Service response
Retrieves approved knowledge and account context, then answers, requests what is missing, or routes to the right person.
The fear is not that AI answers slowly. It is that it answers wrongly, to a customer, in writing. So we start by deciding what it is allowed to say.
Four things we find in almost every service team that has tried a bolt-on chatbot.
Skilled people answering what a published article already answers, while real problems queue behind them.
Routine questions resolved or drafted, so complex work reaches a person sooner.
Order status, contract terms and delivery dates sit in the ERP. The bot sees the help centre.
Answers grounded in the systems that actually hold the fact, not just the knowledge base.
The signals were there for months, spread across tickets, usage and unanswered emails.
Conditions detected against CRM evidence, with the reasoning shown so a human can judge it.
So it is either switched off, or switched on and quietly frightening.
Granted one action at a time, with everything sensitive escalated from day one.
Skilled people answering what a published article already answers, while real problems queue behind them.
Routine questions resolved or drafted, so complex work reaches a person sooner.
Order status, contract terms and delivery dates sit in the ERP. The bot sees the help centre.
Answers grounded in the systems that actually hold the fact, not just the knowledge base.
The signals were there for months, spread across tickets, usage and unanswered emails.
Conditions detected against CRM evidence, with the reasoning shown so a human can judge it.
So it is either switched off, or switched on and quietly frightening.
Granted one action at a time, with everything sensitive escalated from day one.
Swipe to explore
A wave activates two of the four agent capabilities, chosen against your own value case. Enablement and measurement are part of every wave, not alternatives to them.
Retrieves approved knowledge and account context, then answers, requests what is missing, or routes to the right person.
Assembles outcomes, risks, open requests and commitments before the review, so the meeting starts with the facts already gathered.
Detects inactivity, unresolved service and adoption decline, explains the evidence, and recommends an intervention for a person to take.
Connects a request or a changed condition to a credible expansion hypothesis, and prepares the internal briefing behind it.
The people answering customers trained on when to trust a draft, when to correct it, and how to report one that was wrong.
Response time, resolution time, escalation and correction rate, tracked against the position we recorded before anything was built.
HubSpot's own customer agent answers from your published content. It does not know an order status held in your ERP.
Content pulled from outside the knowledge base re-syncs on a schedule, so pricing and lead times can lag reality.
You get the shape of approvals the product ships with, rather than the one your escalation policy actually needs.
Review preparation, retention and growth are not ticket deflection, and no ticket-deflection tool is built for them.
Authority is granted per action, not per agent. Check what your own subscription covers, then decide each of these deliberately.
Refunds, complaints, legal matters, account closure and anything about a vulnerable customer.
Permission to draft is not permission to send. The two are never granted together by default.
A model's confidence score is not evidence. We gate on source and action type instead.
When identity, entitlement or the source is uncertain, it stops and hands over rather than guessing.
For many customer-facing workflows, drafting for a human is the right end state, not a phase.
Every answer, escalation and correction recorded, so a complaint can be reconstructed from evidence.
An in-house engineering team that has delivered more than 300 technology projects. Work with engineers who have built this before.
Agents and integrations designed, tested and shipped by our own engineers. Nothing is subcontracted.
SpotDev is enrolled in the OpenAI Partner Network at Select tier, alongside HubSpot Diamond status and Custom Integration and Onboarding accreditations.
Honest evaluation of the right model for the work. We hold OpenAI Select Partner status and deep expertise in Anthropic's Claude, and we will say when the answer is the AI you already pay for.
It should mostly not be in a position to. Anything consequential is drafted for a person. Where it does answer directly, errors are logged, reviewed and fed back into the evaluation set.
We recommend telling them, because trust is easier to keep than rebuild. That is a design choice we advise, not a legal opinion, and your own advisers should confirm your obligations.
OpenAI states on its enterprise privacy page, updated 8 January 2026, that it does not train on business data by default unless you explicitly opt in. We configure for that and verify it.
Often, yes, and we will say so. It is enough when the answers genuinely live in your published content. It stops when the answer lives in another system.
No. That is the argument of this page. Everything else in this market sells you a chatbot platform to run alongside your CRM, which is how you end up with two versions of the customer.
It changes the first job. Grounding an agent in stale content produces confident wrong answers, so content quality gets assessed before anything faces a customer.
From £10,000 per month as part of AI Accelerator, committed a quarter at a time. Model and platform usage is billed separately. Enterprise programmes are priced on complexity.
Related: AI Accelerator · AI for Sales · AI implementation
SpotDev is an OpenAI Select Partner and a HubSpot Diamond Solutions Partner. SpotDev is a separate company from OpenAI, and partner status does not constitute endorsement of SpotDev by OpenAI. OpenAI and ChatGPT are trademarks of OpenAI. HubSpot is a trademark of HubSpot, Inc. Claude is a trademark of Anthropic, PBC.