SaaS is not dead. It is being repriced. The market is moving value from tools that help a human work to software that does the work itself. If you sell software, that distinction decides whether 2026 is your best year or your last one.
You have seen the line everywhere. Founder feeds, investor memos, conference stages from Lisbon to Helsinki. Public software valuations have split in two: legacy SaaS names repriced sharply, some to their lowest revenue multiples in years, while AI-native software trades at several times that on the same exchanges, watched by the same analysts.
Read quickly, that gap looks like decline. Read carefully, it is an opening. The market is not writing software off. It is telling you exactly where the next decade of value sits. For founders willing to build on the right side of the gap, this is the largest window we have seen in years.
The user of your software is no longer a person
For twenty-five years, every SaaS product assumed a human in the chair. Someone logging in, clicking through screens, reading dashboards. That assumption is breaking.
Increasingly, the user of your software is another piece of software acting on a person's behalf. Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% in 2025. That is not a trend line. That is a rewrite of who your customer's customer is.
We see this in our own pipeline. When we look at how prospects evaluate MiraReach, the first pass is often not a human opening a browser. It is an agent pulling our docs, checking our pricing page, and summarising us for a decision-maker who never visits the site. If your product cannot be read, parsed, and summarised by a machine, you are invisible to a growing share of the buying process.
If an agent cannot call your product, it cannot buy it
Graphical interfaces exist because humans need to see things. Agents do not need to see. They need to call. So the interface layer of software is being rebuilt for a non-human user.
The clearest signal is the Model Context Protocol, the emerging standard that lets agents connect to tools and data. Anthropic, which introduced it, reports it going from near-zero to around 97 million monthly SDK downloads and thousands of public servers in barely eighteen months. Stripe rebuilt its commerce stack so agents could transact directly.
The takeaway for founders is blunt. If an agent cannot reach your product, it cannot buy, use, or recommend it. The front door is no longer a login screen. It is an API, a schema, a machine-readable contract.
This matters for outbound too. When your prospect's team runs an agent to shortlist vendors, your positioning has to survive being summarised in three sentences by a model that has never seen your demo. Most cold email copy fails that test. It is written for a human skimming on a phone, not a model extracting a value proposition.
Per-seat pricing is quietly dying
Per-seat pricing was the financial engine of SaaS. It assumed one thing: a human occupying the seat. When an agent does the work of ten people, ten seats do not follow.
The model is already changing. Roughly 77% of the largest software companies now use some form of consumption pricing, according to Metronome's 2025 pricing study. Cursor, the AI coding tool, scaled to billions in annual revenue with a team a fraction of the size of its legacy competitors. It did not do that by selling seats.
If you run a small sales team, this has a practical consequence. Your buyers are being asked to justify software spend against outcomes, not headcount. The pitch that worked in 2022 ("this saves each rep two hours a day") is being replaced by ("this closes the loop on X, and here is the number").
We have felt this on our own side. When we talk to founders about MiraReach, the question is rarely "how many seats?" It is "what does this actually produce?" That is a harder question to answer, and a better one.
What this means if you are running outbound right now
You do not need to rebuild your product this quarter. But you do need to notice which side of the repricing you are on.
- Your copy has to survive a model. If an agent summarises your cold email into one line, does that line still sell? Test it. Paste your last sequence into a model and ask it to summarise the offer in ten words. If the result is mush, rewrite.
- Your pricing page has to be machine-readable. Clear tiers, clear outcomes, no "contact us for pricing" black holes. Agents cannot fill in a form.
- Your ICP has to be sharper, not wider. When software does more of the work, the cost of a bad-fit prospect goes up, not down. Volume plays are getting punished by deliverability rules anyway, as we wrote about in Google and Microsoft's DMARC volume loophole closure.
- Your proof has to be specific. "Saves time" is dead. "Cut SDR research time from 40 minutes to 6" is alive.
None of this is exotic. It is the same discipline good sales teams have always had, applied to a market that is now scoring them on it.
What does not work, and we tried it
We spent a chunk of last year testing whether AI-drafted emails could go out without a human pressing send. They cannot. Not at the quality bar that keeps a domain healthy and a prospect interested. The drafts got good. The judgment did not. Every time we let the system send unsupervised, reply rates dropped and unsubscribe rates climbed.
So we kept the human in the loop. That is not a limitation of the technology. It is a limitation of the market. Buyers can tell when nobody read the email before it landed. And in a repriced market, the cost of that tell is higher than it used to be.
If you are evaluating tools in this space, the same test applies. Ask any vendor how their system handles the moment before send. If the answer is "it just sends," you have your answer. We wrote about how to spot inflated vendor claims in our breakdown of the 20-hours-saved claim.
What we would do next
Pick one thing this week. Rewrite your top-performing cold email so a model could summarise it in one clean sentence. Then check whether your pricing page survives the same test. Those two moves put you on the right side of the repricing without touching your product roadmap.
If you want to see how MiraReach handles prospect scoring, inbox checks, and human-approved sending, give MiraReach a try. No auto-send. Ever.
— Mira