An agentic CRM is a customer system where software, not the rep, is the primary author of the record and the initiator of the next action. Agents watch the raw evidence — calls, email, calendar, product events — write the account and opportunity fields themselves, and then act inside a scope you granted: draft the follow-up, move the stage, open the renewal task. The system of record stops being a place people type into and becomes something that runs whether or not anyone logs in.
It is not a CRM with an AI assistant in the sidebar. A summarize button, an email drafter, and a “what changed on this account” panel are all request-response features: a human decides something needs doing, asks, reads, and still types the result into a field. That is a system of record with AI features, and every major CRM has shipped one. The distinction is not how good the model is. It is who holds the pen.
The one-question test
Pick a week and ask what your pipeline looks like if no rep opens the CRM at all. In a system of record, the answer is that the pipeline is frozen — stale stages, empty next-steps, no logged calls, because the record only reflects what humans found time to enter. In an agentic CRM, the answer is that the record still moved: the calls got attached to the right opportunity, the stage advanced or slipped based on what was said, and the follow-ups went out.
That test also exposes the marketing. If a vendor’s agentic story collapses into “your reps will type less because we prefill the form,” the human is still the author and the agent is a typeahead. Useful, but priced and governed like a feature, not like a worker.
Two architectures wear the same label
The agent layer on top. Rox sits above the CRM you already run rather than replacing it — its own site describes integrating with “Salesforce, HubSpot, Gmail, Microsoft Outlook, Slack, and others” and frames the product as an applied AI stack for autonomous revenue, with per-account agents handling research, monitoring, outreach, and write-back. Nothing migrates. Your records stay where they are, and the agents work over them.
The AI-native CRM underneath. Day.ai builds its own object model and positions agents as the workforce over it, with named archetypes including a CRM Data Specialist that will “read the transcript, find the Opportunity it belongs to, and move it to exactly the stage the conversation earned,” plus RevOps analyst, sales engineer, and BDR roles. Here the record itself is new, which is a migration, not an install.
The incumbents’ version. Salesforce Agentforce and HubSpot’s Agent Hub — the July 2026 rebrand of what shipped as Breeze Agents — bolt an agent runtime onto the system of record you already bought. The upside is that the data model, permissions, and audit surface are the ones your admins already know. The cost is that the agent inherits every bit of schema debt you have accumulated.
The buying decision is mostly this: are you replacing the record, or renting labor over it? A team with two years of Salesforce customization and a compliance surface is renting labor. A 15-person company whose CRM nobody trusts enough to forecast from is replacing the record.
The billing unit is the tell
Nothing exposes the category shift faster than the price list. Checked 2026-08-06:
- Day.ai charges per agent deployed, not per person: Free, Turbo at $25/month, Professional at $60/month, Executive at $200/month, with 2, 5, and 10 automated skill slots respectively and 20% off annually. Its own page states “no per-seat fees, no usage-based pricing.”
- Salesforce Agentforce prices the Agentforce User License at $5/user/month and marks it “Requires Flex Credits” — the seat is now a small access fee in front of a meter. Flex Credits are fungible across actions, prompts, translations, and voice actions, sold in $500 blocks, and Salesforce’s own worked examples value each credit at half a cent: 40 credits per request, 20 requests a day, 30 days, is 24,000 credits and $120 a month. Conversation-based pricing runs $2 per conversation.
- HubSpot went outcome-based: $0.50 per conversation the customer agent resolves (50 credits) and $1.00 per lead the prospecting agent recommends (100 credits), which puts a HubSpot credit at one cent.
- Attio shows the hybrid clearly — seats still exist at $35 (Plus) and $79 (Pro) per user/month billed annually, but each seat carries a credit allowance of 100, 500, 1,000, and 2,500 per user/month across Free, Plus, Pro, and Enterprise, with extra workspace credits sold separately.
Read across those four and the pattern is that the seat count stops predicting the invoice. Budget the pilot on units of work at your real volume — resolutions, recommended leads, agent-runs, credits — because a headcount-based model gives you no signal about what a busy quarter costs.
Diagnostic questions for a vendor
- Who authored this field? Ask to see a record where agent-written and human-written values are distinguishable in the UI and in the export. If they are not, your pipeline history quietly becomes model output.
- What write scopes does the agent hold, and can I run it read-only first? A dry-run or proposal mode that queues changes for approval is the difference between a pilot and an incident.
- What happens on hand-off? When the research agent passes to the outreach agent, name which one owns state and what the human gate is. Several agents sharing a task list is orchestration; several agents sharing a logo is a feature bundle.
- What does one unit of work cost at our volume? Get the number for your monthly conversation, lead, or run count — not the list price.
- Can I cap it? Ask for per-agent monthly run limits. HubSpot’s own guidance to review estimated credit costs and set monthly run limits is the tell that unbounded runs are a real failure mode.
- What survives an audit? Action logs tell you what changed. You need the evidence the agent acted on, linked to the record, or “the model decided” is your whole explanation.
Watch-outs, each with a guard
Agent-written history contaminates your analytics. If an agent advances stages, your stage-conversion rates become a measurement of the agent’s judgment rather than the buyer’s behavior. Guard: keep a provenance flag per field, and rebuild forecast and conversion reporting to filter or segment on it before you trust a quarter of it.
Autonomy outruns your data quality. An agent acting on a duplicated account or a stale owner field acts confidently and wrongly at machine speed. Guard: run the CRM hygiene pass first; a deduplication backlog is a prerequisite, not a follow-up.
The meter has no ceiling by default. Consumption pricing turns a busy quarter into a budget variance nobody approved. Guard: set per-agent run caps on day one and reconcile the vendor’s usage report against your own event counts monthly.
Stage definitions become model prompts. Once an agent moves opportunities, your written stage criteria are the specification the model is executing against. Guard: rewrite deal stage definitions as evidence tests the agent can satisfy — a stage requires a linked artifact, not a vibe.
When you actually need one
Buy agentic when the gap you are closing is unrecorded work: reps who run good calls and log nothing, accounts nobody has touched in 60 days, renewals that surface a week late. That is the failure an agent fixes directly, and it is a failure a better form layout has never fixed.
Skip it when your problem is process or schema. If deals stall because approvals take nine days, or your pipeline is wrong because three teams define “qualified” differently, an agent will execute that ambiguity faster and produce a more confident version of the same mess. Fix the definitions, then hand them to the agent.
For the general pattern behind this category, see what makes an AI agent for ops and AI agent vs RPA.