How to Invest in AI Without Regretting It Later

Every few weeks, another model appears that is faster, cheaper or more capable than what came before it (and every so often one quietly disappears again, pulled not long after launch over safety concerns, no names, of course).

Features that recently required specialist teams and substantial custom development suddenly become available through a standard API or enterprise product. For companies considering a serious investment in AI, this creates an uncomfortable question:

What if the system we build today is already outdated in six months?

The concern is legitimate. Some prompts, workarounds and processing steps being built today will become unnecessary. Providers will change, models will improve and current technical limitations will disappear.

But that does not mean companies should wait. The better question is not how to build a system that never changes, but how to build one that becomes more valuable when the technology improves.

A resilient AI investment concentrates long-term value in the parts that remain useful: the process, the data, the business rules, the integrations, the quality framework and the operational experience. The rapidly changing model layer remains replaceable.

Two ways to get it wrong

Companies usually respond to technological uncertainty in one of two ways. Both are understandable. Both can be expensive.

Waiting for the technology to settle

Some companies postpone productive projects because the next generation of models will probably be better. They are almost certainly right. But there will always be another generation after that.

While the company waits, it gains no experience with real users, leaves data-access and integration questions unresolved, creates no representative evaluation cases and builds no operational capability. Meanwhile, more active competitors accumulate process knowledge, user feedback and operational data.

Waiting reduces short-term technical risk while increasing the risk of falling behind organisationally.

Overcommitting to today’s technology

Other companies make the opposite mistake. They invest heavily in one model or provider, proprietary features, model-specific customisation or complex workarounds for current limitations. When the technology improves, too much of the investment must be rebuilt.

This is not only vendor lock-in. A company can stay with the same provider and still see much of its custom implementation become unnecessary when a new model generation arrives. The deeper risk is concentrating too much value in the part of the system that changes fastest.

Past experiments can also distort future decisions. A failed test may reveal something important about the workflow, or merely the limitations of the model available at the time. Companies need to separate what they learned about the business problem from what they learned about a particular model.

The better position lies between these two extremes:

Build something valuable now, but keep the rapidly changing technology layer replaceable.

A cautionary example: the fine-tuned chatbot

Consider a company that invests heavily in fine-tuning a customer-service chatbot on previous conversations, FAQs and preferred-response examples. At the time, this may be a reasonable technical decision.

The investment becomes fragile when too much of the project’s value sits inside that particular model. As general-purpose models improve, behaviours that once required custom training may be achieved through clearer instructions, runtime examples, retrieval from an up-to-date knowledge base and access to customer or product systems.

If the project created little outside the model, the company is left with no strong integrations, reusable evaluation framework, maintainable knowledge layer or robust escalation process. A more resilient approach would have invested in those assets first, because they remain useful even when the model changes.

Fine-tuning can still be valuable for specialised language, narrow behaviours, highly consistent outputs or efficient smaller models at scale. The lesson is not “never fine-tune.” It is:

Do not concentrate the entire business value of a system inside a model-specific intervention that may become unnecessary.

The three layers of an AI investment

A useful way to assess an AI project is to divide the investment into three layers: a durable foundation of process, data, integrations, criteria and governance; an adaptable middle of workflows, retrieval, validation and human review; and a volatile top of models, providers and workarounds for today’s limitations.

The principle is simple:

Invest heavily in the problem, the process and the integration. Invest cautiously in the temporary characteristics of today’s technology.

A more resilient example: an AI-powered audit platform

We built an automated audit platform for a large international insurer to evaluate websites and digital channels against defined brand, content and sustainability criteria. Manual auditing was possible, but it was slow, difficult to standardise and expensive to scale.

The objective was not simply to ask a language model whether a page appeared compliant. It was to create an operational audit capability.

How the platform works

The platform collects the relevant pages and channels, captures screenshots and supporting evidence, determines which checks apply, runs deterministic checks where interpretation is unnecessary and uses AI where contextual or visual judgement is required. It then aggregates the findings, stores the evidence and presents completed audits in a reviewable interface.

The key distinction is:

The AI model is one component inside the audit system. It is not the audit system itself.

What may change and what remains

Today, reliable AI evaluation may require carefully engineered prompts, several model calls and model-specific orchestration. A future model may handle more criteria at once, understand text and images together and produce more reliable structured outputs. Some of the current implementation may therefore become simpler or disappear.

That does not make the original investment a mistake. Even a model that is ten times better will not know which channels to audit, how to collect them, which criteria matter, which checks should be deterministic, what evidence to retain or how results should be structured, reviewed and compared over time.

The evaluation model may change. The organisation’s ability to conduct scalable, consistent and traceable audits remains.

The ten-times-better test

Before investing heavily in any component, ask:

If the underlying AI became ten times better tomorrow, would we still need this?

Applied to the audit platform:

Temporary components can still be necessary. A system must work with the technology available today, not with an imagined future model. But components likely to disappear should be kept light, modular and replaceable.

If almost everything in a proposed project would become unnecessary after a major improvement in the underlying model, the company may be investing in a temporary limitation rather than a lasting business capability.

How we approach this at Voviva

At Voviva, we begin with the process: what information and evidence it requires, which decisions can be deterministic, where contextual judgement is genuinely needed, how quality should be measured and how the result enters the operational workflow. We then design the system so the AI layer can evolve without taking the entire capability with it.

Conclusion

Companies do not need to choose between waiting and making a risky long-term bet on today’s technology. They can invest now while remaining flexible by distinguishing between what is temporary and what is durable.

Models will change, and some orchestration, fine-tuning and technical workarounds will become unnecessary. But the process, integrations, criteria, evidence, governance and operational capability can remain. When a better model arrives, a well-designed system should receive an upgrade, not an obituary. That is the difference between investing in an AI model and investing in a lasting business capability.