We design, build and run OpenAI systems that do real work inside your data, your CRM and your internal tools. Most AI projects stall because nothing connects the model to the systems that hold the facts, so that is where we start. SpotDev is a software engineering firm, an OpenAI Select Partner and a HubSpot Diamond partner, working with UK companies from around £3m in revenue upwards.
.png?height=28&name=Gunnebo%20Logo%20(1).png){kind=small}
.png?height=28&name=Closer%20Still%20Media%20(1).png){kind=small}
.png?height=28&name=Ecommpay%20Logo%20(1).png){kind=small}
.webp?height=28&name=Tussell%20Logo%20(1).webp){kind=small}
-4npmo6jxod.png?height=28&name=iconsolutionsrgb%20(1)-4npmo6jxod.png){kind=small}
Most mid-market OpenAI deployments do not fail on the model. They fail on everything around it: nobody owns the workspace, the agent cannot reach the systems that hold the answers, the pilot is built on a product OpenAI is retiring, and the bill moves without anyone being able to say why. Those are engineering and governance problems, and they are the ones we take on.
Customer data, contracts and pricing end up in personal consumer accounts, where conversations are used for training unless someone actively switches it off. Nobody can say who has access, what has been shared, or what to do when a person leaves.
A single business workspace with single sign-on, verified domains, defined roles and an acceptable-use policy your people have actually read. Business and API data is opted out of training by default, and we choose the tier against your real obligations rather than the cheapest line item.
A demo that impressed the board, built by someone who has since left, sitting on a product with a published end date. OpenAI is retiring the Assistants API on 26 Aug 2026, and AgentKit's Agent Builder and Evals are being wound down through late 2026.
We build on what OpenAI currently supports and maintains, and we treat platform churn as our job rather than yours. When something is deprecated we tell you before the notice period bites, and we cost the migration.
It writes well and it is confident, and it is guessing. It has no access to the CRM, the finance system or the document store, so it cannot tell you what a customer is actually worth or what was agreed in the contract.
We connect the agent to the systems that hold the facts, with read access scoped deliberately and any write back to your CRM held behind an approval step. Answers cite the record they came from, so someone can check them.
Usage climbs, the invoice climbs with it, and no one can attribute the spend to a workflow or say whether the work was worth doing. The usual reaction is to ban the tool rather than fix the design.
Model choice per task instead of the largest model for everything, caching of the context that repeats on every call, and batch processing for bulk jobs that do not need an instant answer. Usage is reported per workflow, so you can see what each one costs and what it returns.
Customer data, contracts and pricing end up in personal consumer accounts, where conversations are used for training unless someone actively switches it off. Nobody can say who has access, what has been shared, or what to do when a person leaves.
A single business workspace with single sign-on, verified domains, defined roles and an acceptable-use policy your people have actually read. Business and API data is opted out of training by default, and we choose the tier against your real obligations rather than the cheapest line item.
A demo that impressed the board, built by someone who has since left, sitting on a product with a published end date. OpenAI is retiring the Assistants API on 26 Aug 2026, and AgentKit's Agent Builder and Evals are being wound down through late 2026.
We build on what OpenAI currently supports and maintains, and we treat platform churn as our job rather than yours. When something is deprecated we tell you before the notice period bites, and we cost the migration.
It writes well and it is confident, and it is guessing. It has no access to the CRM, the finance system or the document store, so it cannot tell you what a customer is actually worth or what was agreed in the contract.
We connect the agent to the systems that hold the facts, with read access scoped deliberately and any write back to your CRM held behind an approval step. Answers cite the record they came from, so someone can check them.
Usage climbs, the invoice climbs with it, and no one can attribute the spend to a workflow or say whether the work was worth doing. The usual reaction is to ban the tool rather than fix the design.
Model choice per task instead of the largest model for everything, caching of the context that repeats on every call, and batch processing for bulk jobs that do not need an instant answer. Usage is reported per workflow, so you can see what each one costs and what it returns.
Swipe to explore
OpenAI implementation is design, engineering, integration, governance and adoption. We are a software engineering firm, so we build production systems and put them into the software you already run. You do not need to be a HubSpot customer to work with us, though most of our clients are.
Agents scoped to a real job, built on OpenAI's current agent stack rather than a retiring one. Each has defined inputs, defined outputs, logging you can audit, and a human approval step wherever it acts on something that matters.
The work that makes an agent useful: connecting it to HubSpot, your finance system, your document store and your internal tools, under scoped OAuth access. HubSpot publishes a hosted MCP server for exactly this, and where a system has no equivalent we build the connection ourselves.
Workspace setup, single sign-on, domain verification, roles, retention settings and an acceptable-use policy, then structured training so the seats get used. We apply the same training method we use for HubSpot rollouts, because paying for licences is not the same as adoption.
A model is only as good as what it can reach, so we get the underlying data into a state worth querying: deduplicated, structured, and indexed for retrieval. Where output goes back into your CRM we enforce the schema, so you get clean field values rather than prose.
Your DPA and transfer position, data residency where you need it, retention settings, access control, logging, and a written record of what the agent may and may not do. We tell you plainly which controls your current tier does not include.
If you have something built on the Assistants API or on AgentKit's Agent Builder or Evals, we audit it, decide what is worth preserving, and rebuild it on the supported stack before the deadline forces the decision.
We look at the work you actually want done, the systems that hold the data, and your obligations, then rank the candidates by value and by how hard they are to build.
We agree the scope and the price before we start, then build against your systems on the platform stack OpenAI currently supports.
We ship it into your environment with access control, logging and retention configured, and we train the people who will use it every day.
We measure what it costs and what it returns, tighten the parts that underperform, and handle platform changes before a deprecation date becomes your problem.
A good deal of what businesses want from OpenAI needs no engineering at all, and we will say so. We would rather turn down a project than sell you a build you did not need, because the ones that get commissioned for the wrong reason are the ones that get switched off a year later. Here is the honest split.
If the job is individual work (drafting, summarising, first-pass research, tidying a spreadsheet, getting a document started) then a ChatGPT Business or Enterprise subscription does it out of the box. Buy the seats, turn on single sign-on, write a one-page acceptable-use policy and train people properly. Your cost is licences and a few hours, not a project.
Engineering earns its money when the work has to reach your systems of record, run unattended, produce the same result every time, be auditable afterwards, or write back into your CRM. Off-the-shelf ChatGPT cannot do any of that against your data, and no amount of prompt writing changes it.
If the job sits entirely inside HubSpot, HubSpot's own AI (Breeze and Agent Hub) may already cover it and you are paying for it. If the workload suits Anthropic's Claude better, we will say so and build it there instead. And if your CRM data is duplicated and half empty, the first project is the data, not the model. That one comes up more than any other.
We've delivered 300+ HubSpot implementations as a firm, and we build production AI on more than one platform. Technology only delivers value when people use it, so adoption is built into everything we ship.
Agents and integrations designed, tested and shipped by our own engineers. Nothing is subcontracted.
Transformation projects shipped for B2B teams across the UK and beyond.
We are deepest on Anthropic's Claude and separately an OpenAI Select Partner, so the model is chosen against your workload rather than around what we sell.
<p>It means SpotDev is enrolled in the OpenAI Partner Network at the Select tier. OpenAI describes the programme as a global programme for partners to build, sell and deliver AI solutions with OpenAI, and it has not published what separates one tier from the next, so we will not characterise it beyond that. It does not mean OpenAI has reviewed, endorsed or guaranteed our work, and it does not give us influence over OpenAI's roadmap, support queue or pricing. Treat it as one proof point alongside the rest, not as a reason to skip your own due diligence.</p>
<p>No. You buy your ChatGPT subscriptions and your API usage directly from OpenAI, on your own contract and your own Data Processing Addendum, and you keep control of both. We do not resell licences, we do not take a margin on your OpenAI spend, and we are not paid by OpenAI for putting you on their platform. We charge for the engineering, which means we have no commercial reason to put you on a larger plan than the work requires.</p>
<p>We work with both and choose per workload rather than by preference. We look at what the job actually requires, what your business already runs and has already contracted for, what procurement and compliance will accept, and what it costs to run at your expected volume. We hold deep Claude specialism and Select Partner status with OpenAI, and we are independent of both: we resell neither and take no margin on your usage. That independence is deliberate, because it means the recommendation you get is the one that fits your situation rather than the one that suits our commercial arrangements. Where OpenAI is the better answer we will say so and build it there, and where the answer is the AI you already pay for inside HubSpot, we will say that too.</p>
<p>This is the question most worth getting right, because the gap is larger than most buyers expect. ChatGPT Business holds SOC 2 Type 2, SAML single sign-on, an admin console and no training on your workspace data by default. It does not include SCIM provisioning, role-based access control, Enterprise Key Management, IP allowlisting, the Compliance Logs Platform, UK or EU data residency, or ISO 27001 certification. All of those are ChatGPT Enterprise only, and if your procurement pack or DPIA template asks for ISO 27001 or audit-log export specifically, Business will not satisfy it. One caveat worth planning for: data residency applies to new Enterprise workspaces and new API projects, and cannot be added to an existing one, so the decision has to be made before you set the workspace up.</p>
<p>For ChatGPT Business, ChatGPT Enterprise and the API, OpenAI's stated default is that your inputs and outputs are not used to train its models unless you explicitly opt in. Consumer ChatGPT is the opposite way round, which is the real risk in most businesses: individual accounts are opted in until someone turns training off. For UK customers, OpenAI's Data Processing Addendum covers international transfers using the EU Standard Contractual Clauses as amended by the ICO's UK Addendum, and the contracting entity is OpenAI OpCo, LLC in the United States. You still need your own lawful basis and your own transfer risk assessment. We will help you assemble the evidence, and we will tell you where a control you have been promised is actually sales-gated rather than self-serve.</p>
<p>Our work is fixed price, scoped and agreed before we start, quoted in pounds. Your OpenAI costs are separate and billed by OpenAI directly. In outline: ChatGPT plans are priced per user per month with a usage-based credit layer on top for heavier features, and the API is priced per token with input and output charged at different rates, a discount on repeated context that gets cached, a cheaper batch route for bulk work that does not need an instant answer, and an uplift for regional processing where you need data handled in a particular jurisdiction. Rates change often, so check OpenAI's current pricing page rather than any figure in an agency proposal. Part of our job is designing the workload so the running cost is predictable.</p>
<p>Those need moving. OpenAI is retiring the Assistants API on 26 Aug 2026, and AgentKit's Agent Builder and Evals are being wound down through late 2026, with several older model families already marked legacy or removed. We audit what you have, work out how much of the behaviour is worth preserving, and rebuild it on the currently supported stack, usually with better logging and cost control than the original had. If the original was a prototype that never earned its keep, we will tell you that too, and retiring it is a legitimate answer.</p>
SpotDev is an OpenAI Select Partner and a HubSpot Diamond Solutions Partner. OpenAI is an AI research and deployment company. SpotDev is an independent consultancy: 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.
