A multi-role AI agent platform that unifies customer onboarding, support triage, and invoice reconciliation
Cut the ops team's weekly workload from 11 person-weeks to 3.2; savings reinvested into ARR growth.
Starting point— Challenge
A mid-market B2B SaaS (Europe-focused). In aggressive post-Series-A growth: 35 employees 11 months ago, 85 now. Internal operations — especially onboarding, customer success, and finance/accounting — couldn't keep up. Hiring a junior for each function looked like the most expensive solution.
In our first conversation with the CEO, one line stood out: 'The problem is, people now search across 5 Slack channels + Notion + Linear + email just to ask each other for information. A 1-hour task takes 4.' Classic coordination tax.
We engaged on an Equity Partnership model. Minimal fixed fee, the rest paid in company equity. Operational efficiency KPI tracked on the balance sheet.
Approach— Approach
- step 01
Week 1 — Process mapping
14 hours of ethnographic observation with onboarding, customer success, and finance teams (sat next to operators and recorded what they did). Surfaced repetitive patterns: information looked up in 60+ places across 3 tools and re-entered.
- step 02
Week 2 — Tool ecosystem integration
Unified read/write layer built across 8 systems (HubSpot, Stripe, Notion, Linear, Slack, Intercom, QuickBooks, Snowflake). We called this the 'company brain' — the agents' shared memory and action surface.
- step 03
Week 3-4 — Three-role agent build
Onboarding agent: when a new client signs up, account setup, initial data load, kickoff scheduling. Support triage agent: classifies incoming tickets, gathers context, routes to the right person or answers if simple. Reconciliation agent: invoice/payment matching between Stripe and QuickBooks, anomaly reports.
- step 04
Week 5-6 — Human-in-the-loop setup
Each agent action's approval level configured: low risk (auto), medium (one-click Slack approval), high (routes to a human + brief). For the first 2 weeks everything ran as medium risk while behavioral data accumulated; automation expanded from there.
- step 05
Week 7+ — Ongoing engagement
Post-pilot, the team started treating the 'agent platform' like a product. New use cases ship every month (now 9 different roles). Maksinet continues alongside equity with monthly active engineering capacity.
Results
"They talked to us like investors. Not for equity — they offered to charge nothing unless we hit the revenue numbers."
— Client side — CEO
Technology stack
- Anthropic Claude (Sonnet 4.5)
- OpenAI GPT-4o (for vision)
- LangGraph orchestration
- Temporal (durable workflows)
- Snowflake (single source of truth)
- Slack Block Kit (approval UI)
- Notion API
- Stripe API + QuickBooks API
What came next
Multi-tenant platform version
The internal platform is mature enough that the client is considering licensing it to their own customers (other SaaS firms). We're co-writing the product roadmap for this new line of business.
Where does yours sit in this picture?
In a 30-minute discovery call we listen to your current state and share an initial read on whether a similar engagement makes sense. No commitment.