What happens if the AI technology changes so fast that my custom solution becomes obsolete in six months?
9 July 2026
What happens if the AI technology changes so fast that my custom solution becomes obsolete in six months?
Yes, AI can move fast enough to expose a weak custom build within six months. The honest answer is that the risk is real, but it is manageable. You reduce it by paying for architecture, testing, documentation, and maintenance, not just a polished demo. For most UK SMEs, the safest custom AI systems are designed so the model can change without rebuilding the whole business process.
The honest answer
Yes, your custom AI solution can become obsolete in six months. That is the uncomfortable truth. It usually happens when the supplier builds a clever wrapper around one model, hides the prompts, stores your knowledge base in a proprietary format, skips evaluation, and treats support as a separate afterthought.
But that does not mean custom AI is a bad investment. It means the architecture matters. A custom AI system should be more like a managed operating capability than a one-off software artefact. The model should be swappable. The prompts should be versioned. The test set should be saved. The data should be exportable. The business rules should be documented. The risk controls should sit outside the model, not inside a vendor black box.
The market is moving quickly enough for this to matter. The Office for National Statistics reported that AI was adopted by 9% of UK firms in 2023, with planned adoption projected to rise to 22% in 2024. It also found that 39% of firms saw difficulty identifying use cases as the most common barrier, followed by cost at 21% and AI expertise at 16%. Source: ONS, management practices and AI adoption in UK firms.
That tells you something important. The danger is not just that the technology changes. The danger is that businesses buy before they understand the process, the use case, the support model, and the exit route.
What actually becomes obsolete?
People talk as if the whole AI system expires at once. That is rarely true. Different parts age at different speeds.
| Part of the solution | Six-month risk | What to do |
|---|---|---|
| Underlying model | High | Use an abstraction layer so OpenAI, Anthropic, Google, Microsoft, local models, or specialist models can be tested and switched. |
| Prompt templates | Medium | Version prompts and retest them against real examples after major model changes. |
| Knowledge base | Medium | Keep source documents, chunking rules, metadata, and retrieval settings documented and exportable. |
| Workflow logic | Low to medium | Design around the job to be done, not around a vendor feature that may disappear. |
| Data connectors | Medium | Document CRM, finance, ticketing, email, and file storage connections. Avoid custom integrations nobody else can maintain. |
| Governance controls | Low | Keep permissions, audit logs, approvals, and escalation rules separate from the model. |
The underlying model is the fastest-moving part. The business process is usually slower. A quote preparation workflow, customer enquiry triage process, internal policy search system, or invoice checking process will still matter in six months. What changes is the best way to power it.
If your supplier has treated the model as the whole product, you are exposed. If they have treated the model as one replaceable component, you have options.
Model retirement is not theoretical
This is not a made-up risk. Model providers retire, rename, replace, and restrict models. OpenAI says it regularly retires older models as it launches safer and more capable ones. Its API deprecation policy gives at least six months notice for generally available models, at least three months for specialist variants, and much shorter notice for preview models, sometimes around two weeks. Source: OpenAI API deprecations.
Microsoft is explicit too. Its Azure Foundry OpenAI documentation says generally available models have retirement dates set programmatically at launch to 18 months out, with lifecycle states visible through the Models API. Source: Microsoft Foundry model retirements.
For a UK business, that means you should assume change, not hope it will stop. If your custom system depends on a preview model, a narrow API feature, a specific prompt behaviour, or an undocumented vendor setting, six months is a long time. If it depends on a tested workflow with a replaceable model layer, six months is normal maintenance.
The practical question is simple: if the model disappeared next quarter, would you need a rebuild or a controlled migration? Those are very different commercial risks.
What should this cost to maintain?
For a UK SME, a realistic custom AI pilot often costs £8,000 to £25,000. A working internal system with CRM, document, email, or workflow integrations commonly sits between £25,000 and £80,000. Larger multi-department systems can reach £80,000 to £250,000 or more, especially where security, compliance, data migration, and change management are serious requirements.
The maintenance budget should be visible from day one. As a working rule, allow 10% to 25% of the original build cost per year for monitoring, model evaluations, prompt updates, minor workflow changes, supplier updates, security checks, and user support. On a £30,000 build, that means £3,000 to £7,500 per year. On an £80,000 build, it means £8,000 to £20,000 per year.
If the system is business-critical, budget monthly. A typical SME support retainer might be £750 to £2,500 per month. A more complex system with live customer impact, integrations, and formal service levels may need £3,000 to £8,000 per month. If a supplier tells you maintenance is not needed, ask what happens when the model changes, the API price changes, your CRM schema changes, staff behaviour changes, or the ICO guidance shifts.
The cheapest project is not always the cheapest system. A £15,000 build with no tests, no documentation, and no support can become more expensive than a £30,000 build that is maintainable.
How UK regulation changes the answer
UK regulation makes brittleness more dangerous. If your custom AI system uses personal data, UK GDPR still applies. The ICO says its AI guidance helps organisations apply UK GDPR principles to AI systems and provides a risk toolkit for assessing risks to individual rights and freedoms. The ICO also notes that the Data (Use and Access) Act 2025 received Royal Assent on 19 June 2025. Source: ICO artificial intelligence guidance.
The UK government is also pushing hard on AI adoption. Its AI Opportunities Action Plan describes the UK as the third largest AI market in the world and names Google DeepMind, ARM, Wayve, OpenAI, Anthropic, Microsoft, and Meta AI as part of the UK AI ecosystem. Source: GOV.UK AI Opportunities Action Plan.
That combination matters. Adoption pressure is rising, but obligations do not disappear. If you swap a model, you may change data residency, output behaviour, explainability, logging, retention, accuracy, or bias risk. In HR, finance, healthcare, education, legal, insurance, recruitment, and regulated advice, that is not a technical detail. It is a governance issue.
A future-proof build separates privacy, permissions, human review, audit logs, and escalation rules from the model. You should be able to test a new model before release, compare outputs, review failures, and decide whether the risk profile has changed.
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, code, evaluation examples, workflow documentation, and configuration?
- Can we export the knowledge base, source documents, vector data, logs, and user settings?
- How often will you test newer models against our real examples?
- What happens if OpenAI, Anthropic, Google, Microsoft, or the chosen platform changes pricing, model availability, or terms?
- What is included in monthly support, and what becomes paid change work?
- Do we get a handover pack if we leave?
- How are sensitive outputs reviewed by a human before they reach customers or staff?
A competent supplier should answer those questions clearly. If they cannot, they may be selling a demo rather than a maintainable business system.
You should also ask for an evaluation pack. For most SME systems, that means 50 to 200 real examples from your workflow, including good cases, edge cases, unacceptable outputs, expected answers, and pass or fail criteria. Every major model change should be run against that pack before release. Without it, you are relying on vendor claims and staff impressions.
Finally, insist on plain-English documentation. If your business cannot understand what has been built, you do not really own it.
When this does NOT apply
This advice does not mean every business needs a custom AI solution. Many do not. If your team mainly needs to draft emails, summarise meetings, analyse documents, create marketing drafts, search internal files, or tidy spreadsheets, start with established tools and a usage policy.
It also does not apply if your business cannot describe the process it wants to improve. AI will not rescue a workflow nobody understands. If your CRM is messy, your documents are scattered, your permissions are unclear, and your team disagrees on the correct process, spend money on process mapping and data hygiene before custom AI.
Custom AI is also a bad fit if you want a one-off build with no ongoing cost. In a fast-changing market, the maintenance line is not optional. It is the thing that stops the system becoming stale, risky, or quietly unused.
If you have no internal owner, no appetite for change management, and no budget beyond the initial invoice, do not buy custom AI yet. Use SaaS first. Learn where the value is. Then consider custom work when the problem is valuable enough to maintain.
The practical answer
If you are worried your custom AI solution will be obsolete in six months, do not start by asking for a demo. Ask for the architecture, the maintenance plan, the evaluation plan, the data export route, and the contract exit terms.
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 easier to measure.
A good build should answer five questions before work starts: what business process is worth improving, what outcome will prove value, what data is required, what happens when the model changes, and who owns the system after launch?
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 process is worth improving, and what would it cost if the solution could not adapt?
Is This Right For You?
This advice is right for you if you are considering a custom AI workflow, an internal agent, a customer support assistant, document automation, sales operations automation, or a private knowledge system for a UK business.
It is especially relevant if the project touches customer data, internal documents, regulated decisions, your CRM, finance records, HR processes, or operational workflows that staff rely on every week.
It is not right for you if you only need general writing help, meeting summaries, spreadsheet analysis, or basic search across documents. In those cases, Microsoft Copilot, ChatGPT Team, Claude Team, Gemini for Workspace, or a specialist SaaS tool may be the better first step.
Frequently Asked Questions
Can a custom AI solution really become obsolete in six months?
Yes. It can happen if the system depends on one model, one vendor feature, one prompt behaviour, or one undocumented integration. It is much less likely when the model is replaceable, tests are saved, data is portable, and support is built into the contract.
Should I avoid custom AI because models change so quickly?
No, but you should avoid brittle custom AI. If the workflow is valuable and specific to your business, custom can make sense. If the need is generic, use off-the-shelf tools first.
How often should a custom AI system be reviewed?
For most SME systems, review performance monthly and run a deeper model and workflow review every quarter. Business-critical or customer-facing systems need tighter monitoring and formal incident handling.
What should I put in the contract to reduce obsolescence risk?
Include ownership of code, prompts, evaluation data, documentation, configurations, source files, and logs. Add model migration support, exit assistance, support response times, data export rights, and a clear definition of maintenance versus chargeable change work.
Is off-the-shelf AI safer than custom AI?
Sometimes. Microsoft Copilot, ChatGPT Team, Claude Team, Gemini for Workspace, and specialist SaaS products can be safer for common tasks because they are maintained by large vendors. Custom AI is better when the workflow, data, risk controls, or competitive advantage are specific to your business.
How much should I budget for ongoing support?
Budget 10% to 25% of the original build cost per year. For many UK SMEs, that means £750 to £2,500 per month for a modest system, and more for customer-facing or regulated workflows.
What is the biggest warning sign?
The biggest warning sign is a supplier who cannot explain how the model will be changed, how the system will be tested, who owns the assets, and how you leave. If the answer is trust us, walk away.