XMarts room

Fall 2024

Rulebase

AI customer operations for regulated industries.

Visit website

What they do

AI customer operations with QA and specialist agents

How close to us

3, close

QA layer plus specialist agents owning accounts, with checks; two layers but one function.

Wedges

  • W2 Seeing whether your AI agents actually produce adjacent
  • W5 One layer for one type of small firm adjacent
  • W7 Find and fix margin leaks adjacent
  • W8 The other 90% of an AI skill adjacent
L1 Company brain (knows the organisation) 2, clear
L2 Agent team (runs agents, tools, approvals) 2, clear
L3 Does and learns (finishes, checks, measures, updates the brain) 3, strong

Who buys

Firm size
mid-market
Buyer type
business
Role
COO, chief customer officer, head of compliance or onboarding at fintechs and regulated firms
UK
not known
Their own description (from YC)

We run customer operations for regulated industries with AI. Our QA layer reviews every interaction, human and AI, cutting manual QA by 99%. On top of it sit specialist agents for onboarding, remediation, reactivation, and retention that own accounts from approved to transacting, chasing customers over WhatsApp, email, and SMS

Full teardown

Rulebase (YC Fall 2024): teardown vs XMarts

In one line: Rulebase reviews every customer conversation at fintechs and regulated firms (human and AI), then sends agents to finish stalled work such as onboarding and reactivation [site] [yc] [blog]. Threat: 2, because it is the one rival that claims all three layers on one workflow and is live with named customers, but it is four people, sells to customer operations at fintechs, and reports accounts, not money [yc] [blog] [case-rho].

Check of the brief's claim ("full-loop, recovers stalled revenue"): partly true. The Revenue Agent does own stalled accounts until the customer converts, and it acts rather than advises [blog-agents]. But the number they claim is "120% lift in recovered accounts" and "3x faster activation", with no baseline or method [blog-agents]. No customer case study shows a money result [case-rho] [case-kuda]. No named customer is shown using the Revenue Agent [blog-agents] [press-digest]. The proof they publish is for the QA and monitoring product, not the revenue agent. See section 8.

1. Fact file

Item Finding Source
What they say Home page: "Monitor every agent. Improve every outcome." YC line: "AI customer operations for regulated industries." [site] [yc]
Founded, team, funding YC Fall 2024. Team of 4 on the YC page. $2.1M pre-seed announced Sep 2025, led by Bowery Capital, with YC, Commerce Ventures and Transpose Platform VC. Founders say revenue grows "double-digit" month on month (founder claim). [yc] [press-tc]
Founders Gideon Ebose (CEO): ex-lead product designer, then Microsoft. Chidi Williams (CTO): built the open-source speech-to-text tool Buzz (300k+ downloads, 12k+ GitHub stars) and systems at Goldman Sachs. They met in London. What YC likely saw: a technical founder with a popular open-source tool plus a product founder who sold to banks (inference). [yc] [press-tc]
Customers Logos shown: Qonto, Lesaka Technologies (Nasdaq), Nala, Interswitch, Kuda, Rho. An unnamed "Fortune 50" financial firm in the press. [site] [press-tc]
Price Not public. Job ad says deals run $50K to $120K+ a year. Press says usage-based, per interaction reviewed or workflow automated. [jobs] [press-tc]
Path Voice-fraud defence for finance (earlier launch, date not found) → Beam, a live copilot for support agents (date not found) → "AI Coworker" for back-office work in fintech (Sep 2025) → compliance agents (Jan 2026) → Gretta for compliance testing (Mar 2026) → Revenue Agents (Jul 2026) → home page now leads with agent monitoring. No rename found. [yc] [press-tc] [blog-gretta] [blog-agents] [site]

2. Their proposition, reverse engineered

Part What they do
Who buys COOs, chief customer officers and onboarding leads at payment processors, benefits administrators, retirement platforms, neobanks and cross-border fintechs [jobs]. Compliance heads also buy (Rho's head of compliance is quoted) [case-rho].
The pain QA teams review only 3% to 5% of conversations, so problems show up late. Rho's COO: "we always felt we were behind" [case-rho] [press-rho]. For the revenue side: accounts stall between approval and first use [blog-agents].
The promise Review 100% of interactions, find why AI and human agents fail, fix it, and send agents to finish stalled accounts [site] [blog-agents].
How it works Four steps: Monitor, Detect, Diagnose, Improve. A failed AI journey is paired with the human resolution, turned into a proposed knowledge, prompt or tool change, tested in a simulation, then measured on the same intent [site] [case-rho]. The Revenue Agent keeps one durable task per stalled account and wakes on events (upload, reply, deadline) [blog-agents].
The output Coverage reports, risk alerts to Slack, coaching notes, proposed fixes, and agent actions over email, text, WhatsApp and calls, with every action logged [case-kuda] [blog-agents].
Proof style Counts of coverage, issues caught, escalations, hours saved. Money is claimed once, as "recovered accounts" with no baseline [case-rho] [case-kuda] [blog-agents].
Who sells it Founder-led outbound. First dedicated sales hire is open, to run outbound without an SDR [jobs].
Price model Not public. $50K to $120K+ a year per deal per the job ad. The ad says the hire will refine pricing [jobs].
The story Enemy: random sampling and slow, manual QA. Belief: review everything and let agents act on what they see. The CEO says they "automate workflows that start with a customer interaction" [press-tc].

3. Positioning

  • Category they claim: "AI customer operations for regulated industries" [yc]; now "AI agent monitoring platform" for customer-facing teams [site]; and "AI Revenue Agents for financial services" [blog-agents].
  • Category a YC partner would put them in: AI QA and agent monitoring for customer support in fintech, with an agent upsell into onboarding and reactivation (inference).
  • What they want to be known for: being the layer that watches every customer conversation and then acts on it.
  • Where they sit on the map: L1/L2/L3 = 2/2/3 (claims), buyer = mid-market (inference from named customers), business.

4. Three layers vs ours

Layer Them (score + evidence) Us (Marton's logs show)
L1 Company brain 2. Each stalled account is a durable task holding stage, blockers, history across support, onboarding and risk, plus company policies [blog-agents]. Also "customer intelligence" from conversations [site]. Memory is about the customer and the case, not the whole company. A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration 2. A Customer Agent Runtime with several agent types (QA, Revenue Agent, Gretta), event-driven, escalating to a person with context, with every action and approver inspectable [blog-agents] [blog-gretta]. One runtime, one function. About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning 3 (claimed). The agents act. The improve step proposes a fix, simulates it, and measures the result on the same intent afterwards [case-rho] [site]. The site calls fixes "governed" and does not say if a person must approve each one [site]. No independent proof [press-digest]. About 160 failure records, each with an automatic check. Offer feedback lines fell from 44 to 18 as 22 rules were added. Only about 15% of jobs reached the main code.

5. Same and different

Topic Same or different Them Us
Core idea Same One workflow, agents that act, a loop that checks and improves [site] [blog-agents]. The brief's thesis: the full loop on one workflow with one number.
Checking work Same A QA layer reviews every interaction, human and AI [yc]. A different AI checks each job; automatic tests follow (inventory).
Kind of work Different Customer conversations and the accounts that stall after them [press-tc]. Any commercial or finance workflow, for example margin leaks (W7).
Buyer Different Fintech and regulated customer operations and compliance heads [jobs]. UK SMEs and scale-ups (brief). No customer yet.
Where it learns Different From conversations and human resolutions at the customer [case-rho]. From checked results of its own jobs and feedback (inventory). Only proven on Marton's own work.
Number they own Different Coverage, escalations, "recovered accounts" [case-rho] [blog-agents]. None yet. Target: a money number.
Proof Different Two case studies, one renewal (Rho), a SOC 2 Type II claim [case-rho] [case-kuda] [press-rho] [site]. No outside customer (brief).
Price Different $50K to $120K+ a year [jobs]. Not set.
Delivery Different Software plus founder-led sales; 4 people [jobs] [yc]. Marton builds; Xavier sells.

6. Wedge overlap

Wedge Fit Why
W2 Are my AI agents producing? adjacent Their home page sells monitoring of AI agents: silent failures, policy breaches, root causes [site]. It covers customer-facing agents (chat, voice) and human agents. W2 is about teams running coding and work agents, and what they cost. Same problem, different agents and buyer. A YC partner may still say "Rulebase does this".
W5 One layer for one type of small firm adjacent They are the pattern: one vertical (regulated fintech), one entry point (QA), then agents on top [jobs].
W7 Margin leaks for commercial or finance leaders adjacent Their Revenue Agent chases stalled accounts to recover revenue [blog-agents]. Same shape (find the leak, act, measure). Different leak and different function.
W8 The other 90% of an AI skill adjacent Self Improvements turns failures into fixes and measures them [site]. They also list a Skills library, with no detail found [site]. UNVERIFIED how far it goes.

7. Whitespace (where they stop)

  • Layer gap: their agents start from a customer interaction [press-tc]. Pricing, demand and margin problems do not start there. To own it we need one finance or commercial workflow run end to end by our agents, with a measured result.
  • Buyer gap: regulated fintech customer operations in the US, EU and Africa [jobs] [yc]. No UK mid-sized commercial or finance buyer shown. To own it we need 3 to 5 such buyers saying they have the pain.
  • Number gap: they publish counts (coverage, escalations, accounts), not money. Both case studies say so by omission [case-rho] [case-kuda]. Their own sales ad lists "revenue lift or cost reduction" as an open question [jobs]. To own it we need one before/after money number a finance person would sign.
  • Price gap: $50K to $120K+ a year [jobs]. To own it we need a price a 50-person firm can pay, tied to the result.
  • Speed gap: Kuda's first useful result came about two weeks after launch [case-kuda]. That is fast. We would need to beat it, so this gap is weak. Not claimed.

8. For the YC application

  • Proven today: Marton's logs show his system runs agents that do work and check it: about 5,600 jobs, each checked by a different AI, with about 160 failure records that each added an automatic check. Rulebase's published customer proof is for QA and monitoring, in fintech customer operations [case-rho] [case-kuda].
  • Draft answer to "how are you different from Rulebase?": DRAFT. "Rulebase works inside customer conversations at fintechs. We pick one commercial workflow at a UK mid-sized firm, run it with agents and own the money number." Founders say it their own way. The second sentence is a future claim; see below.
  • Future claim (needs proof): "We find and fix margin leaks and show the money." Needs one customer workflow run for 4+ weeks with a before/after number.
  • What we understand that they don't: a full-loop pitch is only believed when the number is money, and their own sales ad shows they have not settled that [jobs]. needs proof.
  • Copyability (M3): medium. They already have the runtime, agents that act, a measure-after step and a sales motion [blog-agents] [case-rho]. Pricing and margin work on a finance buyer would pull them away from regulated customer operations, and they are four people with a first sales hire still open [jobs] (inference). Not hard, so M3 can not score above 3 against Rulebase alone.
  • Why now (M4): None found in this teardown. Do not use general AI progress.
  • Do not claim: that Rulebase "recovers stalled revenue" as a proven result. Say "they claim a 120% lift in recovered accounts" [blog-agents]. Do not call them a single-layer rival. Do not claim our loop is measured on a customer.

For the pitch, one finding: Rulebase is the live example of "full loop on one workflow", and it still publishes no money number. A customer-proven money number is the gap to fill.

9. Ideas worth taking

  1. Enter through a check (QA) that proves value in about two weeks, then add agents that act [case-kuda] [jobs]. It earns trust before the bigger ask.
  2. Make every agent action inspectable: the action, the reason, the evidence and the approver [blog-agents]. Buyers in regulated firms ask for this.
  3. Show the improvement loop as a pair: failed AI run next to the human fix, then the measured change [case-rho].

10. Unknowns

  • Whether any named customer runs the Revenue Agent. The launch page does not say [blog-agents]. Ask for a customer reference or look for a case study after Sep 2026.
  • What "recovered accounts" and "120% lift" are measured against. No baseline was found.
  • The claim mismatch: the launch page says "3x faster activation" and the press says time-to-activation "cut in half" [blog-agents] [press-digest]. These are not the same number.
  • Whether Self Improvements change prompts on their own or wait for a person [site].
  • UK customers. Qonto is a European bank and they list operations in Ireland [site], but no UK customer was found. Check the customer page and Nala.
  • Dates of the Beam and voice-fraud launches, and the true date of the funding post (blog lists Jan 2026; the TechCrunch link is dated 16 Sep 2025) [blog-list].

Sources

Technical detail
Slug
rulebase
Lens hash
179096c479
Pass
deep
Stale
no