AI agents that qualify leads and update records in HubSpot and Salesforce

Yes. Both CRMs let approved apps read and write records through their APIs. HubSpot organises its CRM into objects such as contacts, companies, deals and tickets, and its MCP server gives AI tools read and write access to them. In Salesforce, the agent signs in through a connected app with its own user and permission set. We decide with you which objects and fields the agent may change. Everything else stays read-only or hidden.

A CRM AI agent is software that reads your CRM, decides on a next step and acts on it. Examples are scoring a new lead, moving a deal stage or drafting a follow-up email. An AI model such as Claude or GPT does the reading and writing. Rules you set decide what the agent may change alone and what a person must approve first.

Breeze and Agentforce are agents the CRM vendors sell inside their own products. They suit teams whose work stays inside one CRM. A custom agent fits when the job crosses systems, such as reading a quote in your ERP before updating a deal, or when you want your own approval rules and code you own. Our Agentforce pricing guide compares the two on cost and fit.

They do different jobs. An AI SDR works at the edge of the CRM: it writes to new prospects, replies to inbound leads fast and books meetings. A CRM agent works on the records your team already has, keeping them accurate and moving deals forward. Many sales teams run both, and they can share the same lead scoring rules.

Five jobs cover most of the requests we get. Qualifying inbound leads against your ideal customer rules. Updating deal stages when an email or call shows progress. Logging calls, meetings and emails to the right record. Enriching contacts and companies with missing fields. Drafting follow-ups for a rep to send. Each one can be built alone, or several can be combined into one agent.

Yes. When a form or chat creates a contact, the agent reads the form answers, the company website and any past activity. It scores the lead against rules you write in plain English, then sets the lead status and owner. It also adds a short note that explains the score, so a rep can read the reason before calling.

Only if you allow it. Most teams start with suggestions. The agent reads a reply or call notes, proposes a stage change with the reason, and the deal owner approves it in one click. Once suggestions prove right for a few weeks, you can let low-risk moves happen without approval, such as Discovery to Demo Booked. Closed Won stays with a person.

Yes. The agent matches each email thread, call summary or meeting note to the right contact, company and deal. It logs it as an activity with a short summary and next steps. If it cannot find a clear match, it does not guess. The item goes to a review list for a person to file.

The agent fills missing fields such as job title, company size, industry and website from sources you approve. Those can be the contact's own emails, the company website, or a data provider you already pay for, such as Apollo or ZoomInfo. It records where each value came from. It never overwrites a field a rep typed by hand unless your rules allow it.

Not by default. The agent drafts the follow-up inside the CRM, using the deal history and the last conversation, and a rep reviews and sends it. Some teams later allow automatic sends for narrow cases, such as a meeting confirmation. You set those rules in writing before launch, and every email the agent sends is logged on the record.

Yes. Some companies run marketing in HubSpot and sales in Salesforce. The agent can read a lead's marketing activity in HubSpot, then update the matching lead or opportunity in Salesforce, using one shared ID so the records stay matched. Before the build, we agree which system owns each field, so the two never overwrite each other.

Yes, and this is often the reason to build a custom agent. It can check open invoices in QuickBooks or NetSuite before a renewal call, read Zendesk tickets before a deal review, or attach a quote from your ERP to the opportunity. Each extra system gets its own permissions, read-only unless the job needs to write.

You decide per action. Common rules: drafts always need a rep to send them. Stage changes need the deal owner's approval above a deal value you choose. Empty fields may be filled without approval, but a field a person typed is never overwritten. Merges and deletes are not allowed. Every action is logged with the old value, the new value and the reason.

The agent gets its own user with the fewest permissions the job needs, so it cannot see or change objects outside its scope. API keys stay in your accounts. The AI model can run in your own Anthropic, OpenAI or cloud account, so records pass through a contract you control. HubSpot's own MCP server also leaves out properties marked as sensitive data.

We build on Claude, GPT or Gemini and pick per task. Tool use, where the model can only call functions we define, suits record updates because each action has a fixed shape that our code checks before anything reaches the CRM. The model sits behind our own code layer, so you can switch providers later without rebuilding the agent.

Yes. We build agents on the Anthropic API and use Claude's tool use, so the model can only call the CRM actions we define for it. FactoryJet is an independent services company, not part of Anthropic. If your company already has an Anthropic account, the agent can run in it, billed to you directly.

Every action is logged with the reason the agent gave. In the first weeks we review a sample with your sales or RevOps lead: how many suggestions were accepted, how many were edited and how many were wrong. Rules tighten or loosen based on that review, and the numbers come from your own CRM.

A single focused job, such as lead qualification in HubSpot, usually goes live one to three weeks after the scope is signed, the same range we give for our other AI integrations. An agent that covers several jobs or two CRMs takes longer. The written scope has dated milestones, and you see a demo inside your own CRM every week.

A short brief: which CRM and edition you use, the job you want done first, how many leads or deals move each week, and who approves the agent's actions. Test or sandbox access helps. If you already have an RFP or a requirements document, send it as it is. We answer the questions in it, so you do not have to fill in a form of ours.

Yes. Before you sign, we run the agent on a small set of your real records, with names removed if you prefer. You see each decision it would make and the reason for it, where it hands off to a person, and exactly which fields it would change in the CRM.

Both vendors version their APIs and announce changes ahead of time. The agent calls the CRM through one small layer of code, so a change is usually fixed in one place. Failed calls send an alert to Slack or email. If you choose support, we watch those notices and quote any needed change before doing it.

Market figures vary widely. Development firm ProductCrafters puts a custom AI agent build at $5,000 to more than $180,000 in its 2026 breakdown. Consultant Justin McKelvey puts a first done-for-you AI install for an owner-led business at $4,500 to $25,000, as of September 2026. FactoryJet quotes a fixed price in writing after a short scoping call. Both sources are linked below the questions.

Not from us unless you choose support. The AI model and hosting can run in your own accounts, billed to you at cost with no markup. Defects found after launch are fixed inside the build price for a cover period agreed before work starts. After that, each change is quoted and approved first, and there is no monthly minimum.

You do, from day one. The code lives in a repository you own. The HubSpot app or Salesforce connected app is registered in your account, and the prompts and approval rules are written down in plain English. If you bring the work in-house later, your team has everything it needs to run and change the agent.