Custom AI workflows that do the reading and typing your team does by hand
Setting up HubSpot gets you the tools HubSpot already makes. Some jobs need a tool HubSpot doesn't make, like reading the PDFs customers send, checking order history in another system or applying your own pricing rules. We build those pieces, plug them into HubSpot and keep a person approving anything that matters.
Free, 30 minutes, and you'll leave with at least one clear next step.
What is a custom AI workflow?
It's a job your team does over and over, handed to software that can read and write like a person. An email comes in with a request attached. Today someone opens the PDF, types the details into HubSpot, looks up the customer and forwards it to the right person. A custom AI workflow does those steps in seconds and puts the result in front of that person to check.
Our HubSpot feature work, like workflows or lead scoring, turns on and configures tools HubSpot already built. A custom AI workflow is new software we write for one job in your business. It reads the documents and data HubSpot can't, then hands the result to HubSpot so your existing setup can act on it.
When HubSpot's built-in AI is enough, and when it isn't
HubSpot's built-in AI (Breeze) is enough when
- The job matches an agent HubSpot already offers, like answering support chats from your help articles or researching companies
- The data the AI needs is already in HubSpot or on the public web
- HubSpot's per-result credit charges still make sense at your volume
- You're fine with HubSpot's settings and how often it changes them
- Agent Builder, HubSpot's tool for making your own agents, can handle the steps
Custom work makes sense when
- The input is a document, such as an order form, contract, application or quote request
- The AI needs data HubSpot can't reach well, such as order history in your accounting system, inventory or a price list
- Your business rules must wrap around the AI, such as territories, approval limits or minimum order amounts
- At your volume, per-result credits would cost more than paying an AI provider directly
- The job takes more than a few seconds, or has to work through hundreds of records in a batch
- Predictions need your own history, such as when a customer usually reorders
Do you have to pick one?
Often the best answer is both. Turn on HubSpot's agents for the jobs they already do well, and build custom pieces for the document-heavy work and the outside data that make your business different. The custom pieces write their results into HubSpot fields, so HubSpot's agents, workflows, lead scores and reports can all use them.
Custom AI agents and workflows we build
"New requests sit in an inbox until someone has time"
The AI reads each request from email, form or PDF. It pulls out the key details, like what the customer wants, how many and by when. Then it creates the record in HubSpot and routes it to the right person. Unclear ones go to a review list.
"Customers send PDFs and we retype every line"
A purchase order, intake form or signed agreement becomes HubSpot records and line items. Each line is matched to your HubSpot product library. Lines it can't match get flagged for review.
"Our reps never fill in the call notes"
After a call, the AI summarizes it and fills in fields like competitor mentioned, timeline, decision maker and next step. HubSpot does part of this on its own. We build it for very specific fields or other call sources.
"We find out a customer stopped buying months too late"
We pull order history from QuickBooks, NetSuite, Shopify or another system and work out how often each account usually buys. When an account runs late, HubSpot creates a task for the rep.
"Follow-ups take forever to write"
The AI drafts a follow-up from the deal, the quote status and the last call. The draft lands with the rep to approve, edit or delete. Nothing sends without a person.
"Only two people know the answers by heart"
An internal assistant answers your team's questions from your own manuals, policies and product sheets, and shows which document it used. It's for your team, so private documents stay private.
"New leads wait days for research"
When a lead comes in, the AI looks up the company, checks whether it's already a customer in HubSpot and writes a short summary for the rep. The rep starts the first call knowing who they're talking to.
How we build them
We keep the parts simple and use HubSpot's own connections wherever we can.
Inside HubSpot. For short steps, we add custom code to a HubSpot workflow that asks an AI model a question and writes the answer to a field. This needs Data Hub Professional or Enterprise, and each step has a 20-second limit.
Outside HubSpot, for bigger jobs. When a job takes longer, like reading a long contract or a stack of documents, the work runs on a small service we host. HubSpot can send it the record from a workflow, which also needs Data Hub Professional or Enterprise, or the service can watch an inbox or run on a schedule. It does the work, tries again if something fails, and writes the result back.
The AI models. We use Claude from Anthropic or OpenAI's models to read, sort, summarize and draft. Answers come back in a fixed format, so each one drops into the right HubSpot field.
The HubSpot MCP server. This HubSpot connection lets AI tools like Claude or ChatGPT read and update CRM records. It signs in as a named HubSpot user and can only do what that user is allowed to do. It suits assistant-style work and quick trial versions. See our MCP server page.
Human approval and data safety. Anything that goes to a customer, changes a price or moves money waits for a person. We send only the fields a job needs, keep keys in secure storage and log every action, so you can see what the AI did and why.
We start with one workflow and prove it on your records
AI projects go wrong when they try to do everything at once. We pick one job, build it, test it on your records and measure it before adding the next.
Stage 1: Find the jobs worth automating
We look at where your team reads, retypes or looks things up by hand, and how clean the HubSpot data is.
We watch sales, service and operations at work, like a request turning into a quote, and note where someone copies data or waits for an answer. We also check the HubSpot fields the AI would read and write, because bad data leads to confident wrong answers.
You get A short list of jobs that could be automated, with the data gaps marked.
Stage 2: Pick one workflow to start
Together we choose the job with the clearest payoff and the lowest risk if the AI gets something wrong.
We weigh how often the job happens, how long it takes and what a mistake would cost. We then write down what goes in, what comes out, where it lands in HubSpot and who approves it, plus how we'll measure success, such as time from request to first reply.
You get A one-page plan for the first workflow, with a clear way to measure it.
Stage 3: Build it on your systems
We connect HubSpot, the AI model and any other system involved, with approval steps where you want them.
We set up the trigger in HubSpot, the AI step, the business rules around it and where the result lands. We pick the model that handles your documents best, and log every action so any answer can be traced to its source.
You get A working workflow in a test setting.
Stage 4: Test with past records
We run it on a batch of your past emails, PDFs or deals and compare its answers with what your team did.
We use your own examples, including the messy ones, and check every answer. Where the AI is wrong, we fix the instructions, add a rule or send that kind of case to a person. Your team signs off before it goes live.
You get A test report showing what the AI got right, what it got wrong and what we changed.
Stage 5: Launch and measure
We turn it on for one team, track the agreed measure, and decide with you what to build next.
We start with one team and watch the review list closely for the first weeks, then compare results with how the job worked before. Once it's proven, you decide whether to give it more room, add another workflow or stop there.
You get A live workflow, a before-and-after measure and a plan for what's next.
What we've connected to HubSpot before
Custom AI only works if the systems underneath talk to each other. We're a HubSpot Platinum Solutions Partner, and much of our work is connecting HubSpot to other software. For Photon Brothers, we connected HubSpot to its design, inventory and finance systems. That saves about 10 hours a week of manual data entry. It wasn't an AI project, but every AI workflow sits on the same kind of foundation.
For AI search and HubSpot's own agents, see our AI agents service.
Custom AI workflows questions
Is our customer data safe with an AI model?
We use business accounts with the AI provider and go over their data terms with you before anything is sent, including how long data is kept and whether it can be used for training. We send only the fields each job needs and log every action.
Which AI model do you use?
Usually Claude from Anthropic or a model from OpenAI. We test on your documents and pick the one that does the job best. Because the model sits behind our own code, we can switch later if a better option comes along.
What happens when the AI gets something wrong?
It will sometimes. We plan for it. Unclear cases go to a review list. Anything sent to a customer or changing a price waits for a person to approve. Every action is logged, so mistakes are easy to find and fix, and we use them to improve the instructions.
Who maintains it after launch?
AI models and HubSpot both change, so a workflow needs some care. We can look after it month to month, or hand it to your team with written notes on how it works and what to watch.
What drives the cost of a custom AI workflow?
The number of systems involved, how messy the documents are, how many records it handles, and how much testing it needs. You also pay the AI provider for usage, and some setups need Data Hub Professional or Enterprise. We scope each project after the first call.
Do we need Data Hub for this?
Usually, if the job starts from a HubSpot workflow. Custom code steps and sending records from a workflow to an outside service both need Data Hub Professional or Enterprise. Some jobs start another way, such as from an email inbox or a nightly schedule, and don't need it. We check your plan before we design anything.
Related
More in AI services
Tell us the job your team hates doing
Describe one task your team does by hand every day. We'll tell you whether HubSpot's built-in AI can handle it, or whether a custom workflow makes sense.