If you’ve been in the orbit of ‘tech’ at any point in the last six months, you’re likely to have caught a bit of the AI fever. From prompt-generated hallucinations to processing legal documents, AI has made great strides. The sandbox of experimentation has already begun to spin out some serious applications for business and industry.

Quite rightly, every problem is looking for an AI solution and every AI enthusiast is looking for problems to solve. It’s an exciting time for entrepreneurship, and it’s likely that many future unicorns will be born during this wave of tech-optimism.

Unfortunately, we’ve also seen early success stories like Jasper, Mutiny and Scale announcing layoffs, as founders are forced to hone that enthusiasm into competitive businesses in an increasingly crowded market. If hype spreads like fire, then 2023 is a relatively low-oxygen environment for startups and a certain amount of pragmatism is required. The internet has not been particularly impressed by GPT-generated articles, OpenAI’s user activity is in decline, and even Midjourney’s moment seems to have passed.

Following this theme, it is even more important that founders scrutinize the value of what they are setting out to build. While generative AI has the potential to reduce some tasks down to a tiny fraction of the original workload, decreasing operating costs only matters if you are making something that people want.

The wrench in the machine for valuation

An AI model could, in theory, look at the past performance of a pool of startups, and derive a set of trends which should be positive signals for the future. In fact, it’s not a new idea; there are a few funds which do exactly that. The caveat is that you need reliable, measurable metrics to produce a model that can offer much confidence. For this reason they tend to focus on Series B and beyond, the point at which it becomes easier to model the future of a startup as most of the fundamental assumptions will have been proven in the journey to that point.

In fact, the removal of those assumptions is why a model can work at all. You cannot ask a computer to predict the future.

This is why the process of valuation for a startup depends on first agreeing a reasonable shared vision of the future. You look at the potential upside in that scenario, and discount it for the perceived risk to get there. Once a founder and an investor can find agreement on that vision, determining the valuation based on that scenario becomes a more objective calculation.

Valuation will always fundamentally be an agreement between people, built on reasoning and understanding. It requires belief, as well as proof.

Sophistication is not always an improvement

Targeting an AI model at determining the appropriate price for a deal based on analysis of market activity is an attractive idea. Making the process of finding similar rounds faster and easier does cut some of the headache involved in determining market-appropriate deal terms.

The danger is that many will confuse this with greater accuracy, when the problems with comparables remain: innovative startups have few direct comparisons, the terms of those fundraising rounds are never fully known, and revenue multiples are still incredibly crude. Using a bigger sample does not necessarily yield a more useful average.

Most of all, it would be heavily procyclical. If you consider the role that lazy practices played in the 2020-22 startup bubble, as investors relied heavily on multiples for pricing, now imagine that applied with all the efficiency of an AI model. It is likely that valuations would have climbed much higher before the market collapsed, with a fall that was much more painful for all involved.

Outliers and intangibles

Any solid valuation methodology (ours included) is built around the idea of using multiple perspectives on valuation to provide a more comprehensive view on future potential. That combination of perspectives should focus at least as much on the properties of the individual startup as it does on the wider fundraising market.

That said, it’s certainly possible that an AI model could provide a useful additional perspective, you just have to be careful about the bias it introduces. For example, if you train a model to identify positive traits in successful founders, and use it to screen the startups you meet in future, you are likely to end up optimizing for an average which indicates ‘competent leadership’. It could identify a good steward for a stable company, not a driven, tenacious founder with real vision.

What a model like that would miss is not the 90% overlap of traits in successful founders, but the crucial 10% which drove them to build that particular business. Early-stage venture capital is a business of looking for outliers, not consensus, precisely for this reason. Success stories often return the whole fund, but you are far more likely to lose all of your money on an investment.

AI can help an investor take more ‘shots on goal’

While AI models will struggle with the reasoning required to assess the unique scenarios presented by early-stage companies, it is likely to have a positive impact on the industry overall.

Already, investors are looking at how AI can help them to streamline their processes for screening and evaluating incoming pitches. This frees up time to ensure that qualifying pitches (by which I mostly mean companies in the right sector and stage) can get more attention and diligence.

Indeed, investors should pursue efficiency and automation in every step of the process, right up until they need to sit down and consider a pitch. At that point, a quiet room with a comfortable chair is likely to be more helpful than an AI model.

A few years ago, data nerd and investor Jared Heyman published an article on the machine learning framework his fund uses to determine the investability of YCombinator startups, On Rebel Theorem. This might be a good example of adding useful perspective, in that it is applied to a group of startups that are already pre-screened via their acceptance into the program.