Field notes, not thought leadership. Each post is one system I built for a real client or for my own company, what it did, and what I would do differently.
The AI agent incidents in the news were containment and configuration failures, not models turning clever. For a ten to sixty person business that translates into five unglamorous controls before an agent touches live customer data: a scoped credential, a staging target, a human write gate, a readable audit log and a kill switch.
Gemini 4 Argon is a model option for n8n Agents and the Google Gemini node from launch day, so upgrading the brain behind a lead scorer or document parser is a dropdown and a credential rather than a rebuild. The work worth doing is making the swap testable.
Open WebUI is a self-hosted chat front end that points at local Ollama models or any OpenAI-compatible API, so a per-seat subscription becomes one server plus metered tokens. The install is about three hours: container, provider key, signup closed, proxy, prompts, rule sheet. Skip it if nobody on staff patches servers.
OpenRig gives coding agents named seats, roles and shared context, so a seat keeps its address and queued work when the agent in it changes. Useful if your team already runs Claude Code or Codex daily. Not a lead intake or reporting tool. Budget about three hours for a first rig.
AgentSDR is an open-source, self-hosted AI SDR for email, LinkedIn and WhatsApp, with an AI CRM that classifies replies. Its setup path is a git clone, Docker and Postgres, which makes it a useful free benchmark for what automated first touch requires, and a poor substitute for a CRM-integrated build.
A fix you have done twice is a procedure, and Claude Code will read that procedure from a single SKILL.md file. The build is mostly writing, not coding, and a sensible budget is about three hours. Skip it when the fix changes shape every time.
NVIDIA and Microsoft have opened a vendor-supported path for AI agents to run on the Windows machine instead of a cloud API, but the privacy and speed claims attach to new RTX Spark systems, not to your existing office PCs. The useful move this month is an audit, not a purchase order.
Ringg and OpenAI report that Ringg's agents resolve up to 65% of routine customer calls without a human. That is a vendor number, so it belongs in your spreadsheet as a ceiling, not a forecast. Here is the worksheet to run on your own call log first.
Businesses pay freelancers to fix GoHighLevel lead follow-up even though the platform documents round-robin assignment, internal notifications and timed reminders. The gap is configuration and ownership, not features. Six checks, about 40 minutes, tell you whether you need a rebuild or one missing notification action.
OpenAI announced dots, agents in ChatGPT built to keep working between conversations, on 2026-09-29. The persistence is real and the rollout is narrow. For a small team the new work is governance, not prompting: scope, one connected app, a log someone reads, and a named stop switch. Run one low-risk dot this month.
Demand keeps showing up for an automation layer beside a CRM the business intends to keep. The licence is the cheap part: n8n's Community edition is free, while the server, the patching and the on-call are the real purchase. Add the layer for timing, shape and hand-offs, not for reporting.
OpenAI will train about 150 Small Business Development Center advisors to run hands-on AI workshops for at least 1,000 small businesses. The product news is minor. The channel news is not: the person raising AI with an owner stops being a vendor and becomes the advisor they already trust.
VoiceStudio is a local, open-source speech tool that clones voices, dubs video and transcribes without accounts or cloud keys. The docs support that. They do not describe an IVR product or promise studio quality, so treat it as an export-and-upload workflow and budget about three hours to find out whether the output is usable.
A GitHub project called Jeff is described publicly as a 0.8B decision model answering in roughly 30 milliseconds, but nothing verifies it as a replacement for a paid routing call. The honest move is a three hour test on your own past routing decisions, judged on accuracy before cost.
Paperclip is an open-source server and UI that puts agent activity, month spend, pending approvals and failed run transcripts on one screen. The operations detail is the reason to install it, not the orchestration branding. Budget about two hours for the first agent, and skip it if nobody will open the dashboard.
A Relevance AI plus Make build for a multi-location front office reliably takes capture, routing and CRM writeback off the desk. It does not replace phone staff, and no primary source says it does. Here is the work split, the arithmetic on your own call volume, and what still needs a person.
HighLevel already documents an official MCP connector for Claude, so the interesting question is not whether AI in the CRM is early. It is who is accountable once a model can read and write records. The work to do this month is permissions, approval gates and logging, in writing, before you approve scopes.
Stripe's writeup of its internal agent platform is worth reading for the operating model, not the model. The copyable part for a small operator is ownership, scoped access to approved documents, and a refresh habit. Chat over a folder of PDFs is the easy half and the half that fails quietly.
Small businesses keep paying for one-off help with lead follow-up, assignment, appointment booking and contract sending. HighLevel's own docs list all four as workflow actions. The gap is not missing tooling, it is that nobody inside the business owns the workflow, so the same task gets bought again next quarter.
Live CRM job postings show buyers are not shopping for software. They are paying for implementation, workflow design and a clean handoff, sometimes inside days. For a small team that means writing down the workflow before buying help, and judging any quote on documentation as much as on build.
The wrong answers usually come from the ingestion step, not the model. A new paper on retrieval-aware chunking normalizes every format to PDF, converts once with a multimodal model, and plans chunks over section identifiers. The reported savings are large, the benchmark is clean, and your scans are not. Audit before you buy.
A public bundle of LinkedIn skills for Claude Code is worth copying for its workflow, not its writing. The draft is the easy part. The de-AI pass, the audit step and the approval gate are what stop a small team from publishing the same sludge as everyone else, faster.
The reply your AI sends is probably fine. The path it travels before sending is where the deal goes. n8n's own latency post puts orchestration, routing, caching and timeouts ahead of prompt quality, and that reframes speed to lead as an architecture problem you can audit in an afternoon.
A LinkedIn carousel is a PDF and an Instagram carousel is a folder of JPEGs, so both come out of one Pillow script instead of a design tool. Here is the two hour build, the slide contract that makes every later post a fill-in-the-blanks job, and the platform limits that catch people.
Usually it is not a rebuild. In a neglected GoHighLevel account the contact data is fine and the automation layer is what has rotted, in four predictable places. Run a 90 minute audit first: sending status, active workflows, who holds the API keys, duplicates, and three traced leads.
A multi-site housing operator ran sales in GoHighLevel and reporting in a legacy back office whose API matched leads by email only and auto-closed them after 72 hours of silence. The bridge that worked stamped the back-office ID onto every CRM contact, looked leads up on demand instead of migrating 14,000 of them, and treated a 201 response as a claim to verify, not a result.
Email, LinkedIn, SMS, voicemail and AI voice calls in one 30-day, 12-step cadence built in GoHighLevel with Clay enrichment, ElevenLabs voice and avatar video personalisation, deployed across nine partner agencies. On the staffing campaign: 505 leads contacted, 72% open, 25.5% reply, 13.2% positive reply.
For an insurance-agency recruiting business I automated candidate sourcing, screening, scoring and interview scheduling across the full funnel. Time to hire went from six weeks to under one, and the qualified-candidate rate rose from 40% to 78%. The same sourcing stack now runs for nine agencies.
The internal system that runs Web AI Automations: 8 primary agents and 35 sub-agents on React and Vite, FastAPI, Supabase and n8n, with model routing across Claude and OpenRouter to keep inference cost under control at scale. Two walkthroughs are on the Watch section of the home page.