AI Platform Options for a Small Consulting Firm
ENTR 560 client project — contact capture and email triage
How to read this page
- Step 1 — Choose a provider: Claude Team, Gemini in Google Workspace, or ChatGPT Business. The vendor hosts and maintains everything.
- Step 2 — Check what the provider can do: connectors (MCP), email triage, contact pipeline, human approval, scheduling and triggers.
- Step 3 — Add a harness only if needed: if the provider's built-in tools can't run a shared, always-on contact pipeline, add Hermes or another MIT-licensed agent on top of the chosen provider's model.
Questions for the next client meeting: Which Google Workspace plan does the client have? Who is the Workspace admin? How many people need seats?
Step 1. Choose a provider
Suggested order: Claude Team → Gemini (check the client's plan first) → ChatGPT Business. Scheduling results in Step 2 may change the order for contact capture.
#1
Claude Team
$20/seat/mo annual ($25 monthly) · Premium $100 ($125)
Recommended
- 2–150 users; SSO, central billing, admin connector controls
- No training on your content by default; SOC 2 Type II
- Small Business plugin: inbox triage, follow-ups, scheduled runs, approval before anything sends
- Custom MCP connectors added by Owner; per-tool Allow / Ask / Block
- Mailchimp connector; Contacts+ official MCP works
- Up to $500/mo pooled API credits (conditions not yet verified)
#2
Gemini in Google Workspace
Included from Business Standard (~$14/user/mo annual — verify the client's plan)
Check first
- the client already uses Google Workspace → Gemini works inside Gmail, Sheets, Docs
- Built-in MCP integrations incl. Mailchimp, HubSpot, QuickBooks (on by default, admin-managed)
- Workspace Studio: no-code AI flows from plain-language descriptions
- Custom MCP (e.g., Contacts+) mainly via Gemini Enterprise / Google Cloud — higher barrier
- Lowest maintenance: no new vendor, no new accounts
#3
ChatGPT Business
$20/seat/mo annual ($25 monthly) · Premium $100 ($125)
Viable
- 2–200 users; SSO/MFA, usage analytics, spend controls
- No training on business data by default
- ChatGPT Work, workspace agents, scheduled tasks, Google Workspace connector
- Full MCP (read/write, beta) — only admins can use developer mode on Business
- Contacts+ official MCP works
Enterprise tiers (Claude / ChatGPT)
Claude: $20/seat + usage at API rates · ChatGPT: custom
Overkill
- Adds audit logs, SCIM, custom retention, data residency
- Not needed at the client's size
Step 2. What each provider can do
Connectors & workflow fit
| Claude Team | Gemini (Workspace) | ChatGPT Business | |
|---|---|---|---|
| Cost (5 users) | ~$100/mo | Maybe $0 extra | ~$100/mo |
| Who maintains | Anthropic | OpenAI | |
| Trains on data | No (default) | Check plan terms | No (default) |
| Custom MCP | Owner adds; per-tool Allow / Ask / Block | Mainly via Gemini Enterprise | Admins only (developer mode, beta) |
| Gmail / Sheets | Yes | Native | Yes |
| Mailchimp | Connector | Built-in MCP | Custom MCP |
| Contacts+ | Official MCP | Enterprise only | Official MCP |
| Email triage | Ready (plugin) | Good | Good |
| Contacts: shared pipeline | Partial (per user) | Studio flows (to test) | Partial (agents) |
| Human approval before send | Default | Manual | Configurable |
Contacts+ has an official MCP server (mcp.contactsplus.com): search, add, tag, notes; every change needs approval, no delete/export.
Scheduling & triggers (most workflows will run unattended)
Most of the client's workflows should run on a schedule or when a new email arrives, not when someone clicks a button. Info from vendor help centers (checked Oct 2026); not tested by us yet.
| Workflow | Best run mode |
|---|---|
| Email triage | Scheduled — every weekday ~8:00 local time: priority brief + reply drafts |
| Contact capture | Event-triggered — on each new email to the shared intake mailbox (hourly check is acceptable fallback) |
| Possible future workflows | Weekly partner news; daily "who is due for follow-up"; meeting prep the day before a calendar event |
Claude Team — Cowork scheduled tasks
Time-based only- Runs in Anthropic's cloud — computer can be off, app closed; uses connectors
- No "on new email" trigger found → contact capture = periodic check
- Cowork on web/mobile still beta for Team
- Avoid Dispatch / Claude Code desktop local tasks — need the computer awake; missed runs only catch up once
ChatGPT Business — Scheduled tasks
Has email trigger- Recurring up to hourly, exact times
- Gmail event-triggered tasks: up to 30 runs/hour, 720/day; events may be batched
- Business limit: 10 active tasks (official; one third-party source says 15)
- Risk: third-party reports that ignored tasks auto-pause — verify in pilot
Gemini — Workspace Studio flows
Native Gmail trigger- "New email" and schedule starters; runs inside Gmail
- Max 25 flows/user, 20 steps/flow, ≤25 active Gmail-started flows
- Daily run cap per user — if hit, all that user's flows pause 24h
- Admin alert rules available (e.g., >100 runs/hour, heavy "Ask Gemini" use)
Hermes — built-in cron
Most flexible- No limit on frequency or task count; any logic; inbox polling or Gmail push
- Requires a 24/7 server; failed runs and expired OAuth tokens must be handled by someone
Rules for any unattended workflow
- Account ownership: managed-tool tasks live under one person's account — if that seat is removed, the task stops. A shared intake mailbox may need its own seat/account (add to cost).
- Silent failure is the biggest risk: include "processed N emails today" in each brief, alert a named owner when a run fails, and review run history weekly (monitoring + handoff runbook).
- No double processing: label handled emails (e.g., "AI-processed") and skip them next run.
- Drafts only: unattended runs must never send; confirm each task writes drafts only.
- Cost & quotas: hourly = 24 runs/day; seat plans don't bill more but consume usage limits, and Gemini has daily caps. Batch daily where possible.
Impact on ranking: email triage works on all three managed tools. For contact capture, Gemini (Gmail starter) and ChatGPT (event trigger) fit better than Claude's periodic checks. Proposed pilot test (week of 10/19): on each candidate platform, run one daily-brief task and one new-email-triggered task for 5 business days; record missed runs, auto-pauses, and duplicates.
Step 3. Harness on top of the provider (only if needed)
A harness runs our own agent workflow using the chosen provider's model. Use it only if Step 2 shows a gap (e.g., a shared, always-on intake-mailbox pipeline: parse → dedupe → Sheets/Mailchimp). Ready-made agents need the least code; SDKs need the most — and code is what the client cannot maintain. Not tested by us except Hermes.
Hermes Cloud (hosted by Nous Research)
Medium $0.56/day (~$17/mo) · Large $1.09/day (~$33/mo) · stopped $0.03/day — plus model & tool usage
Removes the server problem
- One-click deploy; dedicated always-on instance in its own container; Nous runs the server
- Medium: 10 concurrent sessions, 2 GB RAM, 4 vCPU · Large: 20 sessions, 4 GB, 8 vCPU
- Billed daily from Nous Portal credits; Portal Plus $20/mo includes $22 credits (also covers models & hosted tools)
- Same Hermes features: cron scheduling, memory, skills, MCP, Email / Slack / Discord / Telegram
- One shared agent for the whole team (not per seat) — fits a shared, always-on intake pipeline
- Still open: data-handling terms, training policy, server region, SOC 2 — not yet checked; OAuth setup and workflow upkeep are still ours
Hermes Agent — self-run
MIT, free · model from the chosen provider
Only if Cloud terms don't fit
- Any MCP server, skills, memory, cron, messaging gateways (incl. WhatsApp)
- Reusable skills: Anthropic knowledge-work-plugins (Apache 2.0) + Hermes built-ins
- Best fit for a shared, always-on contact pipeline
- Risk: needs a 24/7 server, OAuth tokens, upgrades — nobody at the client to maintain after 12/8
| Harness | Type | License / cost | Models | Strengths | Maintenance burden | Verdict |
|---|---|---|---|---|---|---|
| Hermes Agent (Nous Research) | Ready-made agent | MIT, free | Any (OpenRouter, OpenAI-compatible, local) | Skills, memory, cron, MCP, messaging gateways (incl. WhatsApp); we already run it | Medium: server, tokens, upgrades | Our pick |
| OpenClaw | Ready-made agent | MIT, free | Anthropic, OpenAI, Gemini, DeepSeek, Ollama… | Largest community; 15+ messaging platforms, cron, ClawHub skills marketplace | Medium: same as Hermes | Equal alternative |
| n8n | Visual workflow + AI agent node | Source-available (not OSI open source); cloud from ~$240/yr | Any major provider or local | Drag-and-drop flows a non-engineer can read; 1,000+ app integrations; cloud option removes the server | Low–medium (cloud) / medium (self-hosted) | Best for handoff readability |
| Claude Agent SDK | Code library | MIT | Claude only (Anthropic, Bedrock, Vertex, Azure) | Strong tool use; same engine as Claude Code | High: custom code to maintain | Too much code |
| OpenAI Agents SDK | Code library | MIT | OpenAI-first | Handoffs, guardrails, MCP, tracing | High | Too much code |
| Google ADK | Code library | Apache 2.0 | Gemini-first | Sequential/parallel/loop agents, A2A, web UI; fits Google Workspace | High | Too much code |
| LangChain Deep Agents | Code library | MIT | Any | Built-in multi-tenancy, per-user sandboxes, RBAC (useful for per-client isolation) | High | Too much code |
Takeaway: if a harness is needed, choose a ready-made agent (Hermes / OpenClaw) or n8n — never a raw SDK, since every SDK option means custom code the client must maintain after handoff.
Considered and excluded
Codex
Not for client users- Coding agent — useful for us to build, not for client staff to run daily
MailFlow (self-hosted webmail)
Excluded- AGPL email client, not an agent; adds a second copy of the client mail
- Gmail via App Password = broad, unscoped access
- Single-maintainer project
Sources: claude.com/pricing, openai.com/business/pricing, workspace.google.com/pricing, vendor help centers, contactsplus.com, agent-framework docs and comparison articles.