What happens if the AI technology changes so fast that my custom solution becomes obsolete in six months?

9 June 2026

The honest answer

Yes, a custom AI solution can become obsolete in six months. That usually happens when it is built as a one-off wrapper around a single model, with no evaluation process, no upgrade route, no contract terms for exit, and no owner inside the business. If your supplier hard-codes one model choice, locks the prompts into their platform, hides the retrieval setup, and treats maintenance as an afterthought, you have a problem.

But the better answer is this: a properly designed custom AI solution should not be a frozen artefact. It should be a working system that can absorb change. The model layer should be replaceable. The prompts should be versioned. The evaluation tests should be saved. The knowledge base should be portable. The integration logic should be separate from the user interface. The business rules should be documented in plain English. That is how you stop rapid AI change from turning into expensive rework.

This matters because AI adoption in the UK is moving quickly, but unevenly. The Office for National Statistics reported that around 25% of UK businesses were using some form of AI technology in late December 2025, up by 15 percentage points since the question was first asked in September 2023. For larger businesses with 250 employees or more, the figure was 44%. Source: Office for National Statistics, Business insights and impact on the UK economy.

The risk is real, but it is manageable. Obsolescence is not mainly a technology problem. It is a design, procurement, and governance problem.

What actually becomes obsolete?

People often say the whole AI system will become obsolete, but that is rarely precise. Different parts age at different speeds.

The biggest mistake is treating the model as the product. It is not. The product is the business outcome: fewer manual document checks, faster quote preparation, better sales follow-up, safer first draft advice, cleaner customer service triage, or quicker internal knowledge retrieval.

If the business outcome is still useful, the solution is not obsolete just because a new model has launched. It may need re-testing, re-routing, or re-tuning. That is maintenance, not failure.

What should maintenance cost?

For a UK SME, a realistic custom AI project might cost £8,000 to £25,000 for a focused pilot, £25,000 to £80,000 for a working internal system with integrations, and £80,000 to £250,000 or more for a larger multi-department implementation. Those are broad ranges because data quality, security, user count, integration depth, testing, and regulatory requirements change the work dramatically.

External UK pricing guides broadly support those bands. Fulminous lists UK AI chatbots at £8,000 to £25,000, machine learning models at £20,000 to £80,000, NLP solutions at £25,000 to £90,000, and AI consultant rates at £70 to £200 per hour. Source: Fulminous, AI Development Cost in UK. Phoenix AI Solutions puts custom AI development for UK mid-market firms at £40,000 to £250,000 plus, and says off-the-shelf or semi-custom tools solve many use cases more cheaply. Source: Phoenix AI Solutions, AI implementation cost UK.

Maintenance should not be a vague support line. Budget it explicitly. As a rule of thumb, allow 10% to 25% of the original build cost per year for monitoring, model evaluation, prompt updates, security checks, minor workflow changes, and user support. On a £30,000 build, that means roughly £3,000 to £7,500 per year. On a £100,000 system, it means £10,000 to £25,000 per year.

Compare that with off-the-shelf pricing. Microsoft 365 Copilot Business is listed in the UK from £13.80 per user per month paid yearly, excluding VAT, with a qualifying Microsoft 365 plan required. Monthly commitment pricing is listed at £19.32 per user per month excluding VAT. Source: Microsoft 365 Copilot pricing. ChatGPT Business requires at least two standard seats and API usage is billed separately. Source: OpenAI Help Centre, ChatGPT Business.

If your use case can be handled by those tools, buying custom build first is usually wasteful. If your use case depends on your private data, your approval rules, your operational workflow, and your systems, custom can still make sense, but only with a maintenance budget attached.

How UK regulation changes the answer

UK regulation is another reason not to build brittle AI. The UK approach to AI regulation is currently context-specific, risk-based, and delivered through existing regulators rather than one single AI law for every use case. GOV.UK describes the framework as pro-innovation, risk-based, coherent, proportionate, and adaptable. Source: GOV.UK, establishing a pro-innovation approach to regulating AI.

That sounds flexible, but it does not mean there are no obligations. If your AI system uses personal data, UK GDPR still matters. The ICO says its AI and data protection guidance explains how to apply UK GDPR principles to AI systems, and notes that guidance is under review following the Data (Use and Access) Act coming into law on 19 June 2025. Source: ICO, artificial intelligence guidance.

In practical terms, your custom AI system should keep privacy, auditability, human oversight, and explainability separate from any single model. If a new model launches in six months, you should be able to test whether it changes the risk profile before you switch. That is especially important in finance, legal, healthcare, HR, education, insurance, and regulated advice.

The question is not only whether the model is better. It is whether the new model is better for your use case, with your data, under your risk tolerance, with your required audit trail.

What good looks like in practice

The best UK AI deployments are not treating AI as a one-off installation. They are treating it as a capability that changes over time. NHS England announced in June 2026 that it would provide Microsoft 365 Copilot access to 505,000 clinicians and support staff, after a trial across 30,000 NHS workers found an average saving of 43 minutes of administration per person per day. The agreement includes Copilot Studio and governance through Agent 365. Source: Microsoft, NHS England Copilot rollout.

UK Power Networks is another useful example. It started with 300 Microsoft 365 Copilot licences, scaled to 1,000 by the end of its pilot, identified close to 50 use cases, and reported 96% adoption and 480% ROI. Source: Microsoft UK, UK Power Networks Copilot case study.

Lloyds Banking Group is now moving from broad Copilot adoption into agentic AI, following a roll-out of 40,000 Copilot licences with 97% active users and deployment to more than 10,000 engineers using GitHub Copilot. Source: Microsoft UK, Lloyds Banking Group AI Frontier Suite.

Those examples are larger than a typical SME, but the lesson scales down. They did not bet everything on one clever prompt. They built adoption, governance, security, use cases, and measurement around the technology.

How to protect yourself before you buy

Ask these questions before signing a custom AI contract.

• Can the model be swapped without rebuilding the whole system?
• Who owns the prompts, evaluation data, workflow documentation, and code?
• Can we export the knowledge base, vector index, logs, and configuration?
• How often will the supplier test newer models against our real examples?
• What happens if OpenAI, Anthropic, Google, Microsoft, or the chosen platform changes pricing or terms?
• What is included in monthly support, and what counts as paid change work?
• Is there a documented human review process for sensitive outputs?
• Does the contract include exit support and handover documentation?

A good supplier should welcome those questions. If they cannot answer them clearly, they may be selling a demo rather than a maintainable system.

You should also insist on an evaluation pack. That means 50 to 200 real examples from your workflow, with expected answers, failure cases, and acceptance criteria. Every major model change should be tested against that pack. Without it, you are relying on vibes, vendor benchmarks, and anecdotal staff feedback.

This is also where internal links between your AI work matter. If you have already read our guide on the biggest risks of implementing AI in a small business, this is one of those risks in practical form: weak governance makes a technical change feel like a business crisis.

When this does NOT apply

This advice does not mean every business needs a custom AI platform. Many do not. If your team mainly needs help drafting emails, summarising meetings, analysing spreadsheets, preparing first-draft marketing copy, or searching internal documents, start with established tools and a clear usage policy.

It also does not apply if your business cannot describe the process it wants to improve. AI will not fix a workflow nobody understands. If your documents are scattered, your CRM is messy, your permissions are unclear, and your team disagrees on the process, spend money on process mapping and data hygiene before custom AI.

Finally, custom AI is not right if you are trying to avoid ongoing cost. The cheaper-looking project can become the expensive one if it has no maintenance plan, no owner, and no way to adapt. In fast-changing AI markets, the budget line you remove first is often the one that keeps the system useful.

The practical answer

If you are worried your custom AI solution will be obsolete in six months, ask for an architecture and maintenance plan before you ask for a demo. The plan should show how the model can change, how performance will be measured, how data will be protected, how staff will be trained, and how you can leave if the supplier is no longer right.

The worst custom AI projects sell certainty in a market that changes every quarter. The best ones sell adaptability. They make change cheaper, safer, and more measurable than doing nothing.

If you want to explore whether custom AI makes sense for your business, start with a workflow review and a risk review. No pitch, no pressure. The right first question is not what can AI do. It is what business process is worth improving, and what would it cost if the solution could not adapt?

Is This Right For You?

This applies if you are considering a custom AI workflow, private knowledge assistant, sales automation, document processing system, customer service agent, or internal tool where the business process matters more than the model brand.

It does not apply if your need is basic content drafting, meeting notes, inbox summaries, or general productivity. In those cases, start with Microsoft 365 Copilot, ChatGPT Business, Claude Team, Gemini for Workspace, or another mainstream tool before commissioning a custom build.

Custom AI is right for you when the value sits in your process, data, approval rules, integrations, and governance. It is wrong when the value is simply access to the newest chatbot.

Frequently Asked Questions

Can a custom AI solution really become obsolete in six months?

Yes, if it is built around one model, one vendor, or one feature with no upgrade route. A well designed system should be able to swap models, update prompts, refresh retrieval, and re-test performance without a full rebuild.

Should I avoid custom AI because models change so quickly?

No. Avoid brittle custom AI. Custom work can still be valuable when it captures your workflow, data, integrations, permissions, and governance. The danger is paying for a fixed demo instead of a maintainable operational system.

How often should a custom AI system be reviewed?

Review usage and failures monthly, then run a deeper model and workflow review at least quarterly. Also review immediately after major model, pricing, security, or regulatory changes from a key provider.

What should I put in the contract to reduce obsolescence risk?

Include ownership of code, prompts, documentation, evaluation examples, data pipelines, and configuration. Add export rights, exit support, maintenance scope, model review frequency, security responsibilities, and a clear process for paid change requests.

Is off-the-shelf AI safer than custom AI?

It is safer for common productivity tasks because Microsoft, OpenAI, Google, and Anthropic absorb much of the model upgrade work. It is not always safer for specialist workflows, sensitive data, unusual integrations, or processes that need strict audit trails.

How much should I budget for ongoing support?

A sensible UK budget is 10% to 25% of the original build cost per year. For a £30,000 project, expect roughly £3,000 to £7,500 per year. If the system is regulated, business-critical, or heavily integrated, budget more.

What is the single biggest warning sign?

The biggest warning sign is a supplier who cannot explain what happens when the chosen model, API price, data policy, or vendor terms change. If the answer is vague, the risk is probably sitting with you.