Last checked: 12 Sep 2026. Prices are converted from US dollars at £1 = $1.35 on 11 Sep 2026; vendors bill in dollars and UK VAT applies on top. HubSpot credit prices are taken from HubSpot's UK price list.
Two buyers in the last eight weeks described the same problem in almost the same words. One asked whether an agent could "own a stage of my pipeline and know when its job is done". The other, describing automation already running in their CRM, said it moves records but "doesn't know its own goal", then put it more sharply: "does it know its exit point?"
That is a better test than most of the AI vocabulary in circulation, and it is the design question. Here is how we specify an agent that owns one pipeline stage, using the case we are asked for most often: chasing quotes that have gone quiet.
Why a rules-only workflow cannot own an outcome
A CRM workflow has an enrolment trigger and a sequence of steps. It does not have a goal. It cannot tell a quote that went quiet because the buyer is on annual leave from one that went quiet because they bought elsewhere last week, so it treats both identically and keeps sending.
Three consequences follow, and together they are what people mean when they say automation does not know its exit point:
- It cannot stop on success. The outcome it wants (a reply, a booked call, a signed quote) is not a condition it can evaluate, so it stops on step count instead.
- It cannot vary. Same email, same order, same channel, whatever the record says about the contact.
- It produces no evidence, only a list of sends. That makes it impossible to tune and impossible to defend in a review.
An agent is different only because you define it differently. You start from the exit condition and work backwards; everything else is a constraint on how it may get there.
The five things to specify before any code
One agent per pipeline stage, not one agent for the pipeline. A stage is small enough that its job fits in a sentence, which is what makes the exit condition measurable.
- Entry condition. The query deciding which records the agent owns, in CRM properties an administrator can read and change.
- Exit condition. A state the CRM can evaluate, not a sentiment: contact has replied, deal stage has moved, deal marked closed lost, three messages sent with no reply. If you cannot write it as a filter returning true or false, it is not an exit condition.
- Persisted context. The account's history in a form the agent can read cheaply, plus somewhere to write what it did and what it concluded. In practice that means new CRM properties, and this is the part people underestimate.
- Guardrails. What it may send, to whom, on which channels and how often, and what it must never do. Written as configuration the agent cannot override, not as prompt instructions.
- Hand-off trigger. The signals that stop the agent and put a task on a person, with the context attached.
The agent's contract
We write this table before anything is built, and it is what the client signs off. This is the version for a quote-chase agent on a stage called Quote Sent.
| Element | Specification for the quote-chase agent |
|---|---|
| Entry condition | Deal in Quote Sent, quote issued more than 30 days ago, no logged contact with the associated contact in the last 14 days, deal not marked closed. Both thresholds are CRM properties, adjustable by an administrator. |
| Inputs | Deal and quote line items, associated company and contact, channel preference, last five logged engagements, an overnight-generated account summary, a style sample of the rep's sent mail. |
| Allowed actions | Draft and send one follow-up in the owning rep's voice on the contact's preferred channel. Log the send. Set a next-review date. Write its reasoning to a note on the deal. |
| Prohibited actions | No pricing changes or discounting, no commitments on dates or scope, no contacting anyone not already associated with the deal, one message maximum in any seven-day window, three messages before hand-off. |
| Exit conditions | Contact replies on any channel; deal stage changes; deal marked closed won or closed lost; three messages sent with no reply; rep opts the deal out. |
| Escalation | Any reply, pricing or contractual question, complaint or unsubscribe signal, bounce on the primary channel, or exit on message count. Task raised on the owning rep with the thread and reasoning attached. |
| Logging | Every decision on the deal timeline: why the record was selected, which channel and why, what was sent, which exit condition fired. Enough to audit a month without opening the agent. |
Two details earn their keep. Some contacts will not answer a phone and will answer WhatsApp within the hour, so channel preference belongs on the contact record and the agent should respect it. And 30 days and 14 days are a starting point: a sales manager should be able to move them to 21 and 10 without raising a change request.
What it costs to run
Buyers ask what this costs month to month. The honest answer is a cost model rather than a number, because volume and route drive it. There are two routes.
Route one: HubSpot-native. HubSpot's agents product is Agent Hub, formerly Breeze Agents, and it runs on credits. HubSpot's UK list price is £9.00 per 1,000 credits monthly and £8.10 annually, so a credit is 0.9p or 0.81p. Included credits reset monthly and do not roll over. Rates below are from HubSpot's Product and Services Catalog, checked 12 Sep 2026.
| Item | Credits | Cost at £9.00 per 1,000 |
|---|---|---|
| Custom agent (beta) | 1 per action unit | 0.9p per action unit |
| Workflow AI action | 10 per AI action | 9p per AI action |
| Customer agent | 50 per resolved text conversation | 45p per resolution |
| Prospecting agent | 100 per lead recommendation | 90p per lead |
| Included per month | Starter 500, Professional 3,000, Enterprise 5,000 | No rollover |
Put 300 chased quotes a month through that. As a custom agent at six action units per chase, that is 1,800 credits, about £16.20 a month, inside a Professional allowance. Assembled instead from workflow AI actions at 10 credits each, two per chase is 6,000 credits, about £54 a month, above even the Enterprise allowance. Same outcome, three times the cost, purely because of how it was built.
Route two: a custom agent on a model API. Here you pay for tokens. A mid-priced frontier model currently sits at roughly £1.50 per million input tokens and £7.40 per million output tokens: Anthropic lists Sonnet 5 at that level on platform.claude.com and xAI lists Grok 4.6 at a similar input price with cheaper output on docs.x.ai, both checked 12 Sep 2026. Cheaper tiers cost several times less again.
Assume the agent reads a pre-summarised account brief, the deal and the last few engagements, around 5,000 input tokens, and writes a follow-up plus its reasoning, around 800 output tokens. That is about 1.3p per chase, so 300 chases a month is under £5 in model cost.
At this volume the model is not the cost. The build, the CRM property work, hosting and monitoring, and the sales time reviewing the first fortnight's output are all larger. The credit route has no build cost and a higher marginal cost; the custom route is the opposite. Below a few hundred records a month, the native route usually wins on total cost.
The pattern that changes the arithmetic
One design choice moves per-record cost by roughly an order of magnitude. If the agent re-reads the full account history at the moment it acts, it pays for tens of thousands of input tokens on every record, every run. If a separate overnight process summarises each account once and writes that summary to the record, the agent reads a few thousand tokens instead. On one build we scoped this year, that change cut cost per record by about a factor of ten, and quality improved, because a summary written without time pressure beats one produced mid-task.
Buyers raise the related point themselves: a messy contact database inflates the bill, because duplicates and generic mailboxes mean the agent processes rows it should never have touched. Cleansing and segmenting before the build is cost control, not housekeeping, and we price it as a separate workstream.
The same pattern on the website chatbot
A lead arriving through the website should get what a lead entering the pipeline manually gets: the same contract. An entry condition (a visitor on a commercial page starts a conversation), an exit condition (qualified and routed, meeting booked, or explicitly not a fit), the same CRM context, the same guardrails on what may be promised, the same hand-off to a person.
What changes is tone and length. A chat reply is two or three sentences and answers the question asked; a follow-up on a 30-day-old quote is longer and gives the reader a reason to reply. Same goal, same inputs, different register. This goes wrong when the chatbot is bolted on outside the CRM, which is how firms end up with a fourth disconnected system and a lead pool nobody owns.
When HubSpot's built-in agents are enough
Often. If the job is resolving inbound support questions from a knowledge base, the customer agent does it at 45p per resolved conversation, and a custom build has to beat that on quality and cost. If the job is surfacing lead recommendations, the prospecting agent does it at 90p a lead. Neither needs us.
A custom build is warranted when the trigger logic depends on your own properties and thresholds, when the agent must act across systems the built-in agents do not reach, when the output has to carry an individual rep's voice, when guardrails must be enforced in configuration because an auditor will ask, or when volume makes credit pricing worse than per-token pricing. Quote-chasing usually qualifies on the first and third. If you are unsure which side of the line you are on, count: how many records enter the stage in a month, and what one conversion is worth.
How we work on this
SpotDev is an AI and digital transformation consultancy that builds real software. We are a HubSpot Diamond Solutions Partner, Custom Integration Accredited, an OpenAI Select Partner and a Claude Registered Partner, with no house model: the choice follows the workload.
Related reading: how to spot stalled deals with AI, our AI for sales service, the wider AI implementation approach including the AI and Data Readiness Assessment, and the HubSpot development work the CRM and integration side of these builds sits on.
To scope one of these, Request a Quote.
Frequently asked questions
What is an exit condition, in practice?
A filter the CRM can evaluate as true or false. For a quote-chase agent: the contact has replied, the deal stage has changed, the deal is closed won or closed lost, or three messages have been sent with no reply. If it cannot be written as a filter, it is not an exit condition and the agent will not know when to stop.
Can a workflow do this instead of an agent?
A workflow can send the messages. It cannot judge whether the outcome has happened, vary the message by what the record says, or record why it did what it did. If a stage needs nothing more than a fixed sequence on a fixed schedule, use a workflow: it is cheaper and easier to support.
What does a quote-chase agent cost to run each month?
It depends on volume and route. At 300 chased quotes a month, a HubSpot custom agent at six action units per chase is about £16.20 at £9.00 per 1,000 credits, the same thing assembled from workflow AI actions at 10 credits each is about £54, and a custom build on a model API is under £5 in model cost. Build, hosting and review time are larger than all three.
Can the agent choose between email, WhatsApp and phone?
Yes, if channel preference is stored on the contact record. The agent reads it and sends on that channel within its guardrails. WhatsApp tends to outperform SMS for UK and international contacts. Voice is possible but is a vendor-dependent capability, so it needs testing on the specific workload before it is promised.
Are HubSpot's built-in agents enough?
For knowledge-base support resolution and lead recommendations, usually yes. Agent Hub, formerly Breeze Agents, charges 50 credits per resolved text conversation and 100 credits per lead recommendation. A custom build is warranted when the trigger logic depends on your own properties, when the output must carry a named rep's voice, when guardrails must be enforced in configuration for an auditor, or when volume makes credit pricing worse than per-token pricing.
Can the thresholds be changed without a development request?
They should be. The quote age and inactivity window belong in CRM properties an administrator can edit, not in code. If changing 30 days to 21 days needs a developer, the agent has been built wrong.
Stay Updated with Our Latest Insights
Get expert HubSpot tips and integration strategies delivered to your inbox.
