We asked ChatGPT to scope our CRM migration. Here’s what it missed.

We gave ChatGPT a realistic Salesforce-to-HubSpot migration brief and asked it to write a scope we could send out to vendors. Nine seconds later it came back with a 38-section request for proposal, complete with an evaluation scorecard, a weighted scoring matrix, reference-check questions and a stage-gated contract structure. On paper, it looks like everything you’d need to start engaging with vendors.

Look a little closer, though, and what it produced is a near-perfect template rather than a scope you could actually act on. The template was always going to be the easy part, and ChatGPT wrote a genuinely good one. What it can’t hand you are the practical steps that only reveal themselves once someone has been inside the system, the mess buried in six years of real data, and the experience that tells you which of its confident recommendations will quietly fall over once real vendors, and a real go-live, get involved. A plan you can’t execute isn’t a plan: it’s a reading list.

We built the alternative, our CRM Migration Assessment, for exactly this reason. First, though, here's what ChatGPT gave us.

What we asked it

We used a deliberately ordinary prompt, the kind a busy sales leader might type on a Friday. The business is invented rather than a real client, though the brief itself is the kind we see all the time:

We're planning to migrate from Salesforce to HubSpot, and I want to put together a scope document we can send to potential implementation partners so they can quote against it.

Context:

- B2B company, ~150 staff, been on Salesforce Sales Cloud about 6 years
- Roughly 40,000 contacts and 8,000 open and closed opportunities
- A fair amount of customisation: several custom fields, two custom objects, and a bunch of flows and process builder automations
- Salesforce is integrated with our ERP (NetSuite) and our invoicing system
- Marketing currently runs in a separate tool, and we want marketing, sales and service all on HubSpot
- Sales team of about 30, adoption is patchy
- We'd like to be live before the start of our next financial year

Produce a complete scope of work we can send to vendors, including what you'd want them to deliver and how we should evaluate and choose between them.

There’s nothing unusual in there, and that’s the point: it’s exactly the sort of brief a consultant would treat as the opening of a conversation.

The output and credit where it’s due

We’ll be straight about this: the output was good. It refused to treat the job as a lift-and-shift, telling you in as many words that the goal is “not simply to reproduce Salesforce in HubSpot.” It built a current-state and target-state assessment, a Salesforce-to-HubSpot functional mapping matrix, a data-migration approach with reconciliation and two test cycles, and a list of deliverables running to 31 items. It went further than most people would bother to, adding a weighted evaluation scorecard, a 1 to 5 scoring guide, reference-check questions, a set of red flags to watch for, and a stage-gated contract so you never sign up to the whole build in one go. It even warned, without a flicker of irony, against partners whose proposal is “essentially a generic HubSpot implementation template.”

Hand that document to an experienced team and they would nod along. So the interesting question was never whether ChatGPT knows what a migration RFP looks like, because it clearly does. The question is whether it can scope yours, and that is where it comes undone.

It never asked us a single question

A proper scoping session opens with interrogation, not answers: why are you actually moving, what lies behind the patchy adoption, is NetSuite or Salesforce the system of record for a customer, which of the two custom objects is load-bearing and which is simply habit. A consultant will ask 30 questions like these before committing to anything, because each answer bends the scope into a different shape.

ChatGPT asked none. It took every assumption in the brief at face value and produced 38 sections on top of it, and the cost of that shows up in three ways.

The first is that, because it never interrogated the brief, it cannot tell you whether anything has been missed. You can issue the whole thing and still have no way of knowing that no stone has been left unturned, which is a strange kind of preparation: it feels thorough, and underneath it you are no better informed than when you started.

The second is industry. It never asks what you actually do, describing you only as “a B2B organisation,” and when it turns to choosing a partner it tells you to look for “comparable projects” by size and data volume. That is useful up to a point, but a healthcare provider, a law firm and a manufacturer are not running the same migration; they carry different regulatory and data-handling obligations that a good partner has to design around from the first workshop. Industry-specific experience usually matters more than a matching headcount, and ChatGPT never thought to ask which industry it was scoping for.

The third is the one our own sales team runs into most. Since AI-created content lacks the same level of personal involvement, partners often receive documents from individuals who haven't fully digested the content. So the first conversation is spent unpicking a 38-section RFP the prospect did not author and cannot fully explain, working out together which parts describe a genuine requirement and which are simply template. That is a slower, harder start than a short brief the business actually understands and can stand behind.

It puts everything in scope

There is a quieter problem beneath all that length. The RFP scopes everything at once, Marketing, Sales and Service, a knowledge base, a customer portal, two system integrations and the full reporting suite, and treats the lot as in scope by default. What it never does is help you decide what to do first. Your burning issue is patchy sales adoption, which is a strong argument for getting Sales live and genuinely used before anything else and phasing Service in later, but the model has no idea that is your priority, because it did not ask. It offers textbook delivery phases, discovery, design, build, and calls that phasing, when the phasing that actually matters is strategic: sequencing the work around the outcome your business needs first. Working that out means understanding the business, which is the one thing a template cannot do.

It can’t see your data, so every specific is a guess

A scope is only ever as good as the discovery beneath it, and discovery means getting inside the actual Salesforce org and seeing what is really there. ChatGPT never did that, because it cannot, which is why the RFP is stitched together from blanks: “[Company name],” “[FY XXXX],” and a data table it politely flags for the partner to “validate during discovery.” It has no way of knowing whether this business is carrying 200 duplicate contacts or 20,000, or whether those two custom objects are pristine or a decade of accumulated mess. The most important input to any real scope, the true state of the system, is exactly the thing it cannot see, so it does the only thing it can and hands the question on to whichever vendor you send it to.

It handed us the hard part back

Once you notice that handing-off, you see it everywhere. Read the RFP closely and almost every section asks somebody else to make the decision. The partner “should assess” the two custom objects and “recommend” what becomes of each. The functional mapping matrix, the very thing that decides what to migrate, redesign, replace or retire, arrives as an empty grid for the vendor to fill in. The 16 “key assumptions vendors should validate” are really 16 things ChatGPT does not know and cannot find out. It reads like rigorous procurement, and in a way it is, yet every genuinely hard call has been deferred to a person who has not been hired yet. The thinking, which is the reason you wanted help in the first place, is still entirely undone.

Its confidence doesn’t mean it is accurate

The RFP lays out its phases and its stage gates with real confidence, and it fills the timetable with “[DATE]” after “[DATE]” because it has no idea when your financial year starts. The structure is not wrong, exactly; it is a reasonable industry average, floating free of anything specific to this business. The platform detail is the same: it names the Hubs, defers the tiers, and sounds every bit as authoritative whether or not a given recommendation fits your portal, because it cannot tell the difference. A confident tone is not the same thing as being right for you, and from the inside that gap stays invisible unless you already know the platform well enough to spot it.

That gap is exactly what a consultant closes, and it is why AI output in a migration needs a human to validate it rather than rubber-stamp it. It is also why we built our CRM Migration Assessment: the speed of an AI-generated scope, with a HubSpot consultant checking it against your real system, asking the questions the model skipped, and standing behind what comes out the other end.

Why this is the dangerous kind of wrong

Had ChatGPT produced an obviously poor document, there would be no harm in it, because you would dismiss it on sight. The danger runs the other way. It produced an impressive one, in nine seconds, and it reads as though a procurement team laboured over it for a fortnight, which is precisely why you might actually send it. A business without deep CRM expertise looks at 38 well-structured sections and reasonably decides it now has an RFP and can go to market. But the holes in an AI scope are never in the structure, which is the part it does so well; they sit in the judgement, in the decisions it deferred, the discovery it could not do, the questions it never asked and the industry it never established. And holes in judgement are exactly the ones you cannot see unless you already have the expertise to know they ought to be there. The polish, in other words, is the risk.

Where AI genuinely helps (and we mean it)

None of this makes us anti-AI, and if anything it makes the opposite case. We build on HubSpot’s own AI in Breeze, we run an AI Engine workshop for the teams we work with, and AI runs through the way we deliver projects. Inside a migration it genuinely earns its place, speeding up the data cleansing and deduplication across tens of thousands of rows, proposing field mappings for a specialist to check over, flagging the records that did not come across cleanly, drafting the internal comms and the training, and summarising a tangle of legacy automations so that a human can make the call faster. It takes real hours out of the work.

That is the honest distinction. In expert hands AI is a remarkable accelerator, and as a substitute for the expert it becomes a liability. The value was never in the tool by itself; it lives in who is holding it, and in whether that person can tell when a confident, well-cited plan is quietly missing the handful of things that will sink your go-live.

What AI can’t do

ChatGPT gave us a strong template, but what it could not give us was a scope, because a scope means seeing your system for yourself, asking the questions the brief never answers, making the calls the model left open, and then standing behind the result once it is live. None of that is a prompt away. It is simply the job.

BabelQuest is one of the UK’s most accredited HubSpot Elite Solutions Partners, ranked 7th in Europe and 11th globally in December 2025, with a team of 35 and every HubSpot accreditation available. We use AI heavily ourselves, and we have migrated businesses onto HubSpot from Salesforce, Dynamics 365, Zoho, Pipedrive, legacy systems and spreadsheets. The people who scope your migration are the same people who build it, and the same people who answer for it when it matters.

Book a scoping call with someone who’ll ask the questions ChatGPT didn’t. Get in touch, or start with our CRM Migration Assessment.

FAQ

Can ChatGPT plan a CRM migration? ChatGPT can produce a strong generic migration framework, including phases, a RACI, testing and cutover steps. What it can’t do is scope your specific migration, because it can’t see your data, doesn’t know what your custom objects or integrations do, and can’t ask the questions a consultant would. It gives you a template to fill in, not a plan you can execute, and the decisions it defers are the hard part.

Can AI migrate CRM data? AI can accelerate parts of a data migration, such as deduplication, standardising formats, suggesting field mappings and flagging records that didn’t transfer cleanly. It can’t own the migration itself: assessing the real state of your data, designing the target model, preserving reporting history and validating the result still require expertise and accountability.

Should I use AI to plan a HubSpot migration? Use AI as an accelerator, not a substitute for expertise. It’s genuinely useful for cleansing, mapping suggestions, validation and drafting documentation, and it can give you a helpful starting framework. But AI produces confident, plausible plans without knowing whether they fit your business, so you need CRM expertise to judge what applies to you and to make the decisions it can’t.

Can AI replace a HubSpot partner? No. AI is a powerful tool within a HubSpot implementation, and good partners use it every day. It can’t replace the partner, because a migration needs someone to run discovery on your actual system, ask the questions your brief doesn’t answer, make platform decisions specific to your setup, and stand behind the outcome on go-live day. AI can’t be accountable for a result.

What can AI do well in a CRM migration? AI is strong at the repetitive, high-volume work: deduplicating and standardising data across tens of thousands of rows, proposing field mappings for review, spotting records that failed to migrate, summarising legacy automations, and drafting training and communications. Used this way, in expert hands, it takes real time off a migration.