Every software company says it does AI now. Four questions that check

Every enterprise software company now describes itself as an AI company. Some of them have genuinely changed their product. Most have changed their slides.

Telling those two apart is a solvable problem, and there are four questions that do most of the work.

Why AI washing happened at all

The incentive is straightforward. Software multiples expanded for anything credibly AI-adjacent, so describing an existing product in AI terms was the cheapest possible way to participate.

Nothing about that requires dishonesty. A feature that classifies support tickets was machine learning five years ago and is AI today, and both descriptions are defensible.

The pressure is real too. Google restructured its AI leadership and lost four senior researchers in the same week, knocking 4% off Alphabet. Boards read that and conclude they need an AI story of their own.

What it does mean is that the label carries almost no information, so you have to look at what changed underneath it.

The four questions that expose AI washing

Question	Repositioning looks like	Real change looks like
What did the product do before?	The same thing, renamed	Something it could not do
How is it priced?	Same seats, higher price	New line item or usage pricing
Who is delivering it?	Marketing	Engineers and forward-deployed staff
What is being reported?	Mentions on earnings calls	Revenue attributed to it

Ask what the product could not do last year. If the answer is nothing, you are looking at a rename.

IBM chose the expensive version

The clearest recent example of a company doing something structural rather than cosmetic is IBM, which folded GPT-5.6, Codex and ChatGPT Work into its consulting platform and committed to a practice of thousands of consultants.

Per its own announcement, it also joined OpenAI’s Elite partner tier and is deploying forward-deployed engineers into regulated industries.

Hiring and certifying thousands of people is not a slide change. It’s a capacity commitment that shows up in headcount and costs money whether the demand materialises or not.

Note what it isn’t, though. IBM is selling implementation of somebody else’s models, which is a services business rather than a product transformation. TechCrunch reported no financial terms and no named customers.

Timing matters as well. The move landed weeks after IBM’s chief executive conceded the company “did not move quickly enough” on AI, per Yahoo Finance, which is a strong incentive to announce something large and visible.

Microsoft shows the other pattern

Microsoft has the strongest distribution in enterprise software and still had to restructure its AI product, which tells you something important about how hard conversion is.

It merged its consumer and business Copilot apps and moved Deep Research behind a paywall, retiring podcasts, group chats and Copilot Labs along the way.

TechTimes read the consolidation as evidence of a paid-adoption problem rather than a product-complexity one, and the feature cuts support that reading.

The lesson generalises. Shipping AI features is easy, and getting people to pay for them separately is the hard part that decides whether any of this shows up in revenue.

Why so many pilots never ship

Enterprise AI pilots stall at a rate that surprises people who have only seen demos, and the reasons are boring rather than technical.

Reliability is the first. A feature that works 95% of the time is a great demo and an unacceptable production system if the remaining 5% needs a human to catch, which is precisely the gap our piece on agent reliability covers.

Compliance is the second. Regulated industries need to know where data goes, and the answer for most AI features is somebody else’s infrastructure under somebody else’s terms.

The EU AI Act’s Transparency Code has already forced changes at the model layer, which is why Anthropic began watermarking Claude output. Enterprise buyers now have to reason about obligations that reach through their vendors.

The pricing tell

How a company charges for AI reveals more than any announcement, because pricing forces a claim about value.

Bundling it into the existing seat price says the vendor doesn’t believe customers will pay extra. A separate line item says they do, and the market finds out quickly whether they were right.

Usage-based pricing is the strongest signal of all, because it only works if the feature is genuinely used. Nobody meters a feature nobody opens.

There’s a cost reason behind it too. Unlike ordinary software features, AI features carry a marginal cost per use, so bundling them into a flat seat price means the heaviest users are the least profitable.

That economics eventually forces the issue. Either usage gets metered, or the feature gets capped, or the vendor absorbs a cost that grows with adoption.

The case that repositioning is fine

There’s a reasonable defence of the companies being cynical about here, and it’s worth stating.

Enterprise software has always absorbed new technology by adding it to existing products rather than by rebuilding. That’s how cloud, mobile and analytics were adopted, and calling it repositioning misses that incremental absorption is how enterprises actually adopt anything.

Customers also benefit from AI arriving inside tools they already use, with the procurement and security review already done. A worse model inside an approved vendor often beats a better one that needs a new contract.

The watermarking example cuts this way too. TechCrunch reported that the change applied across the whole Claude platform automatically, so enterprise customers inherited compliance rather than building it, which is exactly the argument for buying AI inside an existing vendor.

And falling model prices help incumbents most. When capability is available from several suppliers at $6 to $30 per million output tokens, the scarce asset is distribution, which incumbents already have.

What to watch next

Attributed revenue is the only number that settles this, and almost nobody reports it. When a vendor breaks out AI revenue as a line rather than mentioning adoption, that’s a company confident enough to be measured.

Watch renewal rates on AI add-ons too. First-year sales prove curiosity, and renewals prove use.

There is a broader version of that argument worth reading alongside this, which our piece on the case against putting AI in everything sets out.

And watch whether the consulting bet pays. If IBM reports consulting revenue traceable to its OpenAI practice in coming quarters, the implementation model works. If it stays a headcount announcement, it was positioning after all.