For years, running a serious education and events business meant logging into everything.
One platform for finance. Another for customer records. Another for online learning. Another for files. Another for meetings. Another for campaigns. Then more people were needed to move information between those systems, check what happened and keep the entire machine running.
At one point, we had more than 30 full-time team members in the Philippines supporting our backend operations.
Today, I rarely need to open most of our business platforms.
I open Slack. I talk to my AI agent. The agent finds the right information, follows the right permissions and either brings the answer back or executes the approved task.
That is the real AI shift.
It is not a better chatbot.
It is a new operating system for the business.
Complexity forced us to think differently
The Speakers Institute ecosystem is not a neat little database.
It includes large libraries of files and media, a global customer database, clients operating across multiple countries and time zones, and thousands of moving records across our platforms.
That scale used to demand a large amount of manual coordination.
Someone had to find the file. Someone had to update the contact. Someone had to check the payment status. Someone had to move the opportunity. Someone had to prepare the campaign. Someone had to chase the missing information. Someone had to report back.
Our AI agent now handles much of that repeatable coordination and execution. It has absorbed a large share of the backend work once performed by our 30-plus-person support operation.
That does not mean humans no longer matter.
It means people should not spend their best hours copying, chasing, searching and reconciling when a properly governed agent can do that work faster and more consistently.
Humans still own judgement, relationships, strategy, sensitive decisions and final authority.
Slack became the front door
Our technology stack still exists.
We use Xero for finance. Dropbox holds heavy files and media. Google Drive holds working documents and lighter operational assets. Ontraport and GoHighLevel support customer relationship management and learning delivery. Zoom supports online meetings and events. Other specialist tools connect into the same ecosystem.
The difference is that these platforms are no longer the front door.
Slack is.
Inside Slack, I have channels for different companies, departments and projects. I speak to the relevant agent in plain English.
I might ask:
- Who has registered for the next event but has not received the correct reminder?
- Which client payment status needs human follow-up?
- Build the draft campaign, but do not send it until it has been approved.
- Find the latest version of a document and prepare a client-ready summary.
- Reconcile what the CRM says against the actual payment or attendance record.
The agent does not merely write an answer. It uses permissioned tools to inspect the relevant systems, perform the approved action, read the result back and report what actually happened.
That last step matters.
An API saying "accepted" does not prove an email arrived, a social post went live or a payment succeeded. Our agents are designed to verify the final state, not celebrate the button press.
The model is not the system
This is where many people misunderstand business AI.
Models and assistants such as Claude, Codex and Grok can provide extraordinary reasoning, writing and coding capability. But a model is not your business operating system.
The model is the engine.
The harness is the control plane around it.
An agent harness can coordinate:
- which model is used for which task
- what information the agent can access
- which tools it can operate
- what memory it can retain
- what actions require human approval
- what must stay in a sandbox
- what gets logged and audited
- how results are verified
We use an open, self-hostable harness as that control layer. OpenClaw and Hermes are two examples. I personally use OpenClaw.
I host agents in both a local sandbox and a cloud-based sandbox. Locally, the system runs on my Mac Mini. In the cloud, it runs inside private rented infrastructure that I control. That gives me flexibility without giving every agent access to everything.
The agents are separated by role, data access and authority. I can add staff and contractors to a specific Slack channel for a specific department, while the agent they use can reach only the information relevant to that role. Broader company access remains with the executive team.
I also do not use the most expensive model for every task. Open-source models can handle lighter work. Paid models are brought in for the more difficult reasoning and production tasks. The harness lets the job determine the model instead of locking the whole company into one AI provider.
This is not the same as saying every byte always remains local.
If a self-hosted agent sends a prompt to a cloud model, that provider still receives the data included in that request under the terms of the specific product or API. A sandbox controls what an agent can reach and change. It does not automatically prevent data from leaving the environment.
Real control comes from architecture, permissions, provider settings and disciplined data handling.
The privacy conversation needs more accuracy
I often hear speakers and experts say, "AI trained on my intellectual property and gave it to someone else."
That concern deserves attention, but blanket claims are not helpful.
Data policies differ between providers, consumer products, business plans, API products and individual settings. Some consumer services may use conversations for model improvement unless the user opts out. Many business and API offerings state that customer data is not used for training by default.
The right question is not, "Is AI safe?"
The right questions are:
- What data is leaving my environment?
- Which provider receives it?
- Is it retained, and for how long?
- Can it be used for model improvement?
- Who can access it?
- Can I remove access immediately?
- Can I audit what the agent did?
Owning the harness gives you more control over those decisions. It does not remove the need to make them.
For speakers, authors, coaches and consultants, this matters enormously. Your frameworks, manuscripts, client records and commercial systems are part of your intellectual property. They should not be poured into random consumer tools without understanding the settings and terms.
How we built it
We did not begin by asking, "Which AI tool is coolest?"
We began with the business.
1. Map the work
We identified the repetitive jobs people were performing across files, CRM records, campaigns, events, payments, client communication and reporting.
2. Define the source of truth
Every important answer needs a controlling record. A pipeline stage is not proof of payment. A registration is not proof of attendance. A folder promise is not proof that the files arrived.
3. Set authority boundaries
The agent needs to know who may approve what. Reading data, drafting a message, changing a workflow and sending money are completely different levels of authority.
4. Connect systems with minimum access
Each connector receives only the permissions required for its job. Finance data is restricted. Private diary information is separated. Franchise locations see only their approved scope. Sensitive actions stop for human approval.
5. Add verification
After an action, the agent reads the live result back. It reports success only when the final state is verified.
6. Scale one proven workflow at a time
We did not automate the entire company in one leap. We built, tested, corrected and expanded.
Open source does not mean zero cost
An open-source harness can be free to download and free of licence fees.
Running it properly is not free.
There can be model usage, hosting, integration, security, maintenance and implementation costs. More importantly, there is a cost to poor governance.
A badly configured agent with broad access can create more risk than a human team ever did.
The goal is not cheap AI.
The goal is owned, governed and useful AI.
Start with one workflow
Do not begin by trying to build an all-knowing digital employee.
Choose one painful, repetitive workflow.
Map the data it needs. Decide what it may read. Decide what it may change. Add an approval gate. Make it verify the outcome. Keep the audit trail. Then improve it until you trust the process.
Only then should you connect the next workflow.
The biggest lesson from building the Speakers Institute ecosystem is this:
You do not need to own the world's most powerful model.
You need to own the way your business uses intelligence.
Own the control plane. Own the permissions. Own the memory. Own the standards. Keep humans in charge of the decisions that matter.
That is how AI stops being another subscription and starts becoming part of the company.
The practical range is broader than most business owners realise. A properly connected agent can help edit video, create images, write copy, publish social content, prepare reports, inspect systems and coordinate work across the business. The important question is not whether an agent can touch the task. It is whether the task has the right instructions, permissions, approval gate and verification.
Sources for the privacy and architecture statements
- OpenAI business-data commitments: https://openai.com/business-data/
- Anthropic commercial-product training policy: https://privacy.claude.com/en/articles/7996885-how-do-you-use-personal-data-in-model-training
- xAI consumer and enterprise data controls: https://x.ai/legal/faq
- OpenClaw architecture and self-hosting overview: https://docs.openclaw.ai/help/faq/what-is-openclaw
Want help building your own AI agent?
If you want to build this properly inside your own business, I am running a 12-week Speaker Business program. We will show you how to build and govern your own AI agent, and there is also a pathway for us to build your agent with you.
AI agent FAQs
What is an AI agent harness?
It is the control layer that connects models, tools, memory, permissions, approval gates and audit records around an AI agent.
Can an AI agent run locally and in the cloud?
Yes. A governed setup can use a local sandbox, private cloud infrastructure or both, depending on the work, security controls and availability needs.
Do businesses need one AI model for every task?
No. The harness can route lighter work to lower-cost or open-source models and use paid models for more difficult reasoning and production work.
How should staff access be controlled?
Give each role and agent only the systems and information required for its work. Keep broader company access with authorised leaders and require approval for sensitive actions.
