The end of SaaS? It depends. (An investment thesis)

Everyone has a view on whether vibe coding, or agent-created software, will replace Software-as-a-Service. I’ve been thinking about it as an investment question rather than a technical one, and the answer is yes, maybe and no. Which of those applies depends on what the product is actually selling.

Category one: the “yes” category

These are the tools where you, or people you know, are the main user. You control the data, and a bug isn’t the end of the world. The software is the whole product, and the paid tiers are mostly packaging plus features you’ll never use.

If you can describe what you want clearly, an agent can build it for you now, and it will fit the way you work better than the off-the-shelf version does.

In about six hours I built a product development tracking system with everything I need from the big commercial offerings and nothing I don’t. It has an API, so agents can use it. It has all the views: kanban, swimlanes, timeline, calendar and prioritisation planning. It has AI embedded in it, so I can say “find all the tickets marked urgent that were assigned to John and assign them to Jane instead, but not the ones about invoicing”. It also does things I haven’t seen elsewhere, such as generating a narrative stakeholder update from whatever filters I happen to have applied.

For the “what happens when it goes wrong” crowd, and the “it’ll be messy and full of bugs I won’t know how to fix” crowd: you’re out of date. The best coding models have largely solved that problem. An LLM built it and an LLM can fix it. Want to share it with colleagues? It will deploy it, monitor it and do your DevOps. If you haven’t watched this happen recently, your objections are probably a year old.

Others in this category: CRMs, deal trackers, dashboards and BI, forms and surveys, time tracking, expense logging, wikis and knowledge bases, internal calendars, workflow and ticketing tools, internal help desks, document extraction, OCR, transcription and the long tail of small utilities. If you want to see it for yourself, the best first project is an internal tool nobody prioritised.

Category two: the “maybe” category

This one is better described as “not today, but probably soon”.

Here the code isn’t the hard part. The difficulty sits around it. Either strangers will use the software, in which case it has to be reliable and work on every device and in every inbox, or it has to connect to services that only deal with established businesses.

Apps with an audience you don’t control are the customer-facing ones rather than the internal ones: help desks, e-commerce storefronts, public booking and scheduling, email marketing, where deliverability is genuinely hard, and many mobile apps.

Gated integrations include bank account access, payroll with its HMRC integration, tax and accounting filing under Making Tax Digital, and travel booking through the global distribution systems that airlines and hotels sell through. Healthcare records belong in this group too.

Both barriers are coming down as the difficult pieces turn into things you can simply rent. The first credit card payment gateway I built, many years ago, took me a couple of months. An agent now wires up Stripe in minutes. For the moment, though, the paid product in this category still earns its fee.

Category three: the “no” category

Or at least, not any time soon.

What these products sell isn’t really software. It’s other people: the buyers, sellers, members or contacts already on the platform. Or it’s data built up from thousands of customers that you will never see. Or it’s a licence or regulatory approval you can’t get as an individual. An agent can copy the app in an afternoon. It can’t copy what makes the app worth using.

Some of these are obvious. You won’t be replacing LinkedIn unless you can persuade more than a billion people that yours is better. That might not be so hard. Convincing them all to jump ship at the same time would be. Nor Netflix, unless you’ve licensed content from the world’s studios. The same goes for marketplaces such as Airbnb, Uber, eBay and Vinted.

Less obvious are the ones where you need everyone else’s data as well as your own: PitchBook, Crunchbase, Experian, Equifax, salary benchmarking, panel data providers and expert networks. So are the security and fraud services that only work because they see traffic across thousands of customers, such as Cloudflare, Sift and CrowdStrike.

Then there’s anything that needs regulatory approval or carries liability. You might need an audit trail a regulator will accept, or insurance that costs real money. Once again, you’re not paying for the software. You’re paying for the things that don’t come free.

What this means if you’re writing cheques

Will Basecamp disappear overnight? Of course not. Plenty of people rely on it and it isn’t expensive. The fact that you could now build something that fits your needs more closely is not, on its own, a reason to do it. That has always been the build versus buy question, and cost was only ever one input to it.

But should you invest in the next version of Trello? No. The next travel booking website? Not unless you’re quick. A business whose value is its network, its data or its licence? Yes, and by the same logic, whoever gets there first will be hard to dislodge.

The question to ask about a software business is no longer how good the product is. It’s what would be left if somebody rebuilt the product in a weekend.