The hidden cost of bolt-on AI in a legacy CRM is the effectiveness tax: agents working on shallow data do shallow work, and the gap compounds every quarter. The subscription line is visible on the order form and small by comparison. What an AI-powered CRM returns on that spend is decided by the architecture underneath it, before the first invoice arrives.
Every legacy vendor now sells an AI-powered CRM. The system you already run, with copilots and agents added on top, priced per conversation, per action, or per seat. The AI-powered CRM pitch is clean, and the expensive part appears nowhere on it. For a staffing firm choosing how its recruitment desk will run for the next several years, that invisible line is the decision.
Why this matters now: teams are choosing the system they'll sell from for the next decade
AI moved from suggestion-layer to execution-layer. Agents schedule, draft, update, and prep. They do real work instead of recommending it, and the salespeople closest to the work felt the change first.
A team carrying a high volume of routine, repetitive sales work, with a staffed motion and a decision already made to raise the AI-intelligence of how it sells, is choosing a system for an existing motion. That choice sets the ceiling on the next decade, because what AI-native means at the category level is an architectural shift, and architecture is the part you can't patch later.
The vendors confirm the urgency in their own pricing churn. Salesforce launched Agentforce at $2 per conversation, then added Flex Credits at roughly $0.10 per action. HubSpot moved two flagship Breeze agents to outcome pricing on April 14, 2026 (HubSpot company news). Two vendors, three pricing rewrites, well under two years: a market re-architecting around agents in public, before any AI-powered CRM order form gets signed.
Companies retrofitting AI onto legacy stacks hit a ceiling. Companies built around agents from day one are still finding theirs.
What do most teams get wrong about the AI-powered CRM?
The trap is treating AI as a feature you add to the stack you already run. Generic AI on a generic record does generic things: it drafts the email the rep already had in their head and surfaces the insight the rep already knew. Legacy CRMs store what reps type; agents on top of legacy data work on shallow signal.
Two patterns produce that shallow signal, and they look different on an invoice. The deeper treatment of the split is the AI-native vs. AI-enabled architecture question.
Bolt-on pattern one: AI features added to a legacy data model
Salesforce with Einstein and Agentforce, HubSpot with Breeze. AI capability attached to a record that was built for reps to log activity after the fact. The category is honest about itself: an AI-powered CRM is a CRM first, with AI in the feature list. A system whose data model is built for agents to read and write in real time is a different category, with a different ceiling.
The agent treats the record as ground truth. And in the AI-powered CRM, the record holds whatever a rep typed three days after the call, three fields thin.
Bolt-on pattern two: AI point tools stitched around the CRM
Clay, Apollo, Lavender, 11x, and Regie automate one slice of the motion, usually outbound, and they are useful in their category. The platform question is a different one. Each point tool holds its own slice of context, the legacy CRM underneath still stores what reps type, and the full picture of the motion lives nowhere. Agents that work across capture, qualification, prep, execution, post-call, pipeline, and renewal need the system of record itself built for them, not stitched around them.
What does bolt-on AI actually cost?
Bolt-on AI in an AI-powered CRM carries three costs. One of them is on the order form.
The pricing tax
Agentforce's original consumption model bills $2 per conversation, defined as a 24-hour interaction window. Flex Credits bill $500 per 100,000 credits, with a standard action at 20 credits, about $0.10 (Salesforce pricing page). Salesforce's own example on that page: 100 users handling three cases a day runs $1,800 a month in credits, before seat licenses.
Per-user routes start at $125 per user per month for add-ons, with higher bundled editions on top, per Salesforce's pricing page. On the HubSpot side, since April 14, 2026, Breeze Customer Agent is priced at $0.50 per resolved conversation and Prospecting Agent at $1 per recommended lead (HubSpot company news).
The numbers that frame the choice:
- $0.10 — Per standard action, Agentforce Flex Credits
- $2 — Per conversation, Agentforce original model
- $125+ — Per user/mo, Agentforce add-ons
- $0.50 — Per resolved conversation, HubSpot Breeze
- ~20 min — Per agent to configure Sonta
Article 10 carries the full Agentforce pricing breakdown. The pattern matters more than any single rate. These are rational prices from the vendor's side. From the buyer's side they share one property: the AI-powered CRM meter sits on top of the seat license, so the AI line grows with exactly the usage you were hoping to scale.
The integration tax
In Sonta's sales conversations, most teams evaluating are running four to six tools across CRM, sales engagement, meeting intelligence, and AI. Each tool holds part of the context, and somebody syncs it back into the AI-powered CRM at the center.
RevOps owns the field mappings, the dedupe rules, and the integration that broke the week a vendor shipped an API change. None of that labor appears on the AI line. It gets paid in hours, and in records that disagree with each other by Friday.
The effectiveness tax
This is the cost that compounds, and the order form hides it completely. An agent is only as good as the record it reads. On a recruitment desk, a candidate who has placed three software engineers with the firm over five years gets the same outreach as a cold LinkedIn lead, because the AI-powered CRM's agent reads a notes field, and the notes field says nothing. On a system where candidate, placement, and client histories are first-class objects in the data model, the same request produces outreach that reads like the recruiter wrote it.
Multiply that gap across every follow-up and every account prep, quarter after quarter, and the per-action price starts to look like what it is. Our position, plainly: the meter is a rounding error. The architecture underneath decides the return on every dollar of the AI line above it.
Run the evaluation as criteria you apply inside your own flow:
| What to test | AI-powered CRM (AI on a legacy record) | AI point tools around the CRM | Agentic CRM |
|---|---|---|---|
| Where the AI runs | On top of the existing record | In a separate tool, synced back | Inside the system of record |
| What the AI reads | What reps typed after the fact | The tool's own slice of context | Records that update as a side effect of work |
| How the AI is priced | Metered usage on top of seats | A subscription per tool | Platform price; AI at model cost |
| What your team maintains | The add-on's configuration | The sync between systems | One platform |
The position: the agentic CRM as the architecture alternative
Sonta is the agentic CRM for AI-first GTM teams, built AI-first from the ground up, not a legacy CRM with AI capabilities added on. Sonta owns three layers on behalf of the customer: the context (every important event in the business, structured), the agents (configured automations executing the work), and the processes (the workflows that connect them). The customer owns all three as portable assets. When the frontier moves, Sonta moves the work onto the right model for the workflow, whichever of Claude, Gemini, or OpenAI that is, and the customer keeps running.
The pricing follows the architecture. No proprietary AI tax. No per-conversation fees defending a 25-year-old margin.
Where the AI-powered CRM meters usage on top of seats, the agentic CRM prices the platform and passes AI through at model cost, so the AI line falls as frontier prices fall. Configuration is part of onboarding; in Sonta's onboarding experience, an agent takes about 20 minutes to set up. The work compounds from week one, because every action an agent takes deepens the record the next action reads.
What does bundling the architecture actually deliver?
One platform replacing CRM, sales engagement, and meeting intelligence. One context, one bill, one place to work.
On a recruitment desk, the daily version looks like this: an interview wraps, and the agent logs the recruiter's voice memo into the candidate record, updates the placement's stage, drafts the client update, and queues the next outreach. Nothing was synced and nothing was retyped. The record got deeper because work happened.
Deep records are the compounding asset. The next account prep starts from more signal than the last one, and the Friday pipeline review compiles itself from context the agents wrote all week. A staffing firm running the agentic CRM built for recruitment teams pays one line item for the whole motion, and the system gets better at the desk every week the desk runs.
The order form will show you the price of an AI-powered CRM. The architecture will show you what it costs.
Frequently asked questions
How much does Salesforce Agentforce cost in 2026?
Between a free Foundations tier and $550+ per user per month for Agentforce 1 Editions, depending on the model. Consumption pricing runs $0.10 per standard action through Flex Credits ($500 per 100,000) or $2 per conversation, and the two models can't run in the same org. Per-user options start at $125 per user per month for add-ons, with higher bundled editions and a free Foundations tier as the entry point, per Salesforce's pricing page. Agentforce pricing has changed multiple times since launch.
Does HubSpot Breeze require an extra subscription?
There is no separate Breeze subscription; the agents ship inside HubSpot's Professional and Enterprise tiers. Since April 14, 2026, Breeze Customer Agent is priced at $0.50 per resolved conversation and Prospecting Agent at $1 per recommended lead, per HubSpot's announcement.
What's the AI cost of an agentic CRM vs. AI-powered CRM?
Sonta uses frontier AI (Claude, Gemini, OpenAI) without a proprietary AI tax and without per-conversation fees. Your AI cost is the model cost, transparent. We don't mark it up to defend a 25-year-old margin. As frontier model prices drop, your AI cost drops with them.
Can I run Sonta on top of Salesforce?
No. Sonta is the agentic CRM itself, the system of record the agents read and write, and that is the point of the architecture. Teams moving from an existing stack migrate in stages: agents start where they add the most value (typically account prep, post-call updates, and multi-thread orchestration), the existing CRM runs in parallel during transition, and the team cuts over once it trusts the agents. Staged migration is designed to reach full cutover in 60–90 days.