TNT Technologies Insights

Why Copilot seats go unused, and what to fix first

2026-10-19T15:00:00.000Z

A company buys Copilot seats. For a week or two, people try it. Someone asks it where things stand with a customer and gets a vague answer. Someone else gets an answer that quotes a file they didn’t know they could open. By the end of the month most of the seats are quiet, and the licenses renew anyway.

The usual reaction is to book a training day. Training is the last thing I’d fix. Copilot isn’t failing because people can’t write prompts. It’s failing because of what it can and can’t see. Here’s the order I’d work in.

1. Data access: can Copilot reach the answer?

Copilot answers from the content a person can already open in Microsoft 365: their mail, calendar, chats and meetings, and files in OneDrive and SharePoint. That’s the point of it, and it’s also the first limit.

If the customer list is a spreadsheet on someone’s laptop, the contracts sit on an old file server and the real history of each account lives in a separate CRM, Copilot has nothing to work with. It isn’t broken. It’s blind.

The fix starts with a short list. Write down five questions people would actually ask Copilot in a normal week, then find where each answer lives today. Anything outside Microsoft 365 either moves in, gets connected, or comes off the list. You’ll learn more from that hour than from a day of training.

2. Permissions: what can it see that it shouldn’t?

The opposite problem stops rollouts cold. Copilot shows each person only what they already have access to, which sounds safe until you look at what people have access to. The HR folder shared with the whole company. The site someone made for a project years ago, open to everyone, with payroll exports in it. Copilot didn’t create those permissions. It just makes them easy to find.

One surprise like that and the owner quietly decides not to buy more seats. So before you expand, look for content shared with everyone in the company or with anyone who has the link, starting with HR, finance and anything holding customer or patient data. Close what shouldn’t be open, and give each of those areas a named owner.

It’s dull work. It’s also the work that lets you say yes to Copilot without worrying about what it will surface next.

3. One task: give it a real job

“Use Copilot” isn’t a job. People don’t change how they work because a new button appeared. They change when one task they do every week gets easier.

Pick one recurring task with a clear owner. The follow-up email after a client meeting. The weekly status summary for the leadership meeting. The first draft of a proposal from the notes of the last three calls.

Sit with the person who does it, write the prompt together, save it where colleagues can reuse it, and check after two weeks whether it saves them time. If it does, pick the next task. If it doesn’t, you’ve learned something cheaply.

How to tell where you stand

Open the Copilot usage report in the Microsoft 365 admin center and compare the seats you’ve assigned with the people who actually used Copilot in the last month. If the gap is wide, start at step one. Not with a training day, and not with more seats.

On our four-stage ladder, this is the work of reaching Assisted: Copilot and AI tools licensed and governed, a security baseline in place, and the data cleaned and reachable. It comes before any agent, because an agent built on scattered data and loose permissions inherits both problems.

If you’d like an outside view first, the free Scorecard includes the value you’re not using yet.

Why AI pilots stall: the four stages, and how to find yours

2026-10-19T14:00:00.000Z

You probably know the story. An AI pilot looked good in the demo. It answered questions, drafted emails, maybe booked a meeting. Then it met the real company and stopped. Nobody decided to kill it. It just never became the way the work gets done.

The model is rarely the reason. More often, the pilot assumed a stage the company hadn’t reached: an agent that needs one clean record of customers, built on top of five spreadsheets, or a Copilot rollout in a tenant where nobody knows who can see what. You can’t skip a stage. You can see which one you’re on, and do the work that moves you up.

We use four stages. Every Scorecard and Assessment places a company on them, and we score TNT on the same ladder.

Stage 1: Pre-AI

Manual, disconnected systems, AI used by individuals in free tools, no policy, no data foundation.

You’re here if people paste work into free AI tools on their own, customer and operational data live in separate systems and spreadsheets, and nobody owns security end to end. That isn’t a failure. It’s where every company starts. What moves you up is a security and insurance baseline, an AI policy, and data you can reach from one place.

Stage 2: Assisted

Copilot and AI tools licensed and governed, security and insurance baseline in place, data cleaned and reachable.

You’re here if your AI tools are licensed with a policy for how they’re used, MFA, backups and endpoint protection meet what your insurer asks for, and the data that matters can be reached from the tools your people use every day. What moves you up is your first agents running defined workflows, with people approving the exceptions.

Stage 3: Operated

Agents run defined workflows with people approving exceptions, on one record, with a control plane.

You’re here if agents draft and act on defined workflows while people approve what falls outside the rules, every action ties back to one record of customers, work and money, and one view shows what each agent did, what’s waiting and what it cost.

Stage 4: Native

The company is designed around agents and people, new work starts as an agent, and the technology function runs as a managed service.

You’re here if new work is designed as an agent with a person accountable for it, agents earn the right to act alone on a measured record and lose it on a single reversed action, and infrastructure, security, data and agents run as one managed service.

Three questions to find your stage

Does someone own security and AI use end to end, with a written policy? If not, you’re at Pre-AI, whatever you’ve bought.
Is the data your people need cleaned up and reachable from the tools they use every day, and could you answer your insurer’s security questions today, with evidence? If so, you’re at Assisted.
Does at least one agent run a real workflow on one record, with a person approving the exceptions and a log of every action? If so, you’re at Operated.

If you can say yes to all three and new work already starts as an agent, look at Native.

Be strict with yourself. A company with an impressive agent demo and nobody who owns security hasn’t reached Operated. It has a demo.

What this means for your stalled pilot

Look at what the pilot needed: clean data, clear permissions, a person who owns the result. If those belong to a stage you haven’t reached, the next move isn’t a better pilot. It’s the stage work underneath it. Once that’s done, the same pilot has somewhere to land.

The four stages page has the full ladder, and the free Scorecard tells you where you stand today.

How our intake works: four ways in, one record

2026-10-19T13:00:00.000Z

The first thing anyone sees of TNT is what happens after they ask us something. A company that sells “one accountable partner” can’t lose an email in a shared inbox. So intake was the first function I built, and it’s the one you can test yourself. Everything I describe as live below was tested on October 16, 2026.

Four ways in

You can reach us four ways: the form on this site, an email to info@tnttech.io, a booking on the 30-minute call calendar, or the Microsoft Form we used before this site had its own forms, which we’re retiring. Each one is a Power Automate flow, and all four write to the same table in Dataverse.

That table is the one record. If you fill in the form and then book a call, you’re one lead with both on it, not two leads in two systems. Every agent we build reads from that record and writes back to it, so nothing depends on someone copying details from one place to another.

An answer in seconds, a reply from a person

The acknowledgement is automatic. In our test runs it went out about two seconds after a website form and about one second after a Microsoft Form response. We promise it within minutes, because a promise should hold on a bad day too.

The acknowledgement isn’t the reply. A person reads every request and replies within one business day. Each new request also posts to our Teams channel with what you told us, so the reply doesn’t depend on someone watching an inbox.

The Teams card

Everything an agent drafts goes to an approval queue in Teams before it goes anywhere else. It’s a card with two buttons, Approve and Reject, and the decision is written back to the record. Nothing leaves until a person taps one. We proved that round trip in production on October 16.

Today the queue is live and the agents that will fill it are not. The Scorecard agent is in build. The content agent and the call-prep agent are planned. When they come online, they start where every TNT agent starts: drafting only, with a person approving each draft. An agent earns the right to act on its own after a measured run of approved work, and one reversed action drops it back a level.

What broke on the way

Two things, both small.

First, email. Our info@ address is an alias, and Exchange resolves an alias to the mailbox’s main address before a flow sees the message. The first version of the flow filtered on “sent to info@”, so it never fired. We dropped the filter and skip our own domain instead.

Second, Bookings. The flow was watching an old booking page, not the calendar the button on this site uses, so a call booked from the site never reached the record. We pointed the flow at the right calendar and booked a test call to prove it.

Neither would show up in a demo. Both would have cost us a lead.

What’s next

The control plane is in build: one view of what each agent did, what’s waiting and what it cost, with one switch that stops them all. So is the Scorecard agent, which will draft each free Scorecard from the request this intake creates, for a person to approve.

None of this is exotic. It’s Microsoft 365, Power Automate and Dataverse, the stack a growing company already pays for or can add. Booking and intake is also one of our quick-win workflows. And if you’ve used the form on this site, you’ve already seen it work.

The full list of what runs TNT, with what’s live, in build and planned, is on How TNT runs.