Every week someone asks me a version of the same question: "Should this be an AI agent?"
Usually the honest answer is no. Sometimes it's yes. Most of the time it's "a bit of both", and the real work is deciding where one stops and the other starts.
HubSpot now gives you plenty of options. There are classic workflows, Breeze AI actions you can drop into a workflow, and Breeze agents that can read context and decide what to do next. Choosing well matters more than choosing the newest thing. Here's the test I use with clients.
Use a workflow when you can write the rule down. If the process sounds like "when this happens, do that", it belongs in a workflow. Assigning an enquiry by postcode. Creating a follow-up task three days after a quote goes out. Sending a booking reminder. Escalating a ticket that's been open too long.
Workflows are predictable and testable. When something goes wrong, you can see exactly which branch fired and why. That's worth a lot.
Use a bounded agent when the context changes the answer. Some jobs need judgement. Reading an email and working out what the person actually wants. Summarising a messy record. Choosing between three sensible next steps. That's where an agent earns its place.
The word "bounded" is doing real work there. A bounded agent has one goal, approved data, a short list of actions it's allowed to take, clear rules for when to stop and hand over, and a named person who owns it. It doesn't get access to everything just because the platform allows it.
If yes, use a workflow. If the right action depends on context you can't express as a few clean branches, consider an agent.
A wrongly tagged enquiry is annoying. A wrong quote, a privacy slip or a message sent to the wrong person is a real problem. The higher the cost of an error, the more you want a person approving and a workflow doing the final step.
An agent can't fix bad data. If you've got duplicate contacts, empty fields and three versions of the truth, the agent will reason confidently from rubbish. Clean up the record first. I've written about what happened when we deleted 300 properties from a client's HubSpot, and the same lesson applies here.
Drafting, tagging, summarising and creating a task are easy to reverse. Sending, deleting, charging and committing are hard to reverse. Start agents on the reversible work.
Every automation hits cases it can't handle: missing information, low confidence, systems that disagree, or a customer who just wants a human. Decide where those go before you launch. A named team, a visible queue and an agreed response time.
For most real processes the answer is a combination. It looks like this:
This keeps the dependable parts dependable. AI goes where judgement adds value, and nowhere else.
A signed agreement triggers a workflow that creates an onboarding ticket. An agent reads the approved documents and drafts a list of what's still missing. The account manager checks the list and approves the request before it goes to the client.
A workflow opens a ticket from a service enquiry. An agent uses your product knowledge and the customer's service history to suggest a category and next step. A coordinator approves anything non-routine, and anything that affects price or scheduling.
A workflow records an intake form and checks the required documents are there. An agent organises the administrative information into a summary. A clinician makes every clinical decision. The workflow records what was approved and routes the next administrative step.
Most of what my clients need is still good workflows. The five workflows that replace 20 hours of manual work a week don't need AI at all. Agents are a great addition when you put them where judgement matters and keep a person on the decisions that count. It's also how we run AI inside DigiKat, with clear human approval rules for everything our own AI does.
If you're not sure whether your next automation should be a workflow, an agent or both, have a look at how we build AI agents in HubSpot and pick one real process for us to map with you.