.

Start with the finished job. “Prepare next week's account-expansion review using CRM, billing and product evidence” is a useful scope. “An agent that runs our entire stack” leaves too much undefined to compare quotes or judge success.

For many teams, configuring an existing platform is the sensible first test. Custom work becomes attractive when you can name a requirement the available options fail, put a value on meeting it and fund the responsibility that comes with ownership.

## First decide whether this job needs an agent

A workflow with known inputs, fixed rules and predictable outputs may need ordinary automation. A weekly query and a short generated explanation can be enough for a report. An agent adds value when the system needs to choose what to investigate or which approved tool to use as new evidence arrives.

Anthropic's [guide to building effective agents](https://www.anthropic.com/engineering/building-effective-agents) recommends starting with the simplest suitable design and adding complexity only when it improves outcomes. It also distinguishes predefined workflows from agents that direct their own steps.

Apply that distinction to your actual bottleneck. If account IDs do not match across systems, an agent does not remove the identity work. If a report's logic is settled, letting a model reinvent the calculation on every run creates another thing to check.

Write down which step needs interpretation. Keep stable calculations, eligibility rules and repeatable updates in components that can be tested directly. The [MCP-versus-API guide](https://www.martechdevs.com/blog/mcp-vs-api) explains how to separate conversational access from dependable execution.

## Compare four ownership choices

### Configure a native agent

Use the existing platform's builder, actions, permissions and interface. Your team owns the configuration and the business process; the vendor operates the underlying product.

This is attractive when users already work there, the necessary data is available and the supported actions fit the job. It can reduce the amount of application infrastructure you must maintain.

You still need someone to manage instructions, field changes, evaluation cases, permissions and exceptions. Buying a platform does not give it your company's definition of an eligible expansion account.

### Extend a native agent

Keep the platform interface and add a narrow tool or data service for the part it cannot do well. For example, an agent could retrieve a governed account-health record from a warehouse and leave task creation inside the CRM.

This can be a good fit when the missing piece is specific. It also creates an integration boundary to own: authentication, tool inputs, freshness, errors and version changes must be maintained.

### Build a custom workflow

Choose your own runtime, context, tools and review interface. Custom can mean a small set of Claude Code skills and approved connectors, a scheduled service, or an application built for a particular team. Those are different deliverables with different operating requirements.

The advantage is control over how the work is assembled and exposed. The obligation is to support it. A repository alone is not a functioning service, and possession of source code does not mean your team can safely change it.

### Keep a simple automation or manual process

Low frequency, unstable requirements or weak input data can make an agent a poor investment today. A documented monthly review may be cheaper to operate and easier to verify.

Defer deliberately: record the volume, error rate or business requirement that would justify reopening the decision. That prevents the same purchasing discussion from restarting every time a new agent product launches.

## What HubSpot Agent Hub and Agent Builder actually change

HubSpot describes [Agent Hub](https://www.hubspot.com/products/artificial-intelligence) as the place to run and manage its agents, including custom agents. The current [Agent Builder product page](https://www.hubspot.com/products/artificial-intelligence/ai-agents) identifies it as the successor name to Breeze studio and includes it in Starter, Professional and Enterprise editions, with custom-agent usage consuming credits. Check the feature and subscription combination you need rather than treating builder access as unlimited use of every HubSpot feature.

The current [Agent Builder setup documentation](https://knowledge.hubspot.com/ai/create-and-customize-agents-in-the-agent-builder) allows instructions, knowledge, inputs and selected actions, including supported external MCP connections. It separates CRM read and write actions and provides agent access settings. Its built-in test process does not consume HubSpot Credits; subsequent run estimates are informational and actual consumption can differ.

That makes HubSpot a candidate for more than a fixed catalogue of canned agents. Test the required external systems, custom properties and outputs before deciding it cannot handle your workflow.

A useful trial is one expansion-review account with a billing discrepancy. Can the configured agent find the right account, show the disagreement and leave the decision unresolved? Can the person who owns the process inspect why it did so? Those answers are more useful than counting the agents in a marketplace.

## What Agentforce changes

Agentforce is also extensible. Salesforce's [MCP connection documentation](https://help.salesforce.com/s/articleView?id=ai.agent_mcp_connect.htm&language=en_US&type=5) describes registering external servers and allowlisting individual tools as agent actions. Connection and management paths differ for third-party, MuleSoft and Salesforce-hosted servers. Required add-ons depend on the agent type, so verify the actual entitlement and connection route.

This matters if Salesforce is already the team's operating environment. A required warehouse lookup does not automatically mean the entire agent must live outside Salesforce.

Ask the implementation team to demonstrate the complete path: the user's identity, allowed tool, record selection, returned evidence and resulting action. Then repeat it with a user who lacks access to the account and with an unavailable external source.

If the design meets the job inside the existing platform, compare its recurring cost and maintainability with an external implementation. Choose based on observed fit rather than assuming either native or custom is inherently more capable.

## Use one acceptance brief for every option

Give vendors and internal engineers the same brief. Specify the audience, input systems, record grain, timing, output and forbidden side effects. Separate requirements that must pass from preferences that can be traded against cost.

Here is an original example for a weekly expansion review. It is a planning document, not configuration that can be imported into a vendor product:

```yaml
workflow: weekly-expansion-review
business_owner: revenue-operations
record_grain: one-account
inputs:
  - crm_account_owner
  - billing_subscription_state
  - workspace_usage_summary
  - unresolved_support_issues
output:
  - account_id
  - evidence_with_source_timestamps
  - missing_or_conflicting_information
  - proposed_next_step
execution_mode: prepare-review-only
must_pass:
  - use_the_approved_account_identity_map
  - distinguish_unknown_from_ineligible
  - retain_the_business_definition_version
  - respect_the_requesting_users_access
  - produce_no_customer_messages_or_record_changes
operating_owner: named_before_launch
```

Use several representative cases: a healthy account, conflicting billing state, a merged account, an unavailable support source and a user without access. Define the expected behaviour before the demonstration. Do not allow the team presenting the tool to quietly substitute cleaner data.

Measure whether the output is correct, complete enough for the decision and quick to review. Keep a record of unacceptable answers as well as successful ones. The [AI-for-RevOps guide](https://www.martechdevs.com/blog/ai-for-revops) provides a fuller pilot method, including missed issues.

An option that fails a mandatory access or output requirement should not win through a high average score. First establish that the workflow is viable. Then compare cost and convenience among the viable options.

![Decision path compares simple automation, native agents, narrow extensions and custom workflows against common acceptance tests and operating costs.](https://www.martechdevs.com/assets/blog/build-vs-buy-martech-ai-agent.svg)

## Compare the billable unit before comparing prices

One request can generate several model calls, tool actions or data queries. “Cost per conversation” and “cost per action” are not interchangeable. Neither tells you how many outputs a reviewer will accept.

HubSpot's [credit-management documentation](https://knowledge.hubspot.com/account-management/understand-hubspot-credits-and-billing) distinguishes included credits, purchased capacity, automatic upgrades and pay-as-you-go overages. Unused monthly credits expire. After buying additional capacity, exceeding it can trigger an upgrade for the remaining commitment term unless the appropriate overage setting is used. Account and feature limits also need attention; a feature limit does not reserve credits for that feature.

For Agentforce, the [public pricing page](https://www.salesforce.com/agentforce/pricing/) lists Flex Credits at $500 per 100,000 credits and standard Agentforce actions at 20 credits each. Other pricing models and action types exist. As an illustrative calculation, 250 runs using eight such metered actions each consume 40,000 credits, equivalent to $200 at that list rate. This is consumed usage value, not a complete quote or necessarily the amount invoiced. Included allowances, commitments, licenses, Data 360 and other services can change the bill.

For a custom Claude Code workflow, [Anthropic's cost documentation](https://code.claude.com/docs/en/costs) distinguishes API consumption from subscription pricing and describes costs varying with model choice and usage. Measure the actual workflow, including repeated attempts and lengthy context. A subscription or an inexpensive model does not make warehouse queries, third-party services or operator time free.

Ask each supplier to map one completed run to its billable events. Include a normal run, an unsuccessful run and a retry. This gives finance something it can forecast and gives engineering something it can optimise.

## A worked first-year cost comparison

The following numbers are fictional planning assumptions for two options that both meet the same acceptance requirements. They are not HubSpot, Salesforce or martechdevs quotes. Existing costs that remain identical under both options are excluded. Incremental subscription or infrastructure requirements belong in the estimate.

Assume the configured-platform option has a $2,000 external setup fee, $400 of additional monthly platform and usage charges, and six internal hours a month for review and maintenance. Value internal capacity at an assumed $75 an hour.

Its first-year cost is $2,000 + 12 × $400 + 12 × 6 × $75 = **$12,200**. Of that, $6,800 is external spending and $5,400 represents internal capacity.

Assume a custom option has a $9,000 external build fee, $150 of additional monthly usage and hosting, and six internal hours a month: four for maintenance and two for review.

Its first-year cost is $9,000 + 12 × $150 + 12 × 6 × $75 = **$16,200**. External spending is $10,800 and internal capacity is again $5,400.

The custom option's recurring monthly cost is $600 versus $850 for the configured platform. Its $7,000 higher initial fee takes **28 months** to recover through that $250 monthly difference, assuming the costs, scope and performance stay constant. This simple comparison excludes discounting, tax and later rebuilds.

Now increase custom maintenance from four to eight hours a month. The extra $300 of monthly capacity cost raises recurring cost to $900 and first-year cost to **$19,800**. The assumed recurring saving disappears.

That sensitivity is the point. A small model bill cannot settle the purchase decision when maintenance is uncertain. Estimate a low, expected and high workload, identify the assumptions that change the winner and test those assumptions during the pilot.

Do not report recovered employee hours as cash savings unless spending actually falls. The time may instead be used for more valuable work. Name that work when presenting the business case.

## Who owns the system after launch?

Assign three responsibilities explicitly. The business owner maintains definitions and decides what counts as a useful result. The system owner maintains integrations, credentials, deployments and recovery. The reviewer resolves cases the workflow cannot handle confidently or correctly.

One person may cover more than one responsibility in a small team. The responsibilities still exist.

For a purchased platform, ask which configuration changes your admins can make, what requires the vendor and how releases affect existing agents. For custom work, ask for the repository, deployment instructions, dependency list, tests, operating costs and the account ownership needed to run it.

Use a handover exercise: someone who did not build the workflow changes a field mapping, reruns the acceptance cases and diagnoses a failed connection. If they cannot do that with the supplied material, identify the missing training or support before calling the handover complete.

Decide who responds when a run fails or a source becomes unavailable. A scheduled weekly process needs a clear missed-run signal and a manual fallback. A command used on demand needs understandable errors and a way to recover without repeating an unknown-outcome write.

## Test portability with the assets that matter

“We own it” can mean ownership of a prompt file, the source code or a complete operable system. Specify which one is being delivered.

Keep business definitions, acceptance examples and data contracts in a form your team can retain. Document platform-specific behaviour that would need rebuilding if you moved. A standard protocol can make connections more portable while leaving authentication, permissions, scheduling and UI tightly coupled to a platform.

Ask what can be exported: instructions, tool definitions, run histories, evaluation results and configuration. Ask how to revoke access, stop schedules and remove credentials at the end of the engagement. For custom software, check the rights to modify and reuse the delivered work and dependencies in the actual agreement.

The aim is an informed maintenance decision. Rebuilding everything for theoretical portability can cost more than the dependency you were trying to avoid.

## What the martechdevs offer includes

Our [custom martech AI agent service](https://www.martechdevs.com/custom-martech-ai-agent) currently includes an agent at no additional charge within a fixed-price implementation. The service centres on Claude Code skills, commands, business context and access to the agreed tools, with documentation and handover.

The implementation still has a price and an agreed scope. “Included” should not be interpreted as unlimited future workflows, included third-party subscriptions or perpetual maintenance. Put the workflows, acceptance cases, ownership, exclusions and post-handover support in the quote.

The [implementation cost guide](https://www.martechdevs.com/blog/martech-implementation-cost) explains our starting price and broader warehouse, visualization and activation scope, alongside the inputs needed for a scoped quote. Use those service ranges separately from the fictional platform comparison above.

You should be able to compare that proposal with configuring your existing platform. A good outcome may be native tooling, a small extension or a custom implementation. Bring the workflow and the evidence requirements to the [quote request](https://www.martechdevs.com/custom-martech-ai-agent#get-quote) so the comparison starts with the same job.

## A buying checklist for founders and marketing leaders

- [ ] Define one completed business job and its users.
- [ ] Confirm why ordinary automation is insufficient.
- [ ] Give every option the same inputs and acceptance cases.
- [ ] Test required records, external tools and permissions directly.
- [ ] Reject failures on mandatory requirements before scoring preferences.
- [ ] Price setup, incremental licenses, usage, review and maintenance.
- [ ] Test sensitivity to retries, volume and operator hours.
- [ ] Separate external spending from the value of internal capacity.
- [ ] Assign business, technical and exception-handling ownership.
- [ ] Verify handover, export, support and the ability to stop the workflow.

## Frequently asked questions

### Should we build our own marketing AI agent?

Build when you have a valuable requirement that configured or extended tools cannot satisfy, and a team capable of maintaining the result. First test simpler options against the same acceptance cases. The ability to build a prototype is not enough reason to take on a permanent operating responsibility.

### Can HubSpot agents work outside HubSpot?

Current Agent Builder documentation includes supported external MCP connections. Verify the specific server, actions, authentication and subscription requirements for your workflow. An available connection does not by itself resolve account identity, conflicting definitions or the reliability of the external data.

### Can one agent use HubSpot, Google Ads and a warehouse?

A workflow can be designed to use several systems when the necessary tools and permissions are available. Evaluate each required read or write separately. Shared campaign identifiers, account relationships, reporting dates and metric definitions determine whether the combined output is meaningful.

### How much does a custom AI agent cost?

The useful estimate includes implementation, model or application usage, hosting, third-party services, review and maintenance. Scope the workflow before requesting a price. The worked comparison above is hypothetical; martechdevs includes an agent within an agreed fixed-price implementation, rather than offering an unlimited standalone system for no cost.

### Who maintains a custom agent after handover?

The owner named in the agreement should maintain it, whether that is your team or a contracted provider. Require operable source or configuration, documentation, acceptance cases and access to the necessary accounts. Verify the handover through an actual change and recovery exercise instead of relying on a folder of files.

## Sources

1. [Anthropic: building effective agents](https://www.anthropic.com/engineering/building-effective-agents).

2. [HubSpot Agent Hub](https://www.hubspot.com/products/artificial-intelligence).

3. [HubSpot Agent Builder](https://www.hubspot.com/products/artificial-intelligence/ai-agents).

4. [HubSpot: create and customize agents](https://knowledge.hubspot.com/ai/create-and-customize-agents-in-the-agent-builder).

5. [Salesforce: connect Agentforce to MCP servers](https://help.salesforce.com/s/articleView?id=ai.agent_mcp_connect.htm&language=en_US&type=5).

6. [HubSpot credit management](https://knowledge.hubspot.com/account-management/understand-hubspot-credits-and-billing).

7. [Salesforce Agentforce pricing](https://www.salesforce.com/agentforce/pricing/).

8. [Claude Code cost management](https://code.claude.com/docs/en/costs).