Last checked 12 Sep 2026. HubSpot credit prices are taken from HubSpot's UK price list, which is quoted in pounds, so no currency conversion is involved. UK VAT applies on top.
Most mid-market buyers of AI work are not short of advice. They are short of running systems. The pattern repeats: a discovery phase, prioritised use cases, a slide scoring each on value and feasibility, a pilot, then nothing a customer or a colleague ever touches. The money is spent and the process is unchanged.
This post sets out a procurement test you can run in a first meeting, without technical knowledge. The questions that separate a build firm from a deck firm are all questions a build firm can answer on the spot.
Why the strategy-only pattern fails a mid-market buyer
Two failure modes account for most of it, and we see both in live conversations with UK companies.
The pilot that never reaches production. A proof of concept demonstrates that a model can do something. Production needs integration with the systems of record, error handling, access control, monitoring, a named owner, and a way to change behaviour without a change request. None of that makes a good slide, and all of it is where the cost and the risk sit. A firm whose delivery model is analysis hands the production problem back to you, usually once your internal sponsor has already declared the pilot a success.
The data problem nobody priced. Across our own client conversations this summer, the most repeated theme among existing engagements was not model choice. It was that inconsistent contact and company records made any AI layer untrustworthy before it was built. In several cases the correct sequence was to stop the AI work, run a data cleanup and mapping pass, then restart. That is an unwelcome answer, and it is the one a firm that has to make the system work afterwards will give you.
Data quality is also the cost lever. In one engagement, moving from an agent that re-read a full account history on every run to a process that pre-summarised that history overnight reduced the cost per record by roughly a factor of ten. That is an engineering decision, made after the system is live and measured. It does not appear in a roadmap.
The procurement test: seven questions for the first meeting
Put the same seven questions to every firm on your shortlist, and write down the answers.
1. Who writes the code?
Name the people, not the firm. Are they employed by the company in front of you, or subcontracted to a partner you have not met? A good answer names individuals and says what each has shipped. A weak answer describes a "delivery network" or moves straight to methodology.
2. What percentage of the people on this engagement are engineers?
Ask for the split between engineering, delivery management and strategy on the specific team you would get, not the firm overall. Firms that build tend to answer with a number immediately. If the proposed team is mostly consultants and project managers with engineering "available as needed", you are buying analysis with a build option attached.
3. Show me a system in production, and the person who built it.
Not a case study PDF. A system that real users log into, and the engineer who wrote it, on the call. A firm that ships can arrange this within a week. This question removes more shortlisted firms than any other.
4. What is the definition of done, and what is the acceptance test?
The answer should be checkable: this agent handles this stage of this pipeline, exits when this condition is met, escalates to a person on these triggers, and is accepted when it processes an agreed sample at an agreed accuracy. Buyers put it in their own words: does it know its goal, and does it know its exit point. If the answer is a list of workshops and deliverables, there is no acceptance test and no moment at which the work is finished.
5. Who owns the intellectual property and the repository?
You should own the code and hold repository access from day one, not at the end of the engagement and not on request. Ask where it is hosted, who has commit access, and what happens to that access if you stop working together. Ask the same about prompts and configuration, which are part of the system even when they are not code.
6. What happens in month two?
AI systems drift. Vendors change models, prices and retention terms with little warning, and a process that worked in week one degrades quietly. Ask what is monitored, who looks at it, how a change is requested, what it costs, and how quickly a model can be swapped if a vendor term becomes unacceptable. A firm with no answer here has priced a launch, not a system.
7. How is the running cost measured, per record or per conversation?
This is the question that separates outcome pricing from hand-waving. The firm should be able to express the cost of the system in the unit your business already counts: per resolved enquiry, per chased quote, per enriched record.
HubSpot's own credit pricing is a useful worked illustration, because the rates are published. HubSpot sells credits at £9.00 per 1,000 paying monthly, or £8.10 per 1,000 paying annually, on its UK price list. Its rate sheet charges 50 credits for a customer agent to resolve one text conversation, 100 credits for a prospecting agent lead recommendation, and 10 credits for a data agent prompt against one record. So a resolved conversation costs about 45p monthly or 40.5p annually, a lead recommendation about 90p or 81p, and a data agent prompt about 9p or 8.1p. Included allowances are 500 credits a month on Starter, 3,000 on Professional and 5,000 on Enterprise, they do not roll over, and account-level and feature-level caps can be set.
Whether or not you use HubSpot's agents, that is the shape of the answer you want: a unit, a rate, a monthly ceiling and a cap. A firm that has run systems in production will give you one for a custom build too, because it has had to reconcile the invoice.
Three engagement shapes, and when each is right
The test above is not an argument that strategy work is worthless. It is an argument for buying the right shape. Costs below are relative, because the honest figure depends on scope.
| Engagement shape | What you get | Relative cost | When it is the right purchase |
|---|---|---|---|
| Strategy engagement | Use case discovery, prioritisation, a target operating model, a roadmap and a business case. No running software. | Lowest as a single fee, highest per unit of change delivered, because implementation is still entirely ahead of you. | A board needs an investment case before releasing budget, or divisions disagree on where to start. Buy it knowing it produces a decision, not a system. |
| Build engagement | A defined system in production, integrated with your systems of record, with an acceptance test, the code and repository in your ownership. | Higher than a strategy engagement, fixed scope where the outcome can be specified. The cost lands once. | The use case is agreed, the data is good enough or can be fixed inside the scope, and someone internal will own the system afterwards. |
| Build plus operate retainer | The build, then monitoring, model and vendor changes, cost management, tuning and a route for change requests without a new contract each time. | The build cost plus a recurring fee. Highest total over a year, and usually the lowest cost of failure. | The system touches customers or revenue, vendor change would hurt you, or no internal engineer can own it. Most mid-market AI agents belong here. |
How to read a consultancy's own website before the meeting
You can do most of the filtering before you speak to anyone. Look for these signals.
- Case studies that name a shipped system, not a framework. The tell is the object of the sentence. A customer portal that handles something is a system. An AI maturity assessment and a roadmap is a document. Both are legitimate, and they are not the same purchase.
- Named numbers. Prices, ranges, timescales, record volumes, response times. A site that describes value only in adjectives has nothing it is prepared to be held to.
- Engineering detail in public. Architecture, integration mechanics, failure handling and cost control, because firms that build deal with these daily.
- Evidence of maintenance. Dated pages, corrections and content that reflects this quarter's vendor pricing rather than last year's. A firm that lets its own published facts rot will let your system rot.
- Who the authors are. Bylined technical writing from people who would appear on your delivery team beats an unattributed blog.
Where SpotDev fits
SpotDev is an AI and digital transformation consultancy that builds real software, an OpenAI Select Partner and a Claude Registered Partner. We are also a HubSpot Diamond Solutions Partner, Custom Integration Accredited and Onboarding Accredited, Cyber Essentials Plus certified, and rated 5.0 across 38 reviews on the HubSpot Solutions Directory. Our engineering team is in-house and fully remote, and it will answer all seven questions above, including question three. Related reading: AI implementation, HubSpot development, and Claude AI agents for business.
If you want the questions applied to your own situation rather than to a shortlist, start with a diagnostic, or Request a Quote.
Frequently asked questions
Is a strategy engagement ever the right first purchase?
Yes, when the organisation has not decided what it wants and stakeholders disagree. Buy it as a decision-making exercise with a fixed end, not as a precursor that automatically becomes a larger programme with the same firm.
What does a good definition of done look like for an AI agent?
A named job, a named exit condition, named escalation triggers, and an acceptance test run against an agreed sample of real records at an agreed accuracy. If it cannot be written in two sentences, it is not a definition of done.
Should we fix our CRM data before starting AI work?
Usually yes, and price it as a workstream rather than an assumption. Poor data affects both the reliability of the output and the running cost, because badly structured records cause the system to do more work per result.
How do we compare running costs between proposals?
Insist on a cost per business unit, such as per resolved conversation or per chased quote, plus an expected monthly volume and a cap. Published rates make this straightforward where a vendor product is involved, as HubSpot's credit rate sheet shows, and a firm that has operated custom systems can produce the same figures for a bespoke build.
Who should own the code at the end?
You should, and you should hold repository access throughout rather than receiving it at handover. Confirm the same for prompts and configuration, which determine behaviour just as much as the code does.
Stay Updated with Our Latest Insights
Get expert HubSpot tips and integration strategies delivered to your inbox.
