The Short Answer
If your team is using ChatGPT to draft, summarise, research and think out loud, a subscription is doing its job and you should leave it alone. We turn away work on this basis regularly.
The line is crossed when you need one of four things a seat-based subscription structurally cannot give you, whatever plan you are on. It has to act rather than answer. It has to serve people who do not have a seat, which usually means your customers. It has to run when nobody is watching. Or it has to answer the same way every time, and be able to prove it did.
If none of those apply, stop reading and keep your money. If one does, the rest of this explains what changes.
What a Subscription Is Actually Built To Do
It is worth being precise about the boundary, because most comparisons on this subject are written by people selling the alternative.
It reads with the permissions of the person using it. OpenAI's help documentation for company knowledge in Business, Enterprise and Edu workspaces is explicit that ChatGPT can only access what each user is already allowed to view, and that after an app is enabled each user authorises their own account, within their existing permissions. That is a sensible security design. It also means the assistant is an extension of one employee, not a service your business operates.
It mostly reads rather than writes. OpenAI's documentation states that its own built-in apps are search-only and do not support write actions, pointing to custom connectors for write or modify capability, and that write actions are unavailable when a request is answered through company knowledge, which is optimised for retrieval. Where write actions do exist, the default is to ask for approval during the run.
It is designed around a person being present. That default approval step is the right behaviour for a tool a human is driving. It is the wrong behaviour for a process that has to complete at two in the morning.
None of this is a criticism. A subscription is extremely good value for the work it is built for. The mistake is assuming that because it can read your CRM in a chat window, it can run your enquiry process.
The Four Lines
It has to act, not answer. Drafting the follow-up email is a subscription job. Writing the outcome back into the CRM, moving the deal stage, issuing the credit note, booking the slot — those are actions against systems, with consequences when they go wrong, and they need permissions the business holds rather than permissions an individual employee holds.
It has to serve people without a seat. Your customers do not have a login to your workspace. An assistant that answers your staff is a productivity tool; a thing that answers your customers is a service you operate, and it needs its own identity, its own boundaries and its own audit trail. Publishing an internal agent outward is not the answer either: OpenAI's own guidance warns that where an agent is published using the creator's personal connections, anyone who can use that agent may access data or perform actions as the creator.
It has to run when nobody is watching. This is the line that surprises people on the Microsoft side in particular, because the billing model changes when you cross it. Microsoft's Copilot Studio licensing guidance sets out that a Microsoft 365 Copilot licence covers interactive use inside Microsoft 365 surfaces, while an agent triggered to run autonomously consumes Copilot Credits from a tenant pack, that building and managing agents requires a per-user Copilot Studio licence, and that an Azure subscription is required. So the honest comparison is rarely "free thing we already have" against "expensive thing we would build". It is metered consumption per action against a fixed piece of work.
It has to answer the same way every time. A subscription gives every employee their own conversation, their own prompt and their own interpretation. That is a feature for thinking and a problem for anything a regulator, a client or a complaint might later examine. If the answer needs to be identical, recorded, and defensible, it needs to come from one place.
The Comparison
| ChatGPT or Copilot subscription | Custom agent or integration | |
|---|---|---|
| Best at | Drafting, summarising, research, thinking out loud | Running a defined process against your systems |
| Who it serves | People with a seat | Anyone, including customers with no login |
| Reads your data | With each user's own permissions | With permissions the business grants and controls |
| Writes to your systems | Limited, and approval-gated by default | Yes, within boundaries you define |
| Runs unattended | Not what it is designed for | Yes, on a trigger or a schedule |
| Consistency | Varies by person and prompt | Same input, same behaviour, logged |
| Cost shape | Per seat, per month, forever | A build, then support, plus model usage |
| Fails badly when | Asked to be a business process | Bought before the process was written down |
| Effort to adopt | An afternoon | Weeks, and access to your systems |
The Honest Test
Ask these four questions about the thing you have in mind.
Does a person have to be sitting there? If yes, a subscription is very likely enough.
Whose permissions does it need? If the answer is "one employee's", a subscription is enough. If it is "the business's", you are building.
Who is on the other end? Staff means subscription. Customers means a service you operate.
What happens if two people ask it the same thing and get different answers? If the answer is "nothing much", a subscription is enough. If the answer involves a complaint, a regulator or a refund, it needs to be one system giving one answer.
Most businesses we speak to answer those four in a way that points firmly at the subscription they already have, and we tell them so on the call.
What Does Not Change
Two things people expect to gain by building, which they mostly already have.
Your data is not training the model by default. OpenAI states that its models are not trained on data from business plans or the API unless a customer explicitly opts in. Building something custom does not fix a problem that was not there.
UK data residency is available without a custom build. OpenAI's documentation lists data residency for eligible ChatGPT Enterprise and API customers covering the UK among other regions, with in-region inference available for some, encryption at rest and in transit, and customer-managed keys on some plans. If UK residency is your reason for building, check your plan's eligibility first — you may already be able to have it. Confirm the current position with the vendor, because this area moves.
Where the Work Actually Goes
When a build is justified, the engineering is rarely the hard part. The time goes into deciding what the thing is allowed to do, which is work only you can do: which systems it may touch, what it must never attempt, who it escalates to, and what happens when a system it depends on is down. We write the first draft of that list from your processes and you correct it, and it becomes the specification rather than an afterthought.
For how that looks against a CRM specifically, see /blog/ai-crm-integration-uk-guide. For the same honesty applied to agency fees, /blog/ai-automation-agency-cost-uk covers what you are paying for when the underlying tools cost very little. Sector context is at /industries/accountants and /industries/law-firms, the two where "the same answer every time, and provably so" carries the most weight. Scope levels are on /services/ai-integration.
Next Step
If you are trying to work out which side of the line you are on, that is a short conversation rather than a project. Book a free discovery call at /contact, message us on WhatsApp, or ring — describe the process, and we will tell you honestly whether the subscription you already pay for covers it.
Frequently Asked Questions
Our team already has ChatGPT seats. Is that not enough?
Usually, yes, and we say so often. A subscription is excellent at drafting, summarising and research, and if that is what you need you should leave it alone. It stops being enough when the work has to act on your systems rather than answer a person, serve customers who have no seat, run when nobody is watching, or produce the same answer every time in a way you can later evidence. If none of those four apply to what you have in mind, you do not have a build to commission.
Can we not just connect ChatGPT to our CRM and be done?
You can connect it, and for reading it works well. The constraint is what that connection is. OpenAI's documentation for Business and Enterprise workspaces explains that each user authorises their own account and ChatGPT accesses only what that person is already permitted to see, typically read-only, and that its built-in apps are search-only without write actions. So your salesperson can ask about a deal. The system cannot update the deal on behalf of the business, and it cannot answer a customer who has no login.
Is a custom build more expensive than the subscription?
It is a different shape rather than simply more, and the comparison people make is usually the wrong one. A subscription is a per-seat charge that continues for as long as you have staff. A build is a one-off piece of work plus support and model usage. There is also a step most people miss: Microsoft\u2019s Copilot Studio licensing guidance sets out that autonomous agent runs draw on Copilot Credits from a tenant pack rather than being covered by an interactive Microsoft 365 Copilot licence, and that building agents needs its own per-user licence and an Azure subscription. The moment the work runs unattended, you are in metered consumption either way. We quote a fixed price after a free discovery call.
Do we need to build something to keep our data in the UK?
Probably not, and it is worth checking before spending anything. OpenAI's documentation lists data residency for eligible ChatGPT Enterprise and API customers covering the UK among other regions, alongside in-region inference options, encryption at rest and in transit, and customer-managed keys on some plans. It also states that models are not trained on business plan or API data unless a customer opts in. If residency is the reason you are considering a build, confirm your plan's eligibility with the vendor first \u2014 this area changes, and you may already have what you are about to pay for.
What actually takes the time if we do build?
Not the engineering. The time goes into deciding what the thing is allowed to do, and that is work only you can do: which systems it may touch, what it must never attempt, who it escalates to, and how it behaves when something it depends on is unavailable. We draft that list from your processes and you correct it, and it becomes the specification the build is tested against. Projects that skip it are the ones that disappoint in month three.