How to Agentify Your SaaS Before Someone Does It for You

To agentify your SaaS means taking the work your customers still do by hand inside your product and giving it to AI that does that work for each customer, with your product's own tools and that customer's permissions. Adding a chat box that answers questions about your features doesn't count. An agentified product acts, keeps acting until the job is done, and is trusted enough that nobody checks its work.

That bar comes from one of the clearest talks on SaaS and AI this year. At MicroConf Europe 2026, Rob Walling gave a talk called The Only SaaS Moats Left in 2026, built with Einar Vollset of Discretion Capital from a month of reading reports by think tanks, private equity groups and analysts. Its bluntest line is aimed at software that sits in front of a human doing a job: "If you don't agentify your software, build some type of agent on top of it, I think you're going to zero in the next couple years."

Most of the writing on this topic stops at "add AI features." The talk is more useful than that, because it gives you a test. It lists the moats that still protect a software business when anyone can clone the code, and several of them are things an agent can give you. So this post doesn't ask how to add AI to your product. It asks how to agentify your SaaS so that it ends up with more of those moats. Watch the talk if you can. Here is the short version and what to do about it.

The talk in two minutes

Walling sorts AI's threats to a SaaS business into two groups. You can get the full framework as a field guide at microconf.com/moats.

Supply shocks happen when there is suddenly more software like yours:

• Your customer builds it themselves. "They don't need your 50 features, they only need the three that they use."
• Competitors reach feature parity in months, not years.
• Open source moves faster, and non-technical buyers can now deploy it with an AI assistant.
• The platform you integrate with builds your feature for free.
• Agencies sell custom software at SaaS prices.
• The model itself absorbs your product.

Demand shocks happen when there are fewer customers, or fewer reasons to pay you:

• Your customers' businesses get wiped out by AI.
• Your seats go away because agents do the work people used to do.
• The thing you manage stops mattering.
• The human who used your software is replaced, and your software goes with them.

His estimate is that 10 to 20% of SaaS companies will decline or go to zero. The talk doesn't say SaaS is dead. It says the old moats are failing: unique features, clever code, UX familiarity and integration count are all cheaper to copy than they used to be.

The moats he thinks still hold are these eight:

1. Buyers know and trust you (brand, partners, champions).
2. You are a system of record.
3. You are allowed to act, not just advise.
4. Customers don't check your work.
5. Someone official already approved you.
6. Everyone is already in the software.
7. Every customer makes the product smarter.
8. It is hard to keep running.

Read those as a product team and something stands out. Moats 3, 4 and 7, and to a degree 2 and 8, are things an agent built into your product can earn for you. That is the real reason to agentify.

The public markets already moved

The talk points to two earnings reports a day apart in August. HubSpot beat on revenue and earnings and fell 19% the next day, according to TIKR, its steepest single-session drop on record. The reported trigger was net customer additions coming in below guidance. Walling reads it as the market doubting HubSpot's inbound marketing engine in a world of AI search.

Atlassian went the other way. Its shares surged more than 35% after a quarter in which usage of its MCP server and Teamwork Graph command-line tool more than doubled to over 1 million monthly active users, and MCP calls rose more than 400% from the previous quarter. Atlassian's software is becoming something agents work through, and Jira already had several of Walling's moats.

You don't need to be public for this to matter. Walling's point is that private equity follows public multiples, and acquirers are now asking one question: will this company still be here in three to five years?

How to agentify your SaaS, step by step

1. Find the job your customers still do by hand

Start with the work, not the model. Every product has a gap between what it can do and what the customer actually gets done: the import nobody finished, the integration waiting on the customer's IT team, the settings that drifted, the new hire who never got set up.

In a learning platform, the customer doesn't want to learn your enrollment screens. They want every new hire trained before their start date. That outcome, not a feature, is what you are agentifying. If you can't name the job in one sentence, you aren't ready to build anything.

This is also your defense against the first attack. A customer can rebuild the three features they use. They are much less likely to rebuild a job your product does for them every week.

2. Make your product something an agent can act through

An agent can only do what your product lets it do. That means an API or an MCP server that covers the actions behind the job, and authorization that lets the agent act as the specific user, with their permissions and nothing more. Atlassian's numbers show where usage goes once a product is reachable this way.

Don't expose every endpoint at once. Expose the actions that the job from step 1 needs. We wrote about what to expose to your customers' AI agents and the tradeoffs in more detail.

3. Act, don't advise

This is Walling's third moat, and it is where most "AI features" fall short. A tool that tells a clerk which payment to make is a reporting tool. A tool that makes the payment is a different kind of business. His examples include Stripe Connect, which moves the money, and Churnkey, which doesn't just predict churn but steps into the cancellation flow and makes the offer.

Apply the same test to your agent. If it can only explain how to fix the stalled integration, you have built better documentation. If it can fix it, retry the import, and message the customer's admin for the missing key, it is doing the job.

4. Own the job until it is done, so nobody checks

The fourth moat is the Gusto moat. Walling has used Gusto for over a decade and never once checked its work. That trust is earned by being right every time, for years, on a job the customer doesn't want to think about.

An agent earns it the same way, and that rules out one-shot conversations. A job like "get every new customer live" stays open for weeks. The agent needs to remember where each account stands, notice when something slips, act again without being asked, and bring in a person when it hits something it shouldn't decide alone. We call this an owned job: a job the AI holds for each customer, with a state you can read (say, 9 of 12 new accounts live) and a log of what it did. Our guide to AI agents for SaaS covers the requirements in depth.

5. Learn across customers, not just within one

Moat seven only counts, Walling says, if every customer makes the product better for every other customer, customers can take their data but not what you learned from it, and the learning keeps going. Stripe Radar is his example: no single merchant could build it alone.

An agent running the same job across hundreds of accounts sees patterns no single customer sees: which integration step fails most often, which instruction confuses people, which fix works. Be honest about how far along you are. Most agents today learn through humans reviewing their runs and updating their instructions, not by retraining themselves. That still counts if the improvement reaches every customer.

6. Price the work, not the seat

Seats are the demand shock that hits first. If an agent does what three people used to do, a seat-priced product earns less as it gets better. Zendesk moved early: in August 2024 it announced outcome-based pricing where "customers will only incur costs for issues that are resolved autonomously by AI." Bain's Technology Report 2025 puts it plainly: "The fundamental shift is to stop charging for access and start charging for work done."

The upside is the flip side of Walling's receptionist example. A customer paying $50 a month for software will pay several times that for software plus the work done. A job your product owns, priced as a subscription, is new revenue rather than a replacement for seats. See outcome-based pricing for software companies for how to set that price.

What agentifying won't fix

To be fair to the talk, an agent doesn't earn every moat. It won't build a partner ecosystem like the thousands of agencies that recommend Shopify. It won't get you board approval or carrier registration. It won't create the network effect of a whole swim league that has to switch at once. If your business has none of those, an agent is still the best move available, but it isn't the whole defense.

And agentifying doesn't make you safe from the first attack by itself. If your agent is just a thin layer over a public API, someone else can build the same layer. The defense is the combination: an agent that acts in your product, owns a job your customers would rather not think about, and gets better across your customer base.

The hard part is running it

Walling's last moat may be the most underrated one: "Is it hard to build or is it hard to keep running?" Scraping isn't hard to build, but keeping it unblocked is a full-time job, which is why ScrapingBee and Cloudflare keep their customers.

An agent working for every one of your customers is the same kind of problem. Building a demo takes a week. Running it in production means acting as each user with the right permissions, controlling what each job costs, seeing every step it took when a customer asks why, keeping jobs alive for weeks, and bringing in your team at the right moments. That burden is real, and whoever carries it well holds a moat.

That is the part we built Aimdoc to carry. It connects to your product through its MCP or API, takes on a job like getting every new customer live, works it for each customer with metered and monitored runs, and lets you sell that work through your own Stripe account. Your team decides what the job is. Aimdoc does the work and keeps it running.

FAQ

What does it mean to agentify software?

To agentify software is to give AI the ability to do the work users currently do in the product, for them, using the product's own actions and the user's permissions. Rob Walling made it the central warning of his MicroConf Europe 2026 talk. A chat assistant that only answers questions doesn't qualify.

Do I need an MCP server to agentify my SaaS?

You need some way for an agent to take the actions behind the job, scoped to each user. An MCP server is the most common way to do that today, and an existing API can serve the same purpose. What matters is coverage of the job's actions and correct per-user authorization.

Will AI replace SaaS?

Not wholesale. Walling estimates 10 to 20% of SaaS companies will decline or go to zero, which is a lot but far from all. The products most at risk sell a screen where a person does a task. The ones that last will deliver the result of that task. We make the longer argument in SaaS is not dead.

Where should a small SaaS company start?

Pick one job your customers repeatedly fail to finish in your product, usually something in onboarding or setup, and agentify that end to end before adding anything else. One job owned well does more for retention and defensibility than ten AI features.

If you watched Walling's talk and found your product on the wrong side of "can your software act, or only advise?", start with one job. See how Aimdoc takes on that work, or start free.