Build Order for Vertical SaaS AI Agents: The Four-Gate Check
By Ryan Vanshur
Open the 2027 roadmap deck and find the slide with the agents on it. Count them, then ask the person who built the slide one plain question about each, which is what that agent will be reading when it runs. In a lot of rooms the honest answer is a spreadsheet, an inbox and a PDF somebody sent in 2024, and that answer is the whole subject of this guide.
A build order for vertical SaaS AI agents settles the question before the first agent ships. Agents are about to be ordinary, and the thing they read is not, so the order we build in decides which of the two a buyer ends up comparing. The four-gate check below gives every build one thing that has to exist first and one pass test, and I'd put it on the first slide of the roadmap, ahead of the agent roster. It takes one sitting to run.
The Example Company
This guide uses one example from start to finish: PlotIQ, an invented software company for cemeteries and funeral homes. PlotIQ is fictitious, but it stands in for a real vertical SaaS company, so every gate has a concrete answer you can check against.
PlotIQ's 2027 roadmap has three agents on it. One schedules interments, one reconciles plot records, and one drafts the paperwork a family needs afterward. Today the plot maps live in scanned PDFs, the deeds sit in a shared drive, and the real status of a plot lives in the head of the groundskeeper who has worked the grounds for twenty years.
Why the Order Matters
According to Gartner's August 2025 forecast, 40% of enterprise applications will be integrated with task-specific AI agents by 2026, up from less than 5% in 2025. That counts applications rather than companies, and it's a projection. Follow it forward a few months anyway, because once every vendor in a bake-off has an agent, the buyer's question stops being whether you have one and becomes what it's reading.
McKinsey's 2025 report on the gen AI paradox, Seizing the Agentic AI Advantage, points the same way. It says about 90% of function-specific use cases remain stuck in pilot mode, and function-specific there means something like a claims workflow, not industry-specific software.
Neither source publishes a build order, and we haven't found anyone who has. The four gates here are our own operating conclusion, not a measured result. I spent a decade inside one unglamorous vertical, from pre-revenue through a private equity exit, and the instinct that a sequence beats a feature list came from watching the feature list lose.
Gate 1: Take Inventory of What You Hold
The first thing that has to exist is a list of the structured data you already hold that a funded entrant can't buy or scrape. Not the data you'd like to have, but what's already in the system because your customers ran their business through you for years.
The pass test is a written list with a named owner for every item. If the list comes back empty, that's a finding too, and it changes which agent we build first.
For PlotIQ the list starts with years of interment records, plot status, deed transfers and the maintenance history of each section. Each line gets one owner, and the plot status line gets the groundskeeper's name beside it, because today nobody else can vouch for it. That's a finding, and it's the useful kind.
Gate 2: Run One Workflow End to End
Pick the single workflow customers already run through you, and structure only what that workflow needs. Not the whole graph, one workflow.
The pass test is that someone who has never run the workflow can complete it from the structured layer alone, with no tribal knowledge borrowed from the person who usually does it. If a new hire can't pass that test, we shouldn't expect an agent to.
PlotIQ picks interment scheduling. A new hire gets a family's request and has to book a date using only the records, which means plot status, the family contact and the cemetery's own opening and closing rules. The first time she has to walk out and ask the groundskeeper, Gate 2 has failed, and the fix is to write his answer into the record so the next person, or the next agent, never has to ask. That's the test.
Gate 3: Ship One Narrow Agent
Now we ship the smallest agent that sits on Gate 2's output, and only that one.
The pass test is a number the CFO already tracks moving in the right direction. It is never logins, sessions or a usage chart, because usage is what the pilot pile is made of, and most of us already have a pile. If you can't name the CFO's number before you build the agent, you've picked the wrong workflow and Gate 2 needs another pass.
For PlotIQ the number is days from a family's request to a confirmed service date, which the finance team already reports. The scheduling agent earns its place when that number drops. A busy dashboard of agent sessions doesn't count.
Gate 4: Close the Capture Loop
Every run of the workflow feeds the structured layer. The agent's output, the human's correction and the customer's exception all land back in the record the agent read from.
The pass test is that the layer is measurably larger this month than last, without anyone scheduling that. This is the gate that makes the whole thing compound, and it's the one that gets skipped, because by the time you reach it the agent is live and we all want to move on to the next name on the roster.
At PlotIQ it looks small, and so does most of what compounds. When the groundskeeper corrects a plot's status, the correction goes into the record and not into a text message, so the next run starts from the corrected version.
Where the Check Falls Short
The strongest counter-argument I've seen comes from a venture firm's 2026 outlook, and it deserves a fair hearing. It says vertical work is inherently multi-party, so the vendor that owns the coordination between the parties may beat the vendor that owns the cleanest records.
PlotIQ has exactly that shape, with a family, a funeral home and a cemetery each holding a piece of every service. Our read is that the coordination layer is a fifth gate, not a replacement for the first four, because coordination with nothing structured underneath it is a shared inbox, and the agent still has to read something.
Two quarters for the four gates is the Guild's own estimate. It has never been measured against a cohort, so treat it as a planning number and not a benchmark.
Steal This: The Four-Gate Check
Print this, fill in the last column for the first agent on your roadmap, and bring it to the planning meeting.
| Gate | What must exist | Pass test | Your answer |
|---|---|---|---|
| 1. Inventory | A list of the structured data you hold that can't be bought or scraped | A written list, one named owner per item | |
| 2. One workflow | The structure that one workflow needs, and nothing more | A new hire completes it from the layer alone | |
| 3. Narrow agent | The smallest agent that sits on Gate 2's output | A number the CFO already tracks moves | |
| 4. Capture loop | Every run writes back to the record it read | The layer is larger this month than last |
Run the Build Order Check This Week
We run it in one sitting, with everyone who will own a gate in the room.
- List every agent on the roadmap, one per line.
- Next to each, finish the sentence "This agent reads ___" with the real system or file.
- Mark every line that reads a spreadsheet, an inbox or a document nobody owns.
- Take the first agent and run it through Gates 1 to 4 in order, one pass test at a time.
- Give every failing gate one owner and one date, and move the agents behind it down the roadmap.
Go Deeper
- Data, Not Features, the essay this guide grew out of.
- The Land Is Being Claimed, on what goes in the structured layer.
- AI Agent Governance Checklist: Five Fields That Matter, for the record you keep once an agent is running.
- Why Agentic AI Projects Fail: The Budget Review Problem, on what happens to the agents that ship without a number.
The Second Agent
Picture PlotIQ a year from now, in a bake-off against a vendor with the same three agents on its slide. The buyer, a manager at a group of cemeteries, asks what the scheduling agent is reading. PlotIQ's rep opens the plot record, shows the owner beside each field and the corrections logged since spring, and the question is answered.
The reconciliation agent ships a few weeks later, because it reads the same layer and nobody has to rebuild it. The other vendor is still explaining its connector to a shared drive.
If you want the next checklist in your inbox, subscribe to the Guild newsletter.