The demo went well. The agent answered a pricing question, booked a meeting and logged the activity to the right contact. Three weeks after launch, a prospect asks a two-part question about contract terms, the agent answers the first part confidently and wrongly, and nobody finds out until the account executive inherits an angry thread. The dashboard still shows a healthy resolution rate, because the conversation never reached a human.
That is the decision behind most Agentforce vs Breeze evaluations, and behind the third option of building custom agents. Each can produce an agent that works in a demo. They differ in how much control you have over its behaviour, how well you can test it before it touches customers, and what you see when it goes wrong. Choose on features and you find out about reliability from your customers.
The CRMArena-Pro benchmark comes from Salesforce's own research team, who tested leading models on realistic sales, service and quote tasks in May 2025. It measures raw LLM agents, not configured Agentforce or Breeze deployments, and it also found the agents showed near-zero awareness of confidential data by default. Gartner's figure is a prediction; it followed a January 2025 poll of 3,412 webinar attendees, in which 19% said their organization had made significant investments in agentic AI. HubSpot's resolution figure is the vendor's own and uses its own definition of "resolved," which matters, as we explain below. HubSpot's product page now shows a 70% average resolution rate; we use the dated April 2026 figure.
How we evaluated Agentforce vs Breeze vs custom agents
VANDFORT publishes this guide, and VANDFORT builds custom agents and agent-driven systems for revenue teams, so we are one of the options and have a stake in the answer. To keep the comparison fair, we fixed six criteria before reviewing either platform and applied them to every option, including our own.
| Criterion | What we look for |
|---|---|
| Behaviour control | Whether you can force fixed steps where a mistake is costly, or only guide the agent with instructions. |
| Pre-launch testing | Whether you can run many test cases at once, with expected outcomes, before customers see the agent. |
| Production visibility | Whether every session can be traced: inputs, reasoning steps, actions, errors and escalations. |
| Failure handling | How the agent hands off to a human, and whether a wrong answer can hide inside a "success" metric. |
| Data reach | What the agent can read and act on: one CRM, or also billing, product and support systems. |
| Pricing model | Published price or pricing model, as of October 2026, and how predictable it is as volume grows. Our own prices are omitted under our editorial policy. |
We weight failure handling most heavily. An agent that fails loudly costs you a ticket. An agent that fails quietly costs you a customer, and you learn about it last. Vendors are described only from their own product, pricing and documentation pages, or dated independent reporting, as of October 2026. Naming a vendor is not an endorsement, and we don't rank anyone first.
The options compared
| Option | Best for | Strengths | Limits | Pricing model, as of October 2026 |
|---|---|---|---|---|
| HubSpot Breeze | HubSpot teams wanting packaged support and prospecting agents | Fast to launch; batch testing and gradual rollout; outcome pricing | Customer Agent needs Professional or Enterprise; "resolved" is a billing definition | Customer Agent $0.50 per resolution; Prospecting Agent $1 per qualified lead; custom agents use HubSpot Credits |
| Salesforce Agentforce | Salesforce teams needing fine control over agent behaviour | Six levels of determinism up to Agent Script; Testing Center; session tracing | More configuration to get right; tracing data stored in Data 360 | Flex Credits $500 per 100,000 (20 per standard action), or $2 per conversation; add-ons $125 per user a month |
| Custom agents, built in-house | Teams with engineers and workflows that cross systems | Full control of logic, data reach and model choice | You own testing, monitoring, security and upkeep | Model, hosting and tooling costs plus engineering time |
| VANDFORT | $3M to $30M ARR teams without an in-house agent team | Custom systems tested on about 20 of your own past cases before go-live | An engagement, not self-serve software; one system at a time | Scoped per engagement; not published here |
HubSpot Breeze: packaged agents inside the CRM
HubSpot lists four Breeze agents: a Customer Agent that answers questions and resolves support tickets, a Prospecting Agent that researches accounts and drafts outreach, a Data Agent that answers questions from CRM records, calls and documents, and Breeze Assistant. Agent builder, for custom agents, is included in Starter, Professional and Enterprise, and those agents consume HubSpot Credits each time they run. HubSpot's credits documentation notes that when included credits run out, features that use them pause until the next reset unless you have bought capacity packs, which is a reliability question as much as a billing one. The Customer Agent requires Professional or Enterprise. HubSpot's page says you can batch test scenarios and roll out gradually before scaling, and that the agent escalates complex issues to your team "with full context."
On April 14, 2026, HubSpot moved the Customer Agent from $1 per conversation to $0.50 per resolved conversation, and the Prospecting Agent to $1 per lead when a prospect is qualified and handed to your team, according to MarTech's report of the announcement. The detail that matters for reliability is the definition. HubSpot counts a resolution when the agent provided support and the conversation isn't handed to a human rep for 72 hours. That is a fair billing rule, but it is not a correctness check. A customer who gets a wrong answer and leaves counts as resolved. Breeze is better suited to high-volume, well-documented questions than to conversations where a wrong answer carries contractual or pricing risk.
Salesforce Agentforce: configurable control on the Salesforce platform
Agentforce gives you more dials. Salesforce's guide to reliable agent behaviour describes six levels of agentic control, moving through instructions, data grounding, agent variables and actions such as Flow or Apex, up to Agent Script, which lets you "hard-code" the reasoning process itself. Testing Center, introduced in January 2025, runs batches of test cases in a sandbox, each with an utterance, expected topics, expected actions and an expected outcome, so you can compare what the agent did with what it should have done. Agentforce Observability, formerly Command Center, tracks error rates, escalation frequency and latency, and its session trace data model logs each session's inputs, reasoning steps, LLM calls and guardrail checks.
Pricing has three models. Flex Credits cost $500 per 100,000, and a standard action uses 20 credits, which works out to $0.10 an action. Customer-facing agents can instead be priced at $2 per conversation, and Salesforce says the two models are not supported in the same org. Per-user add-ons for Sales or Service cost $125 per user a month with unmetered employee use. The trade-off is effort: every level of control is something a Salesforce admin or developer has to design, test and maintain. Agentforce is better suited to teams with Salesforce development capacity than to a lean team that wants an agent running this week.
Custom agents built in-house
The third route is to build agents on an open framework and connect them to your CRM by API. LangGraph, one common choice, says it lets you "mix deterministic, hand-coded steps with LLM-driven agentic steps in the same graph," with agents that persist through failures and human oversight "by inspecting and modifying agent state at any point." A custom build can read billing, product usage and support data in the same run, choose its own model, and follow whatever escalation logic you write. LangChain's survey of more than 1,300 professionals found performance quality was the top concern, more than twice as significant as cost or safety, and tracing and observability tools topped the list of must-have controls.
The limit is ownership. Everything the platforms give you (testing harnesses, session traces, permission checks and audit trails) you build or assemble yourself. We set out what that instrumentation looks like in our piece on why most AI agents fail silently. Custom agents suit companies with engineers who will still be maintaining them a year from now.
VANDFORT: custom systems, tested before they ship
VANDFORT is a custom build that we engineer and run inside your stack, so you don't need an in-house agent team. Each system is tested on around 20 of your own past cases before it goes live, and it ships at 85 percent or it does not ship. We build one system at a time, for example Speed-to-Lead for inbound response or the Pipeline Hygiene Sentinel for CRM data quality, and we often work alongside Agentforce or Breeze rather than replacing them. The limits are real: it is an engagement measured in weeks, not a tool you configure this afternoon, and it doesn't make sense for one well-scoped agent your platform already does well.
How to choose
Write down the job and the cost of a wrong answer. "Answer documented product questions" is a different risk from "quote renewal terms" or "update the forecast category." For each job, write what a wrong action costs and whether it can be undone. Our risk and reversibility matrix is one way to sort them.
Map the data the job needs. If every input lives in your CRM, the native agent for that CRM starts with an advantage. If the job needs billing, product usage or a second CRM, price the integration work before you compare agent features.
Decide where the path must be fixed. Mark the steps that must happen the same way every time, such as verifying identity or checking a discount limit. Those steps need deterministic logic, whether that is Agent Script, a workflow step or code. We cover where that line sits in our agentic vs static workflow framework.
Test on your own past cases before launch. Pull 20 to 50 real conversations or records with known correct outcomes, including multi-step ones and a few the agent should refuse. Run them through each option and score the outcome, not just whether the agent replied.
Define your own success metric. Don't adopt the vendor's billing definition as your quality metric. Sample resolved conversations every week and check them by hand, and track escalations, corrections and complaints next to resolution rate.
Price at three times today's volume. Model cost per outcome or per action at current volume and at three times it, and include the people who review and correct the agent's work.
A rule of thumb, offered as a suggested starting point rather than a benchmark: if the job lives inside one CRM and a wrong answer is cheap to fix, start with that CRM's native agent. If the job crosses systems or a wrong answer is expensive, design the fixed steps and the test set first, then decide whether the platform or a custom build can hold them.
Where VANDFORT fits, and where it doesn't
You launched a Breeze or Agentforce agent, and its success metric looks good while your team keeps fixing its mistakes by hand. The job you want to automate crosses your CRM, billing and support data, so no single native agent sees enough. You are a $3M to $30M ARR B2B SaaS company without engineers to own testing and monitoring. Or you want proof on your own past cases before anything goes live. Our systems page lists the nine systems we build, and our governance guide covers the guardrails and audit trails we put around them.
If you need a support agent for well-documented questions on HubSpot, turn on the Customer Agent and test it; that is what it is built for. If you have Salesforce developers and the job stays inside Salesforce, Agentforce's controls are likely enough. If you want a self-serve tool your admin configures in a day, buy software. And if you already have an engineering team running agents with tests and tracing, extend what you have before you bring in an outside team.
Frequently asked questions
What is the difference between Agentforce and HubSpot Breeze?
Agentforce is Salesforce's agent platform and Breeze is HubSpot's. Both build AI agents that work on CRM data. Breeze offers packaged agents for support, prospecting and data questions, with outcome pricing for its Customer and Prospecting agents. Agentforce offers more configuration, including hard-coded reasoning paths with Agent Script, batch testing in Testing Center and session tracing. The practical difference is control and effort: Breeze launches faster, Agentforce gives you more to tune.
How much does Agentforce cost compared to Breeze?
As of October 2026, Salesforce lists Agentforce Flex Credits at $500 per 100,000 credits, with a standard action using 20 credits, or $2 per conversation for customer-facing agents. Sales and Service add-ons are $125 per user a month. HubSpot charges $0.50 per resolved conversation for its Customer Agent and $1 per qualified lead for its Prospecting Agent. Custom agents in both platforms consume credits.
Are AI agents reliable enough for customer-facing work?
For narrow, well-documented tasks, often yes; for multi-step conversations, test carefully first. Salesforce AI Research's CRMArena-Pro benchmark found leading agents succeeded on about 58% of single-turn business tasks and 35% of multi-turn ones. Production agents with fixed steps, grounding and escalation rules may do better than raw models, but you should measure that on your own past cases before launch.
Can Breeze or Agentforce agents escalate to a human?
Yes, both can. HubSpot says its Customer Agent escalates complex issues to your team with full context. Agentforce Observability reports escalation frequency, so you can see how often agents hand off. Check how escalations are counted: HubSpot treats a conversation as resolved if it isn't handed to a human within 72 hours, so a missed escalation can look like a success.
When should we build custom AI agents instead of using our CRM's agents?
Build custom when the job needs data your CRM doesn't hold, such as billing or product usage, or logic the platform can't express, and when you have engineers to own testing and monitoring. If the job lives inside one CRM and mistakes are cheap to fix, the native agent is usually faster and cheaper. Many teams run both: native agents for simple jobs, custom ones for cross-system work.
Sources: Salesforce AI Research, CRMArena-Pro (Huang et al., arXiv, May 2025). Gartner, press release on agentic AI project cancellations (June 25, 2025; poll of 3,412 webinar attendees, January 2025). MarTech, report on HubSpot's outcome-based pricing for Breeze agents (April 2, 2026). HubSpot, Breeze AI agents and Customer Agent product pages, and HubSpot Credits knowledge base article. Salesforce, Agentforce pricing page; levels of determinism guide; Agentforce Observability blog post (June 2025, since updated); Salesforce Admins blog on Testing Center (January 2025). LangChain, LangGraph documentation and State of AI Agents survey (1,300+ respondents). Vendor pages accessed October 8, 2026.
Last reviewed: October 2026. Vendor features and pricing change; check each vendor's site before buying.
