RealPact (YC Summer 2026): teardown vs XMarts

In one line: RealPact sells voice-driven AI agents to real estate brokerages. The agents pull deed, tax and property records for an address and fill the sales contract [site] [yc]. Threat: 2, because it is the cleanest live example of our W5 idea (one layer for one type of small firm), with named customers, but it is two founders, one US state and one workflow [site] [yc].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "AI-native OS for Real Estate Brokerages" and "intelligent transaction infrastructure" [yc] [site].
• Category a YC partner would put them in: vertical AI for real estate paperwork, in the "AI replaces the form-filler" group (inference).
• What they want to be known for: the agent that does the deal paperwork, not another place to store it.
• Where they sit on the map: layers L1/L2/L3 = 1/2/1, buyer = sme, business.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: one workflow (contract fill and deadlines). No public memory of the brokerage, no check step, no learning loop [site]. To own it we need a customer workflow where checked results change the next run, shown by a rule-change log and falling rework.
• Buyer gap: one US state, one trade [site]. They skip the UK and every other trade. To own it we need a named UK firm type with the pain and a person on file.
• Number gap: they count minutes saved and deal value, not money the brokerage earns or errors avoided [site] [blog]. To own it we need one before/after number a finance person would sign.
• Price gap: price not public [site]. Not clear whether a small firm can pay. To own it we need a published or quoted price tied to the result.
• Speed gap: they show a contract in under 5 minutes, but time to first value for a new brokerage is UNVERIFIED. Their Sotheby's start was a two-hour session [press]. To own it we need first value in days on a customer's real data.

8. For the YC application

• Proven today: Marton's logs show his system already runs agents that do and check work (about 5,600 jobs, each checked by a different AI, about 160 failure records that each added an automatic check). It turns a sales call into a personal offer page with payment links in about 10 minutes, and offer feedback lines fell from 44 to 18 over 3 offers. RealPact has 10+ brokerages, they claim [site]; we have none paying.
• Draft answer to "how are you different from RealPact?": DRAFT. "RealPact does one job well, contracts for New Hampshire brokerages. We are building the same idea for UK firms, where checked results change how the next job runs." Founders say it their own way. This is weak today because they are ahead on customers.
• Future claim (needs proof): "One reusable layer for one type of UK firm, with results that improve each run." Needs: one named firm type, 3 to 5 paid pilots or letters of intent, and one workflow run for 4+ weeks with a before/after number.
• What we understand that they don't: a reusable layer gets better when each checked result updates the rules, not only when a record is fetched. needs proof on a customer; Marton's logs show it on his own offers only.
• Copyability (M3): medium. The tech to add checks and memory is within reach of two strong builders on live customer data (inference). Chasing it would not pull them off their buyer, so it is a feature for them. Moving to the UK or another trade would. Not hard, so M3 cannot score above 3 against RealPact alone.
• Why now (M4): None found. The dated change closest to this space is that county records and MLS data can be called by agents, but this teardown found no dated source for it. Do not use "agents can now fill forms".
• Do not claim: that nobody does "one layer for one type of firm" (RealPact does, live). That our loop beats theirs (nobody has compared). That we have customers. "AI OS" as a headline (they use it).

M2 (what YC funds now) is NOT judged here. It comes from the themes of many companies in Pass 3.

9. Ideas worth taking

1. Win the first customer with a two-hour session on their real workflow, then build to it [press]. Cheap, fast and gives a named logo.
2. Show one filled document live in under 5 minutes as the demo. A visible before/after in one screen [site].
3. Put the claim "stores the work vs does the work" in one line. It is easy for a buyer to repeat [yc].

10. Unknowns

• Price, and whether it is per seat, per deal or per brokerage. Ask a brokerage owner who has seen a quote.
• Whether the "10+ brokerages" and "$35M+" figures are paying customers or pilots. Look for a case study or a second named quote.
• Whether approvals exist between the agents, and whether any learning or checking step exists. Watch the demo video and product film.
• Their plans outside New Hampshire. Watch the blog and films for a second state; other states use other forms.
• Who the "intelligent" Tax & Vision, Property, MLS and Deadline agents really are in product (only the Deed Agent is described).
• The 70% figure comes from a profile site, not from them. Find the original.

Sources

• realpact.ai [site]
• RealPact films, Oct 2026 [site]
• RealPact pre-seed announcement, Jun 2026 [blog]
• RealPact compliance page: "working toward SOC 2 Type II", no timeline [docs]
• YC company page [yc]
• YC launch post [yc]
• The Dartmouth, 21 May 2026: Magnuson competition win [press]
• Extruct profile of RealPact [docs] (third-party aggregator, source of the 70% / 25 minute figures)
• ListedKit: AI transaction coordination [press] (other players in this space)