Why your customers don't care about AI (yet)

Your customers don’t care about AI.

They’re interested in it, sure.

They’ll try ChatGPT, and test drive any new AI-driven features you add to your app.

But few people, today, are willing to pay a premium for AI-powered software.

Microsoft only anticipates a “real revenue signal from [AI]" in the second half of the 2024 fiscal year. That’s after investing over $13 billion in OpenAI for a rumored 49% stake and a share of GPT revenue.

When the company that stands to profit from every product’s use of GPT doesn’t expect to see AI move the needle in the short run, it’s a losing battle to expect AI to make a deep change to almost any other software company’s bottom line this year.

Vendr’s data shows similar trends. We’ve purchased more than $3 billion worth of software for our customers over the past five years. Yet, despite the past year’s rapid AI feature rollouts, we have not seen any material increase in spend on AI features.

Software that added AI saw essentially flat spending growth on Vendr from Q4 2022 to Q1 2023. Same for Q1 2023 to Q2 2023. No significant change.

It’s not that people aren’t using GPT. We’re using it internally at Vendr; our customers are using it. But today, it hasn’t materially enhanced our productivity, not enough to pay extra for it.

{{cta1}}

The race to add AI for AI’s sake

“Every SaaS company in tech is hard at work at an AI strategy, for the benefit of their sales team if nothing else,” remarked tech analyst Ben Thompson recently.

Microsoft was the earliest to the game, launching GitHub Copilot in October 2021 right before OpenAI launched its APIs. Notion followed suit a year later. By this spring, the dam had broken, with Adobe, Zapier, Zendesk, Docusign, and more all launching AI features in May 2023.

AI is the new must-have feature. Much like PDF exports and spell check, AI has quickly gone from a paid, third-party add-on to something customers will quickly assume is a built-in default.

That makes it hard to stand out. ChatGPT already unleashes the majority of the large language model’s power for free. It’s an uphill battle to convince customers to pay more than free for AI features if they offer no more utility than GPT offers on its own.

Some of the earliest AI-focused startups are facing this challenge. AI-powered writing tool Jasper, for example, “saw declining user growth for four consecutive months ended in July,” reported the Wall Street Journal, resulting in their first layoffs.

AI has been commoditized only a couple of years after OpenAI’s APIs were first released. ChatGPT already covers the lowest-hanging fruit.

Build a simple GPT integration that writes for users and lets them chat with AI, and they’ll have little incentive to pay. It’s not all that hard to switch tabs to ChatGPT and do those same tasks for free.

Even worse, a more basic integration that feels less powerful than ChatGPT will make users question the product’s future. Better to have no AI at all, than a worse experience than users could get from the competition.

There’s no reason to build AI into a product today unless it offers more than existing AI implementations.

Imagining a more productive AI

That requires rethinking what AI should help users accomplish in a product.

“I think a lot of people have characterized their foray into language models as experimental, exploratory,” remarked Snowflake CEO Frank Slootman in a recent earnings call.

“People will get tired of that really, really quick,” he said: and that means both customers and investors.

The temptation is to simply add AI chat to a product, see what people do with it, and hope that scratches their itch enough that they won’t jump ship to a new competitor built around AI. A ‘minimum viable AI integration,’ so to speak.

It’s a pitfall that Apple, for one, is studiously avoiding.

“Apple execs use the phrase ‘machine learning,’ which is more popular with academics,” noted CNBC’s Kif Leswing.

Rather than fulfill the expectations that their latest Mac and iOS software would be AI-powered, or that Apple would release a GPT competitor, their 2023 WWDC was focused on “what software does for the user, such as organizing their photos, improving their typing, or filling out fields in a PDF.”

All features are powered by AI but never mentioned as such by Apple.

As Steve Blank puts it, an MVP is about “smart learning.”

A better AI MVP would be fewer AI features, focused on specific productive outcomes. That’s the Apple approach. It’s what would make stakeholders take notice. They’re not going to pay for another AI chat. They might upgrade their team’s devices for improved autocorrect and faster PDF workflows.

CFOs that approve purchases and the teams that use software to get work done care about productivity, about getting more done in less time, and about return on investment for software. Improve their ROI, and they’ll be willing to pay for features—powered by AI or not.

How to build better AI features

Things that people will pay for are those that GPT cannot do on its own.

The first level is unique inputs. GitHub CoPilot can give better code suggestions because it can access the entire codebase—which gives it far more context than a few lines of code copied into ChatGPT could. That’s what every database and big data provider is betting on. The more data they store in one place, the more valuable their AI’s inputs will be, leading to increasing software stack consolidation.

The next level is unique outputs. GPT is great at outputting text answers, and increasingly good at making tables and raw CSV outputs for opening in a spreadsheet app. Go beyond that. More unique GPT integrations use the AI’s smarts to create forms and websites and auto-edited photos, to create outputs that will make users more productive. Those provide enough incentive to use AI in the app rather than jumping over to ChatGPT. Zapier’s AI builds automated workflows, Framer’s GPT codes websites, and Fillout turns text into ready-to-fill-out forms. All are steps in this direction.

–

“The sweet spot for AI is a context where its choices are limited, transparent, and safe,” wrote Isaac Lyman for Stack Overflow's blog. “Not a decision-maker or unsupervised agent tucked away from the end user, but an interface between humans and machines.”

That’s where the more unique GPT-powered software shines. Zapier’s AI reduces the complexity of finding the right app for a task from the automation platform’s thousands of integrations. Framer’s AI reduces the time involved in framing up a basic website. Fillout’s AI reduces the time otherwise spent dragging and dropping form fields to an interface.

Small tasks, perhaps, but each with a material time savings. With the smaller problem surface exposed to AI, users can directly see how to use the AI, while developers can gain visibility into time savings and productivity gains the AI actually brings. That’s data you couldn’t get, with a more generic generative text or chat AI integration.

Beyond that are the frontiers yet to be imagined. “What AI offers us is the opportunity to profoundly augment human intelligence to make all of these outcomes of intelligence much, much better,” wrote Marc Andreessen.

Augmented human intelligence would surely bring massive productivity gains, but will only come through purposefully building AI into products in ways deeper than a cursory integration ever could.

The case for moving slower

Teams who started building AI features when GPT first came out are in a sunk-cost situation. They could abandon their dated plans for something based on newer capabilities. Or, they can stick with their initial investment, and risk a subpar AI integration when other tools ship deeper, more thoughtful AI integrations.

What you don’t want to do is ship a Newton.

One of the rare product categories where Apple was first to market, the Newton was intended to make touchscreen portable computing a reality. This was the iPad and iPhone grandfather, in late 1993.

Yet a month after launch, a Doonesbury comic strip and Newton's errant handwriting recognition that prompted it doomed the tablet. “Egg freckles?” asked the tablet, after Michael J. Doonesbury had written, hopefully, “catching on?”

The rush to market and the technology constraints of the day had led to compromises. Shortcuts. Instead of full handwriting recognition, the Newton team resorted to a dictionary of words the device could recognize. That was far too limited for the wide range of human expression, no matter how accurate the text detection.

Palm, on the other hand, reinvented handwriting. Computing wasn’t advanced enough to recognize real handwriting; the Newton had made that clear enough. A simplified alphabet that made each letter more distinct could work. Or so went the bet.

Launched as Graffiti, Palm’s simplified handwriting recognition software was quickly the best-selling app for Apple’s beleaguered Newton. When Palm launched a touchscreen device of their own in 1996—with Graffiti baked in—the writing was on the wall for Apple, which killed the Newton less than a year later.

Both Apple and Palm worked under the same hardware constraints. Yet Palm’s more customized handwriting approach won the day, versus Apple’s broader handwriting implementation.

The same may be true for AI. Everyone has access to the same GPT APIs. Everyone has already tried GPT, ChatGPT, and third-party implementations. A quickly-added GPT integration won’t win any hearts and minds. What may, what has the best chance of success, is doing something different, something more thoughtful, something created specifically for a product’s target users, something that works around AI’s constraints to aid productivity.

Everyone else has already rushed to the market. Now is the time to ship a better AI-powered tool, something that boosts workplace productivity and has stakeholders clamoring to pay for it. And when you’re buying software, don’t look for AI features—look for deeper AI integrations that do things you couldn’t in other software.