The vendor demo was slick. Their AI understood context, generated human-sounding responses, integrated with Salesforce in "just three clicks." Six months later you are on their third escalation engineer and wondering why the integration requires six weeks of professional services.
This is the gap between off-the-shelf promise and reality. Here is how to navigate it.
The off-the-shelf fantasy
The pitch: Best-in-class AI, continuously updated, enterprise-grade security, no engineering required.
The reality: It does 80% of what you need. The remaining 20% is critical to your business and impossible to customise. You either abandon your requirements or pay their professional services team to build workarounds on top of their platform.
The numbers: £30–100k per year base licence. Implementation: £50–200k. Annual support: 20%. Five-year total: £300k–800k. For something that never quite fits.
The custom build fantasy
The pitch: Exactly what you need, integrated how you want, owned by you.
The reality: You are now an AI engineering company. You have recruitment challenges, technical debt, model drift, and a system that only the person who built it understands. When they leave, so does your capability.
The numbers: £200–500k to build. £100–300k per year to maintain. Plus the hidden cost: management attention diverted from your actual business.
The actual decision framework
Stop asking "buy or build?" Start asking these questions.
Question 1: Is your use case genuinely unique?
Common scenarios (buy): email classification and routing, basic chatbot responses, document summarisation, sentiment analysis. These are solved problems. Off-the-shelf works.
Uncommon scenarios (consider custom): multi-step reasoning across proprietary data sources, integration with legacy systems that have no APIs, regulatory requirements that vendor platforms cannot satisfy, competitive differentiation that requires unique capability.
If your use case is unusual, off-the-shelf will frustrate you. If it is standard, custom build is vanity.
Question 2: Do you have the capabilities to maintain it?
Custom AI requires a data scientist or ML engineer, infrastructure to host and monitor models, a process for retraining on new data, security review of model inputs and outputs, and regulatory compliance documentation.
If you cannot name who does each of these, you cannot maintain a custom build.
Question 3: What is your real timeline?
Off-the-shelf: live in 8–16 weeks. Custom build: live in 6–12 months, if you already have the team. If you do not, add six months to recruit.
Most organisations underestimate this. They want custom because it sounds better, then discover their board wanted results this quarter.
The third option: guided implementation
There is a middle path. Use off-the-shelf for what it does well. Customise only where it fails your critical requirements. Have someone experienced hold the vendor to account and know when to build workarounds versus when to walk away.
This requires technical credibility to challenge vendor claims, commercial ability to negotiate scope and pricing, and pragmatism to accept good enough over perfect. Most organisations lack all three. They either accept vendor limitations or over-build.
The cost matrix (3-year view)
| Approach | Year 1 | Years 2–3 (annual) | Total | Risk |
|---|---|---|---|---|
| Off-the-shelf, straightforward | £100k | £50k | £200k | Low |
| Off-the-shelf, complex integration | £250k | £80k | £410k | Medium |
| Custom build, in-house team | £400k | £150k | £700k | High |
| Custom build + consultancy | £600k | £200k | £1m | Medium–High |
| Hybrid (off-the-shelf + custom bridges) | £200k | £75k | £350k | Medium |
The hybrid approach usually wins: fast start, targeted customisation, no massive build.
Red flags
"Our AI learns your business automatically." Translation: it does not work without six months of expensive configuration.
"We should build this ourselves." Translation: they want to hire a team and manage an empire. Check if they have done it before.
"We need to own the IP." Translation: they do not understand AI. You cannot own a vendor's model weights. You can only own your prompts and your data.
The right answer
There is no single right answer. There is only the honest answer.
If you are standard, buy. If you are unusual and have capabilities, build. If you are unusual and lack capabilities, hire someone who has done both and can tell you which side of the line you are on.
Not sure which you are? We will tell you in a 45-minute conversation — and sometimes the answer is to do neither yet, because the use case is not defined enough.