How to add AI on top of HubSpot or Salesforce without replacing it
Updated
The short answer. To add AI on top of HubSpot or Salesforce without replacing it, keep the CRM as the record your forecast runs on and add a layer that reads every call, email and Slack thread, does the work that follows from them, and writes the results back through the CRM's own MCP server with a rep approving what matters. Day AI is that layer, whether teams run it in place of their CRM or, as here, on top of Salesforce or HubSpot: it builds a customer memory from your team's conversations, gives every rep agents that prep meetings, draft follow-ups and keep records current, and writes back with each rep's own CRM permissions.
Say you run revenue at an audit-software company with about 23 AEs on HubSpot. Your team built its own workflow to update deals after calls, but it only saw the one call in front of it, and sometimes it put the update on the wrong deal. So you add Day AI on top of HubSpot. Now, after each recorded call, Day AI reads the call alongside everything else on the deal and proposes updates to its MEDDPICC scores. The rep looks over the proposal and approves it, and the new scores are written to the deal in HubSpot. The people who were on the call get attached to the right deal, too. Nobody typed a field, and the deal in HubSpot now says what the call said.
Where the usual answer goes wrong. Most advice says to switch on the CRM's built-in assistant, or to wire a recorder into fields with an automation tool. The built-in assistant sees only what someone typed, and a recorder wired into fields sees one call at a time. The conversations are the record; the CRM is where their results should land, with agents doing the updates and reps approving them. If leadership has ruled out a migration, that's a reasonable call, and you may already have a recorder and a few automations. That was the right start.
Keep, capture, act, write back, with Day AI
A working setup has the same four parts, whatever tools it uses. Use them as the plan, and as the checklist for any vendor that says it "adds AI to your CRM."
- Keep. Decide what stays in the CRM. It remains the record for pipeline stages, the forecast, reports and the integrations that read them. Nothing in this plan moves those. In Day AI, the CRM stays connected through its own MCP server (HubSpot, Salesforce), and agents read its records live.
- Capture. Put the conversations the CRM never held in one place: recorded calls with transcripts, email, calendar and the Slack channels you share with customers, each linked to the right person, company and deal. In Day AI, that means connecting Google Workspace, the notetaker or Gong or Granola, and the Slack app.
- Act. Give agents jobs that run on those conversations: a brief before every external meeting, a follow-up draft after it, a proposed CRM update, an answer when a rep asks about an account. In Day AI, each agent has a job and skills that run on a schedule, when something happens, or on request.
- Write back. Send the conversation-derived values to the CRM's fields through its MCP server, with rules for which fields write on their own and which wait for a person. In Day AI, you connect the HubSpot or Salesforce MCP, and the skill's instructions say which updates it writes and which it proposes for the rep to approve.
Where HubSpot or Salesforce ends and Day AI begins
This is the decision to make first, before choosing tools. A rule that holds up: the CRM keeps anything another system or report depends on, and Day AI owns anything that comes from a conversation.
| Stays in the CRM | Lives in Day AI | Written back |
|---|---|---|
| Pipeline stages and the forecast | Call transcripts and recordings | Next step and close date, proposed |
| Reports and dashboards | Email threads and calendar history | New contacts and their roles, proposed |
| Quotes, products and billing links | Slack conversations with customers | Call notes and summaries |
| Integrations that read CRM fields | Account research and meeting briefs | Tasks from what was agreed |
| Permissions and territories | Follow-up drafts | Competitors and risks mentioned |
The third column is the point. The layer doesn't become a second record people have to check. It makes the first one accurate.
Three ways to build the layer in Day AI
You can use all three, on the same memory and the same CRM connection.
1. Skills that run on a schedule or when something happens
In Day AI, a skill is a written job an agent runs on a schedule, when something happens (a recording is ready, an email lands) or on request. The post-call update and the meeting brief in the rollout below are skills. Every skill keeps its run history, and when you tell the agent what to change, later runs use the change.
2. Agents a rep asks in chat
Each rep has a Day AI agent with a name, a job and skills, and it reports to that rep. A rep can ask it about any account before a call ("what did they push back on last time?") and get an answer that cites the call clip or the message it came from. A rep can also ask it to update a deal; in chat, write tools ask first.
3. Day AI's MCP, from Claude, Claude Code or Cursor
Day AI exposes its customer memory over MCP with OAuth at https://day.ai/api/mcp, and day.ai/mcp walks through setup for Claude, Claude Code and Cursor. Builders keep their own tools and skills and work from the same calls, email and Slack threads as the rest of the team. Day AI's own marketing team runs its Claude Code repo this way: its skills live in the repo and read and write the same memory the rest of the company uses.
A 30-day rollout with Day AI
Day AI is the layer in this plan: each rep's agents work from every call, email and Slack thread and write back to HubSpot or Salesforce through the CRM's own MCP server.
The plan assumes a 60-person company on HubSpot with Google Workspace, a RevOps lead, and reps who record calls. It's an illustration, not a customer story. The same steps work on Salesforce.
Week 1: draw the line in Day AI and pick five fields
Fill in the table above for your own CRM: the middle column is what Day AI will hold. Then pick the five fields that are most often wrong and most often come up on calls. For most B2B teams that's next step, close date, stakeholders, competitor and the deal risk. These are what Day AI's post-call skill will propose in week 3. Leave everything else alone for now.
Week 2: connect the conversations to Day AI
In Day AI, connect Google Workspace for email and calendar, turn on the notetaker or connect Gong, and add the Day AI app to the Slack channels you share with customers. Gong brings in your past calls, so agents work from your history from the first day.
Week 3: give Day AI's agents two jobs, with rep approval
First, your HubSpot admin registers the OAuth app for Day AI, and each rep connects the HubSpot MCP in Day AI with their own login. Then turn on two skills reps feel immediately. A meeting-brief skill sends a brief before every external meeting, with what was said last time and what's still open. A post-call update skill runs when the recording is ready and proposes the five fields, with the source behind each change, for the rep to approve. The proposal arrives in Slack or email, and the rep approves it with a reply. Attach the HubSpot connector to that skill, with instructions to write nothing the rep hasn't approved. When reps ask their agent to update a deal in chat, leave the HubSpot write tools on Needs approval.
Week 4: let Day AI write the low-risk fields, and measure
Update the skill's instructions: write notes and tasks directly, and keep proposing stage, amount and close date for the rep to approve. Then describe a new property, such as "main competitor", and the agent backfills it from the conversations already in Day AI. Then ask your Day AI agent which open deals have a current next step, and which close dates moved because of something said on a call, and compare its answer with what you saw in week 1.
Where a homemade layer breaks
A homemade version of this layer works for one rep. These are the places it breaks as the rest of the team uses it.
Whose conversations an agent can see. One rep's inbox is easy. Ten reps, a manager and an agent that works for each of them is not. Someone has to decide which threads stay private, and the rule has to hold for every agent, not just the one you built first. In Day AI, each person sets sharing rules for their email, and when attendees' settings for a meeting conflict, the most private one wins.
Matching conversations to records. A call with "Sam" and a thread from a personal address both need to land on the right contact and the right one of two open deals. Wrong matches write wrong values, and reps stop trusting every field the layer touches. In Day AI, every call and email shows on the contact, company and deal it's about, and each proposed update cites its source, so a rep sees a wrong match before approving it.
Keys and permissions. A shared admin key in an automation tool lets every workflow write anything. Writing as each rep, with their CRM permissions, is safer and harder to set up. In Day AI, each rep connects the CRM with their own login, through an OAuth app your company registers, and agents act with that rep's permissions.
Skills that drift. The prompt that drafts follow-ups changes every time the pitch does. If every rep has a copy, nobody knows which version is running. In Day AI, an admin can publish a skill once as a Managed skill, and every rep's agent follows that source and picks up each edit.
Knowing why a value changed. When a close date moves, the rep needs to see who changed it, a person or an agent, and which call it came from. Without that history, nobody can check the layer or undo it. Day AI keeps that history on every value, including when and why it changed and the conversation it came from.
None of this means you can't build it. The first version takes weeks, not months. Most of the work starts when the second person relies on it. If you're weighing that up, see build vs buy for customer context.
What Day AI adds
Day AI provides the context graph and agent control plane to quickly and securely upgrade your company brain. Teams run it in place of their CRM, or on top of Salesforce or HubSpot. Day AI syncs with Salesforce or HubSpot through their MCP servers, which allow reads and writes when the write tools are turned on.
If you later decide to move off the CRM, the conversations and the agents stay where they are. For the bigger picture, see what comes after CRM. To connect Claude or another agent to the same memory, see how to give AI agents access to your CRM data.