Agentforce
Agentforce vs. Lyzr: the CRM-native agent layer against the deploy-anywhere platform
Lyzr is the Agentforce alternative regulated enterprises keep shortlisting: full-stack agent infrastructure that runs in your own VPC, routes to any model, and prices per run. Agentforce is the agent layer of the CRM your sellers and service reps already live in. This is the comparison across hosting, grounding, model freedom, build model, governance and pricing, and the one question that decides it.
Lyzr keeps turning up on Agentforce shortlists next to much bigger names. The other three are usually Amazon Bedrock AgentCore, OpenAI and Sierra, and I have compared each of them already.
Lyzr is the interesting one. It isn’t a hyperscaler or a model lab, and it doesn’t solve one department’s problem. The Bengaluru-founded company positions itself between open-source frameworks like LangGraph and closed ecosystems like Agentforce.
It sells the whole agent stack as software you run in your own cloud, and it has spent the past year collecting the kind of backing that gets a startup through a bank’s procurement desk: an Accenture Ventures investment and a reported nine-figure round.
That makes it a fair fight, because the two products answer different questions. Agentforce asks what agents should do inside the CRM. Lyzr asks what infrastructure an enterprise should run its agents on, wherever they are.
The one-line difference
Agentforce is the agent layer of Salesforce. Agents reason over CRM and Data 360 records, act through Flows, Apex and prompt templates, and inherit the org’s sharing model, all metered on Salesforce’s consumption model and hosted by Salesforce.
Lyzr is full-stack enterprise infrastructure for building, evaluating, deploying, and governing AI agents: studio, SDK, orchestration, memory, retrieval and guardrails, which you can run as SaaS, inside your own VPC, or hybrid, against whichever model provider you choose.
With Agentforce, the system of record is the platform. With Lyzr, the platform is deliberately not a system of record. It sits across the ones you have.
Hosting and data ownership: Lyzr’s strongest card
Lyzr wins this dimension outright, and it is the reason it appears on regulated shortlists at all.
Lyzr’s documentation describes three deployment modes: SaaS, where Lyzr hosts the platform; on-premise, where the platform runs inside your own VPC; and hybrid, where Lyzr orchestrates agents in the cloud while your data connectors stay on-premise. The platform is described as HIPAA and SOC 2 compliant with regional hosting available.
Accenture’s press release frames the pitch as open-source flexibility with managed-platform security, within the customer’s environment, for complete data privacy and IP ownership.
For a bank whose model-risk team will not let prompts and retrieved documents leave a controlled boundary, that settles the conversation. Agentforce’s answer to the same objection is the Einstein Trust Layer plus, since Claudeforce, inference contained inside the Salesforce trust boundary.
A strong answer, but it is Salesforce’s boundary, not yours. If the requirement reads “our VPC, our keys, our region”, Lyzr can say yes to the letter of it and Agentforce cannot.
The counterweight is that running the platform yourself means running the platform yourself. Upgrades, capacity, the vector store, the observability stack. Lyzr’s self-hosted price is lower because you carry the infrastructure. Agentforce’s hosting is invisible because Salesforce carries it. Which one you want is a staffing question before it is a technology one.
Grounding: whose data, and how close
Agentforce’s grounding advantage is proximity. Agents retrieve against CRM records with sharing enforced, against Data 360’s unified profiles and unstructured content, and against knowledge articles, without an integration step.
The Atlas Reasoning Engine plans over actions that already exist in the org. When the work is a service case or a sales opportunity, the data is already where the agent is.
Lyzr’s grounding is built in but brought in. Its stack includes enterprise RAG with secure retrieval and a Cognis memory layer for persistent context, and its product page describes wiring retrieval to your own vector stores and data sources through native connectors, naming Salesforce, SAP and ServiceNow. The retrieval layer is complete on its own.
For an agent that must reason across a core banking system, a policy admin platform and the CRM at once, Lyzr’s shape is the more natural one, because none of those systems is privileged. The cost is that every source is a connector you own, and the freshness and permission model of each retrieved document is yours to enforce rather than inherited from a sharing model.
Model freedom: routing against a menu
Lyzr is model-agnostic in the strong sense. Its product page describes routing each agent to OpenAI, Anthropic, Gemini, Bedrock, or DeepSeek, and because the platform can run in your VPC, a self-hosted open-weight model is a routing target too.
Per-agent routing is a real architectural advantage. The classification agent runs on something cheap, the reasoning agent on something capable, and switching either is configuration rather than a rebuild.
Agentforce has moved a long way here: a model menu, bring-your-own model through Model Builder, per-subagent overrides, and Claude as a default in several products since Claudeforce. But the menu is curated by Salesforce and the models run through Salesforce’s path. If your AI governance standard says “we choose and host the model”, Lyzr fits it and Agentforce accommodates it.
Build model and prebuilt agents: two kinds of speed
Both platforms sell speed and they mean different things by it.
Lyzr’s speed is breadth. Agent Studio spans a visual no-code studio, a Python and TypeScript SDK, and raw REST APIs. Orchestration is deterministic through SuperFlow or dynamic through a Manager Agent, and voice agents deploy over Twilio, Telnyx or Plivo.
On top sits a catalogue Lyzr describes as 150+ prebuilt agents and 200+ integrations, with industry blueprints. Accenture’s release names insurance claims processing, policy renewals, endorsements and mid-term changes, and banking loan approvals, onboarding and GRC audits.
Lyzr’s pricing page commits to eight weeks from use case to agents in production with field development engineers included. An unusually specific promise for a platform vendor to put in writing, and one of the more credible things about the company.
Agentforce’s speed is adjacency. The actions already exist. Every Flow, invocable Apex method and prompt template is a candidate, and Agent Script gives the deterministic control a regulated conversation needs without leaving the platform.
The prebuilt layer is narrower and deeper: Sales, Service and the industry clouds ship agents that know the objects they act on. What Agentforce does not do is build the claims agent for a policy admin system that is not Salesforce. What Lyzr does not do is inherit five years of Flow automation for free.
Where Lyzr deserves particular credit is evaluation. The platform covers building, evaluating, deploying and governing as one lifecycle, with a Hallucination Manager for verified outputs and built-in guardrails, and the company has been willing to eat its own cooking.
Its reported $100 million Series B was run in part by its own agent, which fielded questions from more than 130 investors and drafted investment memos. The founders were careful to say the agent started conversations rather than closed them. Which is the right claim about what an agent does, and rarer than it should be.
Governance: inherited against configured
Agentforce’s governance is inherited. Sharing rules, field-level security, the Trust Layer’s masking and audit trail, and the Testing Center exist before the first agent does, and a Winter ‘27 org adds Custom Scorers for session-level evaluation. The weakness is the flip side. It governs what is in Salesforce, and an agent that acts on three other systems carries three other governance models.
Lyzr’s governance is configured, and designed as a control plane rather than a feature: audit trails, responsible-AI controls, safety guardrails, and a compliance function that apply to every agent on the platform regardless of what system it touches.
For an enterprise standardising agent governance across many systems, one control plane is what the architecture review asks for. The obligation is that “configured” means someone configures it. The identity and permission model each agent runs under, against each system, is a design decision rather than an inheritance.
Pricing: per run against per conversation
Lyzr publishes its numbers, which in this market counts as a feature. Its pricing page lists $0.08 per agent run on Lyzr Cloud and $0.03 per run on VPC or on-premise, with model and compute costs passed through at usage rates, and a run covering knowledge-base queries, tool calls, orchestration, memory, guardrails and security policies.
The worked example on that page, a 20-agent KYC workflow at roughly a dollar per end-to-end run, is arithmetic a CFO can check. There is no free tier. Enterprise software, priced like it.
Agentforce prices on conversations and Flex Credits, with a per-outcome option for some prebuilt agents, and the model cost sits inside the metered unit. Neither is cheaper in the abstract. Lyzr is more legible, and for a high-volume back-office workflow in your own VPC it is likely lower per unit. Agentforce’s price includes hosting, the trust layer and the data adjacency you would otherwise build.
So which one
| Dimension | Agentforce | Lyzr |
|---|---|---|
| Where agents run | Salesforce-hosted, inside the Salesforce trust boundary | SaaS, your VPC or on-premise, or hybrid |
| Data ownership | Salesforce’s boundary; Trust Layer masking and zero data retention | Your boundary, your keys, your region |
| Grounding | Native to CRM, Data 360 and knowledge; sharing enforced | Built-in RAG and memory over connectors you own |
| Model choice | Curated menu, BYO model through Model Builder, Claude defaults | Per-agent routing across OpenAI, Anthropic, Gemini, Bedrock, DeepSeek, self-hosted |
| Build model | Builder plus Agent Script; actions from Flow, Apex, prompts | No-code studio, Python and TypeScript SDK, REST; SuperFlow and Manager orchestration |
| Prebuilt | Sales, Service and industry-cloud agents | 150+ agents and industry blueprints (Lyzr’s figures) |
| Governance | Inherited from the org; Testing Center, Custom Scorers | Configured control plane across all agents and systems |
| Pricing | Conversations, Flex Credits, per-outcome options | Published per run: $0.08 cloud, $0.03 self-hosted, plus model pass-through |
One question decides it. Is the system the agent must act on Salesforce, or is Salesforce one of several?
If the agent’s job is a case, a lead, an opportunity or a customer record, and the data it needs lives in CRM and Data 360, choose Agentforce. Every hour spent rebuilding sharing, actions and the trust layer on another platform is an hour you did not need to spend, and the readiness work is mostly done already.
If the agent’s job crosses systems where Salesforce is a participant rather than the centre, such as claims adjudication against a policy admin system, KYC across a core banking platform and a document store, or a procurement workflow in an ERP, choose Lyzr and connect it to Salesforce as one of its tools. The same goes if nothing may leave your VPC.
Accenture is investing behind that shape for financial services, and it is the right shape for that problem.
For a large regulated enterprise the mature answer is increasingly both, with a seam. Agentforce now delegates to any A2A-compliant agent and consumes external tools over MCP. Lyzr exposes its agents programmatically through an Agent API.
A Lyzr claims agent in the bank’s VPC and an Agentforce service agent in the contact centre can be two halves of one workflow, each governed where it lives.
What both get right, and what neither fixes
Both have converged on the same lifecycle: build, evaluate, deploy, govern, observe. Both treat guardrails as a product feature rather than a prompt.
Neither fixes your data. Lyzr’s retrieval is only as clean as the documents you connect. Agentforce’s grounding is only as good as the CRM hygiene underneath it. Neither fixes the permission model you bring, Lyzr because you configure it, Agentforce because you inherit it with its warts. And neither replaces the judgment about which decisions an agent should make alone, which is the question every one of my industry agent guides ends up being about, whatever the platform.
Understanding the basics
What is Lyzr?
Lyzr is an enterprise AI agent infrastructure company founded in 2023 in Bengaluru by Siva Surendira and Anirudh Narayan. Its Agent Studio and SDK let teams build, evaluate, deploy and govern agents, with orchestration, memory, retrieval and guardrails built in, and it runs as SaaS, inside a customer’s VPC, or hybrid. Accenture Ventures invested in October 2025, and a $100 million Series B at a reported valuation of around $500 million was reported in July 2026.
Is Lyzr a replacement for Agentforce?
Only when the agent’s work is not centred on Salesforce. For agents that act on CRM and Data 360 records, Agentforce’s native grounding, actions and inherited governance are hard to justify rebuilding. Lyzr replaces Agentforce most convincingly for cross-system, back-office or on-premise workloads where Salesforce is one system among several.
Can Lyzr agents work with Salesforce?
Yes. Lyzr lists Salesforce among its native connectors, and Agentforce can delegate to external agents over the A2A protocol and consume external tools over MCP, so the two platforms can share a workflow with each governed where it runs.
How is Lyzr priced compared with Agentforce?
Lyzr publishes per-run pricing: $0.08 per agent run on Lyzr Cloud and $0.03 per run self-hosted in your VPC, with model costs passed through. Agentforce prices on conversations and Flex Credits, with per-outcome options for some prebuilt agents, and the model cost sits inside the metered unit. Lyzr is more transparent. Which is cheaper depends on volume, hosting, and how much of Agentforce’s included platform you would otherwise have to build.
The comparison the shortlist is really making
When a Lyzr-against-Agentforce evaluation stalls, it is usually because the team is comparing feature lists when the real disagreement is about where agents should live. Lyzr’s answer, your infrastructure, your model, your control plane, priced per run, is a coherent bet, correct for a growing class of enterprise, and the company has earned the credibility to make it. Agentforce’s answer, inside the system of record, inheriting its governance, priced per conversation, is equally coherent for the work that lives there.
Decide where the work lives first. The platform follows, and for most large organisations it is not a single choice.
If you are weighing Lyzr against Agentforce for a specific workload, the useful first step is a map of which systems the agent must act on and what the data-residency requirement says. Talk to us and we will draw it with you, or size the CRM-side case first with the Agentforce ROI calculator.
