GROWTH SCIENCE CAPITAL
Internal · Strategy Memo · May 2026
AI Agent Platform
Backbone.
Evaluating raia and the field. A backbone for our internal team and the portfolio's client-facing agentic workforce.
01 · The Question
What we're choosing.
We need a platform — or stack — that powers AI agents across our internal operations and our portfolio companies' client-facing deployments. raia is one credible option. The question is whether it's the right one.
Internal
GSC + brand operations
TeamCentral, BrandBot, Meeting Assistant, recruiting, knowledge sharing across the Connected Companies Network.
Portfolio
5 brands, multi-tenant
TMA, Levantage, Glial, GSC, CCS. Each needs isolated agents, branded experiences, and per-firm governance.
Client-facing
Productized agentic work
Levantage's law-firm clients, TMA's authority programs, Glial's consulting engagements. White-label matters.
02 · The Lens
Seven criteria that actually matter.
Translating "flexible, developer-friendly, user-friendly, economical" into the buying tests we'll apply.
Multi-tenant / white-label
Isolated agents per brand and per client; brandable surfaces.
Model flexibility
OpenRouter-native. No single-vendor lock-in.
Code escape hatch
Drop into JS/Python when logic exceeds the visual builder.
Predictable cost at scale
Flat or owned infra, not credit roulette.
Observability + governance
Audit trails, per-tenant logs, RBAC, SOC 2 / HIPAA paths.
Fits our existing stack
Complements TeamCentral, n8n-class orchestration, Netlify.
Non-technical ramp
Portfolio leaders configure their own agents without devs.
03 · The Real Decision
Before products, three architectures.
The real choice isn't "which logo" — it's which shape of stack we're committing to. Each has a different long-term trajectory.
A
All-in-one SaaS
raia · Lindy · Sierra · Decagon · Agentforce
Buy a turnkey platform. Vendor owns the runtime, the UI, the upgrades.
Trade-off
Fast to start. Locks you into pricing, models, and product roadmap you don't control.
B
Self-hostable orchestration
n8n · Botpress · Flowise · Langflow
Open-source builders you self-host. Visual canvas for non-devs, full code escape.
Trade-off
Own the data plane and economics. You run ops; you control destiny.
C
Code-first SDK + custom infra
Claude Agent SDK · OpenAI AgentKit · LangGraph
Build the platform yourself, embedded into existing infrastructure (TeamCentral).
Trade-off
Maximum flexibility and brand alignment. Highest engineering investment up-front.
Option A
All-in-one SaaS platforms.
raia · Lindy · Sierra · Decagon · Salesforce Agentforce · Microsoft Foundry · Google Gemini Enterprise
Where they win
- Genuinely no-code. Wizard to live agent in hours.
- Multi-channel out of the box — chat, SMS, email, voice.
- Compliance posture pre-baked (SOC 2, HIPAA, audit trails).
- Vendor handles upgrades, scaling, infra, security patches.
Why they fail for us
- Single-tenant by design — no clean way to isolate 5 brands + clients.
- Vendor model lock-in (raia = OpenAI Enterprise; Agentforce = Salesforce).
- Credit pricing or per-seat fees compound brutally at portfolio scale.
- Limited or no white-label of the configuration UI itself.
Option B
Self-hostable orchestration.
n8n · Botpress · Flowise · Langflow
Where they win
- Visual canvas + code escape — fits both tech and non-tech.
- Self-host: no per-execution tax. Your data, your infra.
- Native AI agent nodes (memory, tools, vector stores, MCP).
- Multi-tenancy via folders, workspaces, and isolated runtimes.
What you own
- Ops burden: uptime, backups, upgrades, secrets management.
- Builder UIs aren't white-labelable to clients (back-office tooling).
- Steeper ramp than Lindy/Make for non-technical leaders.
- Need separate front-end (TeamCentral) for client-facing UX.
Option C
Code-first SDK + custom infra.
Claude Agent SDK · OpenAI AgentKit · LangGraph · CrewAI
Where they win
- Total flexibility — agents become first-class TeamCentral entities.
- MCP standard for tool reuse across every agent and channel.
- Model-agnostic via OpenRouter (route Claude, Gemini, GPT freely).
- We turn TeamCentral into the platform raia is selling — but ours.
What it costs
- Engineering investment — but we're already 60% there with TC.
- No visual builder for non-technical users (solve in TC's UI).
- Maintenance cost grows with surface area — set thresholds.
- Slower time-to-first-agent than buying off the shelf.
04 · Landscape
The competitive field, mapped.
Where each platform sits and what it specializes in.
| Platform | Architecture | Strength | Fit for us |
|---|
| raia | All-in-one SaaS | Multi-channel, no-code, SOC2/HIPAA | Weak — OpenAI lock-in, no white-label |
| Lindy | All-in-one SaaS | Best-in-class personal AI EA | User layer only — not backbone |
| Botpress | Self-hostable | LLM-native chat + white-label widget | Strong — client-facing chat layer |
| n8n | Self-hostable | Open-source orchestration + AI nodes | Strong — replaces Make, owns ops |
| Voiceflow | Self-hostable | Voice + chat design, collaborative | Add later if voice goes core |
| Claude Agent SDK | Code-first | Embed agents into TeamCentral, MCP | Strong — extends what we have |
| OpenAI AgentKit | Code-first | Visual builder + Evals + ChatKit | Skip — OpenAI-only |
| Sierra / Decagon | All-in-one SaaS | Enterprise-grade CX/support agents | Skip — wrong shape, wrong price |
| Agentforce / Foundry | Hyperscaler suite | Deep CRM/M365 integration, governance | Skip — $50K+/yr commitment, lock-in |
05 · Spotlight
Honest read on raia.
raia is a credible product. It's not the wrong choice for a single mid-market firm with no engineering capacity. It's the wrong shape for what we're actually building.
What's good
- Genuinely no-code wizard for spinning up agents.
- Multi-channel (chat, SMS, email, voice) from one config.
- Constellation Software-backed — not going anywhere.
- Honest pricing: flat platform fee + token pass-through.
Why it's not our backbone
- Locked to OpenAI Enterprise — undoes our OpenRouter strategy.
- Wizard-first; we're past needing one. We have Claude Code.
- Multi-tenant story is weak across 5 brands + client deployments.
- Duplicates TeamCentral. Paying to rent a competitor to our product.
06 · Economics
Cost at portfolio scale.
Rough monthly run rate to support 5 portfolio brands + 10 client deployments + internal use, ~25 agents, moderate volume.
| Stack | Monthly | Note |
|---|
| raia Enterprise (estimated) | $5K – $10K + tokens | Quote-based; multi-tenant story is weak |
| Sierra / Decagon / Agentforce | $15K – $50K+ | Enterprise floor; not viable at our scale |
| Lindy Enterprise (across 25 agents) | $5K – $15K + credits | Credit overages are the killer |
| Hybrid: n8n + Botpress + Claude SDK | $500 – $1.5K + tokens | Plus our engineering time |
| Hybrid (managed enterprise tier) | $3K – $5K + tokens | If we want hands-off ops |
The hybrid stack saves an order of magnitude — and the engineering capacity is already paid for.
07 · The Call
Build the backbone. Don't rent it.
A three-layer hybrid stack — owned by us, model-agnostic, multi-tenant by design, and aligned to TeamCentral as the front door.
Layer 1
TeamCentral
The customer-facing front door — agent management UI, per-org isolation, our brand. We already own this.
Layer 2
Botpress · Client-facing chat surface
Where white-label matters. Branded webchat, knowledge bases, human handoff. Plus tier or self-hosted for compliance-sensitive clients.
Layer 3
n8n + Claude Agent SDK · Orchestration & code
n8n self-hosted as the workflow engine (replaces Make). Claude Agent SDK + MCP for code-driven agents inside TeamCentral. OpenRouter for model flexibility.
08 · Roadmap
30 / 60 / 120-day rollout.
Validate the architecture small. Migrate the legacy. Productize at the portfolio level.
30 days
Validate
- Stand up self-hosted n8n on a $50/mo droplet.
- Spin up Botpress Plus workspace; build one branded agent for TMA or Levantage.
- Sketch "agent-as-entity" extension for TeamCentral with Claude Agent SDK.
60 – 120 days
Consolidate
- Migrate Make.com workflows to n8n. Cost win compounds immediately.
- Standardize MCP server library (OpenRouter, Cloudinary, Twilio, ClickUp, TeamCentral).
- Position TeamCentral as the agent-management UI for the portfolio.
6 – 12 months
Productize
- TC + Botpress + n8n becomes the GSC offering: "AI agentic workforce in 120 days."
- Sell self-hosted Botpress + on-prem n8n as a Levantage premium tier.
- Lindy stays in the picture as a personal-AI-EA perk for portfolio leaders.
09 · Pressure Test
Open questions before we commit.
Each has a default answer baked into the recommendation. If we disagree on any of them, the architecture changes.
Voice strategy
Is voice agent capability core to any portfolio firm in the next 12 months? If yes (Levantage law-firm intake?), pre-decide Vapi vs Retell vs Voiceflow now.
Compliance ceiling
Glial deals with sensitive org data. Some Levantage clients may require HIPAA-equivalent handling. Self-hosted Botpress + n8n on SOC 2-attested infra is the answer if so.
Who maintains it
TeamCentral + n8n + Botpress is a stack we can run. The handoff UX (TC's UI) for non-technical portfolio leaders is the most important thing to nail.
Build-vs-buy threshold
If maintenance exceeds 4 hrs/week of ops work, we reassess managed offerings. Set the threshold up front; revisit at 60 days.
What we're deciding
Don't sign with raia.
Build the backbone we'll own.
Three asks for the team
- 01Approve the 30-day pilot — n8n droplet + Botpress Plus.
- 02Confirm voice strategy timeline (Levantage intake — yes/no in 90d?).
- 03Name a portfolio firm + workflow as the first real test case.