Moving up is a count
The record keeps each agent's run of approvals, so a promotion is a number on the record, not a judgment call.
How TNT runs
We run on what we sell.
TNT is its own first client. It runs on the Microsoft stack it builds on, behind the same approval queue it sets up for you. Here is what runs TNT today, with dates, what TNT built to sell, and what is still being built.
In short
Three things to know before the detail.
One AI app is live
The meeting assistant prepares Thiago for his meetings and drafts what comes after. TNT's drafting agents are being built behind it.
A person approves
Every agent draft waits in a Teams queue until someone approves it, and that queue is live.
Status, not promises
Each system below says live, tested, in build or planned, and nothing is described as working before it does.
What's true today
Facts from our own records, each with the date we checked it.
No estimates and no targets dressed up as results. When a fact changes, so does its date.
2 seconds
From a website enquiry to its acknowledgement email, in test runs on TNT's own intake flow.
As of October 16, 2026
A person approves
Agent drafts go to a Teams approval queue; nothing acts until someone approves it.
As of October 16, 2026
What runs TNT
Live, in build, planned.
Everything that runs TNT, or will, with its real status. Nothing here is described as working before it does.
| System | What it does | Status |
|---|---|---|
| Website, email and booking intake | Turns every enquiry into a lead on one record and acknowledges it in minutes | Live |
| Teams approval queue | Every agent draft waits here for a person to approve, edit or reject | Live |
| Meeting assistant | Briefs Thiago before his meetings and drafts the notes, action items and follow-ups after, from Teams transcripts | Live |
| Control plane | One view of what each agent did, what's waiting and what it cost, with one switch that stops them all | In build |
| Scorecard agent | Drafts the free AI-Era Readiness Scorecard from a company's answers, its public footprint and any figures it sends us | In build |
| Content agent | Drafts the newsletter and LinkedIn posts for approval | Planned |
| Call-prep agent | Writes a one-page brief from the lead's record before every call booked on the site | Planned |
| Proposal agent | Turns an Assessment into a scoped proposal and statement of work | Planned |
| Website assistant | Answers questions about stages, offers and the two published prices, and books calls | Planned |
Live: running for TNT today. In build: being built and tested now. Planned: on the plan, not running for TNT yet.
What TNT built to sell
Two apps we built, and exactly where each one stands.
Neither has run for a client yet. Each says what it does and how far it has been proven.
| App | What it does | Where it stands |
|---|---|---|
| Meeting assistant | Walk into each Teams meeting briefed. Walk out with the notes, action items and follow-up already drafted for you to send. It's set up in your own Microsoft 365 and works from your Teams transcripts. | Live Live at TNT: Thiago uses it every day for meeting prep, notes and follow-ups. Not yet set up for a client. |
| Google Workspace migration | Moves a company from Google Workspace to Microsoft 365 with tooling TNT built: Drive files to SharePoint with links fixed and sensitivity labels applied, Chat spaces to Teams, and Gemini chat to Copilot Chat. | Tested Built by TNT and tested on test tenants. Not yet run for a client. We'll say that again before you sign. |
Live: in use at TNT today. Tested: built by TNT and tested on test tenants, not yet run for a client.
The meeting assistant Google Workspace migrationAutonomy levels
No agent starts with autonomy. Each one earns it on a measured record.
Every agent has a level, kept on the record and read before every run. Every agent starts at L1, drafting for a person to approve, and moves up only on its own numbers.
| Level | What the agent does | What a person does | To move up |
|---|---|---|---|
| L0 Manual | Nothing | Everything | Where calls and partner work start, and stay |
| L1 Draft | Prepares the work and posts it to the Teams approval queue | Approves, edits or rejects it with one tap | 20 approvals in a row with no edits, or 50 with fewer than 1 in 10 edited |
| L2 Act, sampled | Acts, logs every action and flags exceptions to the queue | Reads a daily digest and reviews 1 output in 10 | 30 days with no reversed action and no complaint |
| L3 Autonomous | Acts and monitors itself | Reads the weekly report and acts on alerts | Stays here until an incident drops it to L2 |
Dropping a level takes one mistake
A single reversed action or complaint drops an agent one level, whatever level it has reached.
What stays human
Six things never go to an agent, whatever its record.
These stay with a person every time, at every level.
Discovery calls
The first call and the Assessment conversation are always a person talking with you.
Partner relationships
Partners and referrals are handled by a person. LinkedIn messages are typed and sent by a person.
Contracts and prices
Signing a contract, and any price outside the published menu.
Your acceptance
Whether delivered work is accepted is your decision, made with a person.
Payments
Any payment or refund outside the standard invoice flow.
Naming a client
Anything that names a client publicly, which also needs that client's written approval.
Control
One queue to approve. One switch to stop every agent.
The queue is where a person says yes. The global stop is where a person says stop.
-
The approval queue in Teams
LiveEach draft arrives as a card in a Teams channel. A person approves, edits or rejects it, and the decision is written back to the record. Those decisions are what move an agent up a level.
-
The global stop
In buildBefore every run, each agent reads its own pause flag and one global stop from the record. Set the global stop and no agent starts another run, with nothing to redeploy. The switch is set from the control plane.
Architecture
Six layers, with one record in the middle.
How TNT is designed, in plain words; the table above says which parts are live. Nothing passes between layers except through the record, and the same design goes into your own Microsoft 365 and Azure when we build for you.
Channels in. Where work arrives: the website, email, Bookings and forms, and later partner referrals and the website assistant.
Events. Power Automate turns every event into a record that agents pick their work up from, and routes approval cards and alerts.
Record, agents and models. Dataverse is the one record and the only data an agent acts on. Drafting agents run as short jobs on Azure Container Apps; the website assistant runs in Copilot Studio. Each agent's model is set in configuration: Azure OpenAI by default, the Claude API as an alternate.
People. The Teams approval queue, and the control plane: a dashboard in our Power Apps operations app, where each agent's pause and the global stop are set.
Actions out. Only connectors and the deployment pipeline touch the outside world, and only after an approval or at a level an agent has earned.
Identity and platform. Entra ID, with managed identities and certificates in Key Vault instead of stored passwords. Logging, a budget per agent, and every change deployed from GitHub through a pull request.
A lead, traced
The website form posts to the intake flow. The flow writes the lead to the record and sends the acknowledgement email, and a person picks it up from there.
An approval, traced
A person approves a card in Teams. The decision lands on the record, the action runs through a connector, and the run log counts toward the agent's next level.
Client workloads never run in TNT's systems. They run in systems the client owns, its own Microsoft 365 and Azure, reached through delegated, least-privilege access.
See how it works
Book a call and we'll show you how TNT runs.
Thirty minutes with Thiago, without showing anyone else's meetings or details. Then, if you want to know where your own company stands, start with the free Scorecard.