Almost every SaaS product is currently adding an AI assistant. The pattern is predictable: a chat icon in the corner, a sidebar that summarises whatever the user is looking at, maybe a "generate" button on the most-used form. This is the bolted-on era, and it is buying time while the real reorganisation is being figured out.
To be fair to the pattern: bolting-on is the correct first move. It ships in a quarter, it teaches the team what the models can do against real usage, and it satisfies the board slide. The mistake is not starting there. The mistake is mistaking the start for the strategy, and spending the next two years polishing the sidebar while someone else redesigns the job.
Why bolted-on AI is fragile
Bolted-on AI assumes the existing product is correct and AI is a layer on top. But the product was designed in a world where the user did the cognitive work. AI changes what the user comes to your product for. If the AI can do half the manual work in a competitor's product, the right design is not to bolt the AI onto the existing UI; it is to redesign the workflow around what the AI now makes free.
The economics are what make this dangerous rather than merely aesthetic. The incumbent's price is anchored to the old amount of user labor: seats, because humans sat in the tool. The AI-native entrant prices the outcome, needs a fraction of the screens, and demos in minutes because there is nothing to configure. The incumbent cannot match that price without detonating its seat-based revenue, and cannot match the product without deprecating the surface area its existing customers depend on. That squeeze, not model quality, is what makes bolted-on a six-month strategy: the timer is set by how long it takes a focused team to rebuild your most valuable workflow without your constraints.
The real question to answer
If a smart intern with perfect knowledge of your domain could do this task for the user, what changes about the product? The answer is rarely "add a chat icon." It is usually: half the screens go away, the dashboard becomes a goal, and the interface becomes a conversation about the work, not a tool for doing the work.
Run the intern test across your top five workflows and score each one honestly: does AI merely assist the user's work here (drafting, summarising, autocompleting), or does it absorb the work entirely? Assist-workflows can keep their bolted-on features indefinitely; they are genuinely fine. Absorb-workflows are the exposed flank, because in those the user never wanted the workflow at all; they wanted the outcome, and the first product to sell the outcome directly wins the account. Most products discover one or two absorb-workflows, and they are usually the ones the pricing depends on.
What to do this quarter
Pick the one workflow where AI most changes the user's job, and redesign it from scratch with AI at the centre. Keep the bolted-on features for the rest, for now. Buy yourself time to be the AI-native company in the workflows that matter most.
Structurally, treat the redesign like an internal startup, not a feature ticket: a small team, sheltered from the main roadmap, building the workflow as if the company had no existing UI to protect, with permission to end in a product that cannibalises the current one. The measure of success is uncomfortable on purpose: would a new customer, shown both versions, ever choose the old one? Ship it to a segment where the stakes are survivable, learn against real usage, then expand. The companies that navigate this era keep the bolted-on layer as a bridge while rebuilding the load-bearing workflows underneath it, one at a time, before someone else does it for them.