AI-assisted enquiry processing

The problem

Enquiries from my Futuretrace website initially arrived in Outlook, where they could be missed or needed organising by hand. I wanted to manage them in HubSpot CRM, but when the same customer sent more than one enquiry, the new one could overwrite the previous information in the contact record.

The approach

I used Make to preserve each enquiry as its own note on the customer’s contact record, and to use AI to categorise and summarise it.

How it works

• A webhook receives the enquiry. The Make webhook is designed to receive data when someone submits a form, so the automation can run immediately rather than waiting to check for new submissions. I tested the workflow independently with fictional data sent to the webhook using cURL.
• The email address identifies the customer. HubSpot uses it to find or create the contact and returns a unique Contact ID. The automation uses that ID to attach the note to the right contact.
• AI reads the free text. I used the Make AI Toolkit’s “Extract information from text” module to return structured fields: category, summary, suggested service, missing information and next action. The enquiry is classified into one of four predefined categories, so later steps receive structured fields rather than a paragraph of generated text.
• A HubSpot private app creates the note. The automation retrieves the Contact ID, prepares the enquiry as JSON and makes an API call that creates a new note associated with that contact.

Decisions

• Why AI rather than rules. Rules-based automation follows predefined conditions, such as keywords. Customers describe what they need in different ways, so I used AI to interpret the meaning of free text.
• A human stays in the loop. The AI makes a recommendation; I review it and decide how to respond. AI classifications and recommendations aren’t always reliable.
• Least privilege. Make’s standard HubSpot integration asked for permissions I didn’t need (deals, tickets and files) when I tried to add notes. I used a private app with narrower permissions and an API request instead.

Testing

I tested the workflow with fictional enquiry data sent to the Make webhook using cURL. This allowed me to check the HubSpot contact creation, AI processing and creation of associated notes independently of the website.

I submitted two different enquiries using the same email address and confirmed that both were kept as separate notes under the same contact, without overwriting the earlier enquiry.

The tests demonstrated that the automation worked as a prototype. I did not complete the connection to the live Futuretrace website form, which continues to submit directly to HubSpot.

From prototype to production

Although building the automation in Make was relatively straightforward, connecting it to a public website introduced further considerations.

I would not want to expose the Make webhook directly in the website’s JavaScript, where it could be used to generate unwanted submissions and consume automation and AI resources. A production version would need a secure server-side connection, input validation and bot protection. I explored an approach using PHP and Cloudflare Turnstile, but decided not to deploy it at this stage.

Reliability would also need further work. If Make or HubSpot becomes unavailable, I would need a way to identify failed operations, recover enquiries and avoid creating duplicate notes when retrying submissions.

There are other areas I would want to evaluate before deployment:

• AI accuracy: test the classifications against a set of fictional enquiries with expected categories, and review the quality of summaries and suggested actions.
• Data protection: minimise the personal data sent for AI processing, review access permissions and retention, and update the Privacy Notice to reflect the additional services involved.
• Contact management: consider how to handle customers using different email addresses and enquiries from several people within the same organisation.

For now, the existing HubSpot integration meets the website’s immediate needs. I’ve therefore kept the Make workflow as a tested prototype rather than introducing additional complexity without a clear business benefit.

The project has helped me consider not only whether an automation can be built, but whether it is practical, secure and worthwhile to maintain.

Throughout, I treated AI as a tool to support decisions rather than replace them. I remain responsible for reviewing each enquiry and deciding how to respond.

← Back to Work