Build an AI SDR on Top of HubSpot: Complete Architecture Guide
The architecture behind an AI SDR on top of HubSpot: data, prioritisation, contact, write-back and measurement, plus common mistakes and what the human keeps.
Table of contents
How do you build an AI SDR on top of HubSpot?
An AI SDR is not one product you switch on. It is a collection of tasks that an SDR (sales development rep, the person who finds and qualifies leads and books meetings) normally carries out, where the AI takes over the parts that are routine: preparing a lead list, working out who should be contacted first, writing a first draft and putting the result back into HubSpot. The human keeps what requires judgement, and everything that goes out to a customer.
That is also the order the architecture should be built in: first data, then prioritisation, then contact, then write-back, then measurement. Most projects that go wrong skip the first step.
TL;DR:
- An AI SDR on HubSpot consists of five layers: data foundation, prioritisation, contact, write-back to the CRM and measurement.
- Data quality in HubSpot determines the result. Do not call anyone until the list has been cleaned.
- Let a human send and call. Where an AI itself may speak to a customer, the rules are not clear enough to promise anything without a legal assessment.
What is an AI SDR, and what is its task?
An SDR's job is to create conversations that a salesperson can close. Typically this means finding the right companies, contacting the right people, qualifying interest and booking a meeting. It is a job with many repetitions and many small decisions, which is why it is an obvious candidate for AI to take over parts of it.
But "AI SDR" is used about very different things. Some people mean a tool that writes emails. Others mean a voice that calls. It is worth distinguishing between three levels:
- Preparatory: The AI gathers information, prioritises and writes drafts. A human makes the contact.
- Assisting: The AI suggests next steps and writes back to the CRM, while the human uses it in their dialogue.
- Self-running: The AI contacts the customer itself. This is the most debated level, both technically and legally.
We recommend starting at the first two levels. They deliver most of the time savings and carry far less risk.
What is the architecture behind an AI SDR on HubSpot?
The architecture consists of five layers. Each layer has one purpose, and each layer can be tested on its own.
Five layers in an AI SDR built on top of HubSpot
| Layer | Purpose | What can go wrong | Who has the final say |
|---|---|---|---|
| 1. Data foundation in HubSpot | Clean, consistent contacts and companies | The AI contacts the wrong people or those who have opted out of contact | You, before you start |
| 2. Prioritisation | Who should be contacted first, and why | Scoring based on random signals | The sales manager, who sets the rules |
| 3. Contact | Draft email or call preparation | Wrong claims, wrong tone | The human who sends |
| 4. Write-back | Results are written to HubSpot | Noise in the CRM, duplicates, overwritten fields | The rules for what may be written |
| 5. Measurement | What works, and what should change | Measuring activity instead of meetings | The sales manager |
Layer 1: HubSpot as the data foundation. What needs to be in place before you start?
An AI SDR is only as good as the data it works from. Go through these points before you set anything up.
Data quality. Contacts should as a minimum have a name, a company, a title and a usable contact point. Remove duplicates (HubSpot has built-in tools for handling them). Clean up contacts who have left their jobs, and companies that have closed or been acquired. If the list is full of old contacts, an AI will just make the mistakes happen faster.
Source and consent. Every contact should have a recorded source, so you know where the information came from and what basis you have for contacting the person. This applies under both GDPR and the Danish Marketing Practices Act. Create a marker for those who have opted out of contact, and let the AI layer read it before any draft is written.
Common definitions. What is a lead, an MQL, an SQL? When does a deal move to a new stage? If two salespeople use the fields differently, no automation can use them.
Scope. Decide precisely which contacts the AI layer may work with. This is best done with a saved list or a filter, for example contacts in a particular stage with no activity for a period, in a particular company size, with no open agreements. The filter must be visible and changeable by the sales manager.
Access. For the AI layer to read and write in HubSpot, you typically create a so-called private app or an OAuth integration. Give it the fewest possible rights: read contacts and companies, write notes and tasks. Store the key securely, and agree who can revoke it.
Which automation features you have access to in HubSpot depends on your subscription. Check your own account before you plan, and ask HubSpot if you are in doubt.
Layer 2: Prioritisation. Who should be contacted first, and when?
Most people have more leads than time to work them. A prioritisation layer helps choose the order. There are two kinds of signals.
Fit: Does the company match the customers you usually succeed with? Size, industry, geography and possibly technology.
Interest and timing: Has the contact opened emails, visited pages, replied earlier or been in dialogue before? Has something happened at the company that makes the timing right?
Keep it simple at the start. A score with a few understandable rules, which the sales team can explain, is better than a complicated model no one can question. Let each prioritisation come with a short reason, so the salesperson can see why a contact is high on the list, and can disagree when the reason is weak.
Layer 3: Contact. What may the AI do, and what does the human do?
This is the most important decision in the whole project, and it should be taken deliberately.
Email drafts and call preparation. The AI can write a draft based on what is in HubSpot and on publicly available information. The draft must mark what has been taken from where. An SDR reads it through, corrects it and sends it. The quality is then a human's responsibility, and the tone can resemble yours.
Phone. The salesperson can get a short overview before the call: recent activity, what was said last time, and what is relevant to ask about. After the call, the AI can write the note.
AI that calls by itself. It is technically possible to let a voice agent contact people. But the rules are not clear enough for us to promise anything. The Marketing Practices Act sets requirements for automatic calling systems and for calls to consumers, GDPR requires information and a valid basis, and there is no established practice for how fully automatic AI conversations should be treated. If you want to explore that route, bring in a lawyer from the start and first choose a small, delimited trial. We work with JesperAI on voice agents, but we promise nothing about what is lawful for you; that must be assessed case by case.
Also be careful about volume. One error in a draft sent to a thousand people is a thousand errors.
Layer 4: Write-back. How do you write back to HubSpot without noise?
Many people forget this layer. The result is a CRM full of activities nobody understands, and a sales team that stops trusting the system.
Good rules for write-back:
- Write as notes and activities first. Let the AI write a note with source and time before it is allowed to change any field at all.
- Change only the fields you have given it permission to change. Price, stage and probability should require approval.
- Mark what was written by the AI. Then you can always see the difference, and you can correct systematically if something proves to be wrong.
- Avoid duplicates. Look up before creating. Write to existing contacts instead of creating new ones.
- Build a way back. If a batch goes wrong, can you undo it? If not, limit how much is written at a time.
When a meeting is booked, HubSpot should be able to show it in the same way as when a salesperson has done it themselves: a meeting activity, a task for the person who is to hold the meeting, and an updated stage if your rules say so.
Layer 5: Measurement. What works, and what should change?
Measure what matters to the business, not just activity. The number of emails sent says little. It says more how many replied, how many meetings were booked, how many attended, and how many became an agreement. Follow the whole path, and compare with your figures from before.
Also be honest about what cannot be measured. A draft can be so good that a meeting ends in an agreement, or so bad that the customer never replies again. Quality is seen in the replies and in the salespeople's experience, not only in the numbers.
There are no universally valid figures for how many meetings an AI SDR books. Such figures depend on industry, target group, offer and the quality of the list. Be sceptical of anyone who mentions a single number without knowing your situation.
What does it cost to build?
There are two routes. The first is to build it yourself, with your own IT department or an internal resource, possibly with a tool such as n8n or Make to connect things together. It is cheap to get started, but remember to budget time for maintenance: APIs change, models change, and someone has to watch for errors.
The second is to have a partner build it on top of HubSpot. At Agent360 we build AI solutions in the customer's own systems, and the price depends on the task: how many systems must be connected, how much approval logic must go in, and how much data must be cleaned up. There is no price list. We give an estimate after a short conversation.
The most relevant example on the use cases page is "The meeting is over. The CRM work begins", where the agent writes notes and updates CRM fields (HubSpot, Salesforce, Pipedrive or Microsoft Dynamics), but asks first before anything goes out to the customer.
What are the most common mistakes?
Mistake 1: The data quality is not in place. Companies jump straight to the contact layer and only discover afterwards that the list was out of date. The AI then contacts the wrong people, or the same people twice.
Mistake 2: The write-back is missing or messy. An AI tool that sends emails but does not write back properly leaves the CRM in disorder. After a few months the team loses trust.
Mistake 3: Scaling too fast. The volume is opened before the drafts have been tested. Start small, read the results, correct, and then scale.
Mistake 4: No human reads. Drafts that are sent without review are cheap to produce and expensive to undo.
FAQ: Build an AI SDR on top of HubSpot
Q: Does it only work with HubSpot Sales Professional, or can I use Starter? It depends on what you want to automate. The connection to HubSpot via the API is not limited to the most expensive subscriptions, but features such as advanced workflows and code actions may require a higher subscription. Check your account, and ask HubSpot if you are in doubt.
Q: Can the AI call in Danish, and does it sound natural? The technology has become markedly better, but quality varies between suppliers and with background noise, dialects and technical terms. Ask every supplier to show the result on your own examples. See also our guide to testing an AI language model on Danish sales.
Q: Are AI calls lawful in Denmark under GDPR and the Marketing Practices Act? We cannot answer that unambiguously. It depends on who is called (consumer or business), whether consent has been given, whether the call is fully automatic, and what is said. We recommend that a human makes the first calls, and that you have a lawyer assess your specific setup. Read more in our guide to lawful AI calls in Denmark.
Q: How long does it take to set up an AI SDR? It depends on how tidy your data is and how much needs to be built. Clean up first, and you will build faster. We build one layer at a time and finish the first one before we move on.
Q: What happens in HubSpot when a meeting is booked? Your rules decide that. Typically a meeting activity is created, a task for the person who is to hold the meeting, and possibly a changed stage on the contact or the agreement. Agree in advance which fields the AI may touch.
Q: Can it be connected to CRM systems other than HubSpot? Yes. We also build on top of Salesforce, Pipedrive and Microsoft Dynamics.
Q: What is a realistic expectation for meeting bookings? Nobody can say that without knowing your market, offer and lists. Measure your current figures first, and compare afterwards.
Conclusion: The architecture is the foundation
The order is the most important thing: first data, then prioritisation, then contact, then write-back, then measurement. Let humans send and call, and let the AI prepare and write back. Start small, and finish the first layer before you move on.
Book 20 minutes of free advice if you want to hear whether your sales task is worth building on top of HubSpot. We will also tell you if it is not.
Sources
- HubSpot, Developer documentation (CRM API and private apps)
- The Danish Data Protection Agency (Datatilsynet), rules and guidance
- The Danish Consumer Ombudsman, Marketing Practices Act
#hubspot#ai-sdr#architecture#crm#sales
Contents
Should your CRM keep itself up to date?
We build the layer that moves data into the CRM you already use, without anyone having to type it in. We look at your current set-up first.
Book a reviewShould your CRM keep itself up to date?
We build the layer that moves data into the CRM you already use, without anyone having to type it in. We look at your current set-up first.
How we built it · CRM & customer overview
Customer CRM that updates itself
An internal page per customer with contacts, the agreement, the projects and a timeline where emails, meetings, email drafts and tasks land by themselves. Above the pages sits the board with all customers, owner and latest contact, and the system suggests next steps based on fixed rules.
- Email, meeting, task
- Lands on the customer’s page
- You take the next step
Related articles

Sales Ops Automation in Python: Automating What No-Code Cannot
When is no-code not enough for sales ops? Three Python automation examples (bulk sync, scoring, reports), pitfalls, requirements and who should own them.

Headless Sales AI: What Happens When You Separate Intelligence from the Interface
What is headless sales AI? Separate the AI logic from the user interface so it works in CRM, Teams and email. Benefits, risks and when it is worth doing.

Webhooks for sales teams: how to build a real-time data pipeline
Webhooks for sales teams: what they are, webhooks versus polling, the most useful triggers, the five typical mistakes, and how to start without overbuilding.