The AI strategy trap

Right now, AI is cheap. Suspiciously cheap. Ok, maybe not to the average user who thinks £19.99 a month is quite a hefty amount to pay out for a glorified search engine, but for organisations embedding it into their ways-of-working, it's the kind of cheap that should make you somewhat nervous, not relieved.

Right now, we can get access to the latest and greatest models for the price of three to four bougee coffees a month. Because of this, people are wiring these tools into everything, from customer service to content production. So, a £20 bill at the end of the month to be a "10X developer" all seems pretty reasonable.

This is because, you're not really footing the bill.

Usually, people don't sell you something for less than it costs to make. This is a terrible business model, and you wouldn't stay in business long if this was the only choice you made. So, when they do sell you something for less than it costs to make, it's not out of kindness.

This might be a stretch of an analogy, but I'm starting to view this predatory pricing by AI companies as being somewhat similar to heroin dealers. They give you "the good stuff" for a low price to get you hooked. Then, seemingly out of nowhere, the cost sky-rockets! But, by this point, you're dependent, and being without it feels impossible. So, you either pay the higher cost, or you pay whatever you can muster up for the old, dried-up, sub-quality product.

Like anything addictive, AI, or any other third party product or service, once it becomes wired so deeply into your operations, processes and ways-of-working, ripping it out and feeling the loss of it can feel like a death sentence. So, you just keep paying, whatever the cost.

If this sounds far fetched, it's not. We've all seen this playbook before. We've all seen it play out, and we've all probably felt it first-hand, with Uber.

How Uber actually worked

When Uber suddenly appeared on the market, it felt like sorcery! You just opened an app, set a pin, and you could watch a car drive to you within minutes. And, most importantly, it was cheap! It was way cheaper than getting a black cab or phoning your local taxi firm. For a while, it genuinely felt like we'd been let in on a secret as to how much we were actually being ripped off, and Uber was our way of making a stand!

The problem is, the original Uber prices were never real! They were charging us less than the rides actually cost them to provide, and cutting their losses by paying the difference with billions of pounds of invested venture capital.

Powerswitch Action estimates that back in 2015, passengers were only paying around 41% of the true cost of their trips. Investors were happy to pour money in and just burn it, because the goal was never to make a profit early on. The goal was always to capture the market and hike prices later.

And, it worked! The artificially low fares let Uber undercut and outlast the competition for long enough that, in a lot of cities, the other firms had either gone bust, or cut back their operations so much that they were no longer a viable alternative.

Then, the prices moved. Between January 2018 and July 2021, the cost of an Uber or Lyft ride in the US went up by around 92% according to CNBC. The cheap introductory pricing had done its job, it had everybody hooked on a product they now couldn't live without, so they started charging what it was really worth in order to make a profit.

This is the bit that matters. By the time the prices sky-rocketed, a lot of people no longer had a viable alternative. The local taxi ranks had closed. Public transport hadn't improved. So you were left paying whatever Uber decided to charge, because the whole point of the strategy was to make sure you had nowhere else to go!

Uber did not give you low prices because they are generous. They gave you low prices to collect your data, improve their product, and ultimately to clear the field of any competition. The cheap fare was the bait, and we all fell for it.

The AI business model looks suspiciously Uber shaped

Swap a Taxi app for a large language model, and the structure is almost identical.

The current pricing for the AI tools we use, does not reflect what it actually costs the AI companies to provide it. Training a new model and running inference at scale costs an extortionate amount in hardware, power and infrastructure. The compute consumption, the electricity, the water used for cooling the data centres, the salaries of some of the most expensive engineers on the planet, none of that is being recovered by us paying a £20 a month subscription!

The gap, like Uber, is being filled by billions of dollars in venture capital and the deep pockets of a handful of very large organisations. They are deliberately running at a loss to capture the market, to capture your data, to create the best product they can, and to make themselves indispensable to you before all their investments dry up.

The strategy for these AI companies is to get their product embedded into as many of your workflows, products and habits as possible before anybody starts seriously asking, who is going to pay for it all?

The monetary, environmental and ethical costs of AI are real, and somebody eventually has to pay them. Right now, that somebody paying for it is an investor betting on future dominance. But, make no mistake, once all that venture capital money dries up, we as the users of the product will be footing the bill!

What happens when the subsidy ends

The pattern from here is fairly predictable, because it has played out across pretty much every venture-backed industry that ever offered you something for almost nothing.

First, the free or cheap tier gets worse, gets capped, or quietly disappears. Then, the paid tiers go up in price. Then, the popular features we all rely on get moved into a more expensive tier. And, by the time this happens, we are not really customers making a free choice anymore. We are just a user with a dependency that has to keep paying, because unpicking AI from our operations is now a huge project in its own right.

We've watched taxi apps do this. We've watched food delivery apps do this. We've watched streaming services do this. They all splinter into a dozen subscriptions where the user is left facing a choice between what they're prepared to pay, and the quality they're prepared to accept. Because, the cheap introductory price was never the real price. It was just the hook!

The difference I'm seeing with AI is the depth of the dependency. We can all cancel a streaming service or a food delivery app, and we still have choices. The loss doesn't feel that devastating. But, if you're an organisation who has embedded a model they don't own and can't host themselves, into your content pipeline, support functions, code reviews, and all your internal tooling, the cost to switch or walk away will be huge!

Why this matters for accessibility

Here's where I guess I try to stop this post from being an abstract rant about economics, and try to tie it to something tangible I actually worry about.

There is a strong push right now to solve all problems, including accessibility, with AI. It can write your alt text. It can review your code. It can remediate your entire codebase without you lifting a finger. A lot of organisations find this incredibly tempting, because doing accessibility properly takes investment in people and culture. So, a tool that promises to make the problem go away automatically is an easy sell. I have written before about organisations treating accessibility as a sticky plaster rather than a cultural commitment, where they try to save face by recruiting for impossible accessibility job roles. Unfortunately for us, AI remediation looks great, but it is the ultimate sticky plaster!

Now, what if we apply the Uber model to that?

Right now, the models aren't good enough to do accessibility work. But, let's say a year down the line they are. On the surface, this looks great. Suddenly, creating accessible websites and applications is within the grasp of every developer. But, if an organisation is outsourcing all of its accessibility work to an AI tool, and that tool follows the pricing curve I've described in this post, then the cost of making your product or service usable is no longer within your control! It sits with a vendor whose prices are always going to rise year-on-year.

People with disabilities almost always pay the costs for the work or technologies which empower them. Whether it's paying for a wheelchair, or a screen reader, they're often left out of pocket. My fear is that with a reliance on AI, accessibility skills will atrophy. We'll essentially outsource all of our human-centred design to something that is not human. And when the house of cards falls, and everything's a mess, the solution will probably be another AI tool that can paper over the cracks. Which, will be another expense for those people that need it. I've already heard arguments for this, which I wrote about in my post screen readers do not need to be saved by AI.

Final thoughts

I'm not saying AI has no value, and I'm not saying every tool is a trap. Plenty of this technology is useful, and some of it is genuinely impressive.

What I am saying is that the price you see today for the most impressive models is a marketing ploy, not a real price. Like a heroin dealer, this price exists right now, because we're still not entirely reliant on what it is they're selling. It's only once we lose the ability to walk away without serious withdrawal, that we can expect to truly feel the cost of that dependency.

So, before you jump at the chance to wire AI into the very core of how your organisation works, and especially before you let it stand in for the accessibility work you should be doing properly, ask yourself a few simple questions:

• What does our budget look like if the cost of AI doubled tomorrow?
• How much work will it take to undo this if we can no longer afford to pay?
• What happens to us and our organisation if the cost of this increases 10x?
• What might we need to cut to re-direct money at increasing AI costs?
• What safeguards are we putting in place to make sure we do not become dependent on a technology we don't own?
• Will we still have the right skills in the organisation to continue to function if AI became unavailable tomorrow?

The reason for asking these questions is not just to be negative, or to throw doom-and-gloom on an exciting technology. We just need to step back and take a moment to gather our thoughts. Because, the cost of AI is always going to go up. That, we know is inevitable! And, the people who plan for that will be in a far better position than the ones who assumed the baited price was going to last forever.

I think it's easy to forget, that we don't own the skills an AI can loan us. We don't own the perceived productivity it provides us at a cost. We're simply leasing the appearance of competence, and we have no say in the terms of that lease, so it can be taken away at any moment! Anthropic abruptly turned off it's Fable 5 model for all customers outside America.

I guess the questions are… Are we happy to being totally dependent? And how much are we prepared to pay for it?

Thanks,
Craig

Post details