ChatGPT vs custom AI: when to buy, configure or build

Buy a licence first. ChatGPT Business or Microsoft 365 Copilot will cover drafting, summarising and research for most teams, for a monthly fee per person. Configure an assistant over your own documents when people keep asking the same questions of the same files. Build custom only when the AI has to act inside your systems, at volume, with checks you can audit.

Our rule of thumb is blunt. If an off-the-shelf tool does the whole job, buy it and don’t look back. If it does 80% of the job, look hard at the other 20%, because that’s where the pain lives.

What are the three options?

Off-the-shelf means a general assistant your team logs into. ChatGPT Business, ChatGPT Enterprise and Microsoft 365 Copilot are the common ones in UK offices. You pay per seat and configure almost nothing.

Configured means the same kind of model, set up for one job. That might be a custom GPT with your instructions and files. Or it might be a retrieval-augmented generation (RAG) assistant: a tool that searches your documents first, then answers from what it found.

Custom build means software written for your business. It connects to your CRM, finance system or database, takes actions, and keeps a record of what it did. Some custom builds are AI agents: systems that carry out steps, not just answers. Many are simpler, like a pipeline that reads and tags documents.

What does off-the-shelf AI cost?

These are the vendors’ own list prices at the time of writing, in US dollars, before VAT. UK prices differ by currency and change often, so check the pages yourself before budgeting.

• ChatGPT Business: OpenAI’s business pricing page lists a standard seat at $20 per user per month billed annually, or $25 billed monthly. The minimum is two seats.
• Microsoft 365 Copilot Business: Microsoft’s Copilot pricing page lists it at $21 per user per month, billed yearly. It’s an add-on to an eligible Microsoft 365 Business plan. It’s on offer at $18 for the first year until 31 December 2026.
• Microsoft 365 Copilot for enterprise: the enterprise page lists $30 per user per month, paid yearly. It needs a qualifying Microsoft 365 licence on top.

Microsoft also includes a web-based Copilot Chat at no extra cost with eligible Microsoft 365 plans. If you already pay for Microsoft 365, start there. ChatGPT Enterprise is priced by negotiation, so there’s no public number to quote.

How do you choose between ChatGPT and Copilot?

It mostly comes down to where your team works.

• Mostly in Outlook, Word, Excel and Teams? Copilot sits inside those apps and follows your SharePoint permissions. If those permissions are loose, it will surface files people shouldn’t see, so tidy them first.
• Across lots of tools, or heavy on research and writing? ChatGPT usually suits a mix of tools better, and gets OpenAI’s new features first.
• Not sure? Give five people each tool for a month. Ask them which one they opened more.

Whichever you choose, the bigger risk is staff using personal accounts for work. OpenAI says ChatGPT Business data isn’t used to train its models by default. Free personal accounts don’t carry the same business terms. Paying for proper seats is partly a data protection decision. Our guide to AI, GDPR and governance covers the policy side.

Buy, configure or build: the decision table

When does off-the-shelf win?

More often than consultancies will admit. Off-the-shelf wins when:

• the work is personal productivity, like drafting emails, summarising meetings or rewriting copy
• each person’s task is different, so there’s no single process to automate
• a person reads every output before it goes anywhere
• volume is low enough that copying and pasting is fine

If that’s you, don’t hire anyone. Buy the seats, write a one-page usage policy and run a lunchtime session on good prompts.

What if it only does 80% of the job?

This is the question that decides most of these choices. A tool that covers 80% of a process looks like a bargain on the pricing page. Then your team spends every day working around the missing fifth.

“Sometimes if it only does 80% of what you’re doing, the 20% still becomes hell.”

So test the gap, not the headline. List the steps of the real job, end to end. Mark which ones the tool does, which it does badly and which it can’t do at all. Then ask what the workarounds cost in hours, errors and staff patience.

If the gap is small and rare, live with it. If it’s daily, central to the job, or forces people back into copy and paste, that’s your case for the next step up.

When does a configured assistant win?

When the same questions hit the same documents over and over. HR policies, a product catalogue, contract templates, a technical manual. A configured assistant answers from those files and shows which page it used.

Configured tools hit a ceiling in three places:

• Permissions. If different people may only see different documents, platform tools can struggle to enforce it reliably.
• Freshness. If documents change daily, someone has to keep re-uploading them.
• Actions. They answer questions well. They don’t usually update your CRM or raise an invoice.

Before you configure anything, check your documents are current and consistent. An assistant over a messy shared drive gives confident answers from out-of-date files. Our guide on getting your data ready for AI has a checklist.

When does a custom build win?

When at least two of these are true:

• Volume. Hundreds or thousands of items a week, every week.
• Integration. The AI needs to read from and write to your own systems.
• Control. You need to log every decision, check outputs against rules, or keep data in a specific place.
• Conditions. It has to work where a general assistant can’t, like offline, on a phone or inside your product.
• Ownership. The AI is part of what you sell, so you need to own the code.

At Everyone’s Invited, a safeguarding charity was getting over 100 submissions a day. Each had to be stripped of names, dates and locations before review. Pasting survivors’ stories into a chat window was never an option. The custom system went on to handle more than 50,000 submissions and saved around 12,500 hours of review.

Pristine Condition needed AI that checks a person’s lifting technique through a phone camera, in real time, with no internet connection. A spot check might happen in an aircraft cargo hold. No off-the-shelf assistant does that. We built it with TensorFlow, running on the phone itself, and the platform is now used by around 150,000 warehouse staff.

Is a thin “wrapper” around ChatGPT ever enough?

Sometimes it’s the smartest product in the market. A wrapper is an app that sends your users’ requests to someone else’s model, like OpenAI’s, and wraps a better experience around the answer.

Take calorie counting. The old way is scanning a barcode on every packet. A newer kind of app lets you photograph your plate, and a model estimates what’s on it. The model isn’t the clever part. The win is that it’s faster and easier for the user.

So a wrapper can win on experience. Where wrappers go badly wrong is in how they’re built. Here’s the checklist we use.

• Locked to one model. Models change every few months. If yours is wired tightly to one vendor, every model change means rework. Even with a clean design, a swap needs re-testing.
• No test set. Without a fixed set of real examples and the right answers, you can’t tell if an update made things worse.
• No monitoring. Nobody watches what goes in and what comes out, so problems surface as customer complaints.
• Open to jailbreaks. A jailbreak is a user tricking the model into ignoring its instructions. Assume someone will try.

Every AI company says it’s an AI company now. Ask anyone selling you a build how they’d swap the model, and how they measure accuracy. Vague answers to both are your hype check.

How do you build custom AI so it lasts?

Build it so the model is a part you can replace. Keep the instructions, the test set and the scoring outside the model. When a better or cheaper model arrives, you run the test set against it and switch if it wins. Expect to adjust the instructions for each new model. If you’ve fine-tuned, you’ll need to repeat that step on the new one.

“Your accuracy is never going to be 100%, because that’s the wrong target.”

The right target is “better than the person doing it now”, counting what each kind of mistake costs, not just how many there are. To know that, measure how the person does today, on the same examples. Then you have a number to beat, and a number to watch after launch.

What does a custom AI build cost?

We won’t give you a number in an article, and you should distrust anyone who does. Three things drive the cost:

1. How many systems it touches. Each integration is design, build and testing work, and old systems are slower.
2. How clean your data is. Messy data is the most common reason AI projects overrun.
3. How much checking it needs. A system that drafts for a person to approve is far simpler than one that acts alone.

There are also running costs: model usage, hosting and someone keeping it healthy. Budget for those before you start.

If AI is part of a new product rather than an internal tool, it fits our fixed-price LaunchPad. Growth is £60k over three months and includes complex data and AI features. Scale is £100k over four months and adds data and AI strategy support. Both exclude VAT, and you own the code and IP. See LaunchPad pricing for the full list.

For internal tools, our process is a free chat, then a paid discovery that ends with a scoped plan and a price.

What you’re signing up to after launch

A custom build is a system you run, not a file you receive. Models get updated and retired, so someone has to run your test set against new versions. Your documents and rules change, so the checks need updating too. Usage costs move with volume, so someone should watch the bill.

We agree how that work is covered before the build starts. Whoever you pick, ask who’ll do this work a year after launch, and what it costs.

When you shouldn’t hire an agency like us

• Your needs are personal productivity. A licence does it. You’d be paying us to tell you that.
• You haven’t tried off-the-shelf yet. A month of real use is the best brief you’ll ever write. Our AI readiness scorecard helps you rank what you find.
• A configured tool covers it. If a RAG platform answers your team’s questions well enough, keep it.
• Nobody will own the system. A custom build needs someone inside your business who checks it and cares.

If you already have something half-built and you’re unsure about it, ask for our free architectural review. We’ll tell you what we’d keep and what we’d change.

If you’ve outgrown the licence and want to know whether configuring or building is the right next step, talk to us. We’ll look at the actual job and tell you which of the three it needs.