Automated delivery operations

Everything was manual.

On a Monday morning each project manager had to draft the report for their projects by hand and send it to a lead for review. Once it was approved, it was pasted into an email and sent to the client. Updates lived in group chats, emails and trackers, and sometimes the only way to get a straight answer was to call the project manager. For every report, the plan also had to be updated by hand in Microsoft Project and exported to PDF.

There was visibility, as long as you could find the right person. Nothing went wrong every day. It was just slow, and it depended on people remembering.

The company had become AI-native, and each leader was asked to train an agent for their own team. I had been waiting for this for years.

For about two weeks we taught it how we work. Then one day, for fun, I gave it a made-up project and asked for a plan. It came back with the plan, the risks and assumptions, the milestones, the parties involved and a RACI tab. A complete first attempt.

I told the whole team that day. That same month I asked our infrastructure team to cancel our Microsoft Project licence. The agent was now running our project documentation and reporting.

We did not give it one big briefing. On day one we told it who our clients are, what we deliver, what a delivery to a bank looks like and what one to a fintech looks like, how big the team is and how we work.

After that, we fed it a little every day. It saw the reports going out and the emails coming in. It joined our group channels and our meetings. It had access to our document library: old project plans, test scripts and project reports. The project managers kept giving it updates, and I kept checking that it was catching up.

It made mistakes. I once asked it for a slide and it copied an existing one word for word. I told it to follow our branding and to make each design its own. Each correction made it better.

In the early days it was shaky with reports. A report would be approved, and it would send a different one. Sometimes it sent to the wrong recipient. Sometimes I caught it and sometimes my team did.

So we added rules, one at a time. It must show the exact version it will send. Once a person approves that version, nothing can change it. It cannot send anything that has not been explicitly approved. Every report uses the same template and banner, and goes out from the team's address, never the agent's own. Over time it settled into a rhythm.

At the start the effort fell on the team. Everyone had to spend time training it, on top of a full workload, and I kept running orientations.

Then the project managers saw the difference. They stopped chasing documents and tracking projects by hand. They now build documents with the agent and ask it for help when they are stuck.

Some people worried it would take their jobs. I told them what I believe. The agent handles the admin: documentation, reporting and analysis. Project management still needs judgement, and with institutional clients you still have to talk to people to get things moving. The time the agent gives back goes into clearing blockers.

When a project starts, the agent builds the plan, the risks and the RACI. Updates arrive all week through emails, chats and meetings, and it files them in a running project register, so it always knows what a project has been through.

Tasks are shared. The agent hears someone commit to a call or a document in a meeting and sets the task for them. I can ask it to set one for myself or assign one to someone on my team. When a project moves from one project manager to another, both are told and the handover happens properly.

It spots projects that are slowing down, at risk of missing their date, or missing something it needs. It chases people for it in their chats. It writes the client reports, and a project manager reviews them. Blockers can be escalated by the agent or by a person.

In meetings it takes the minutes, stores them in the register and turns commitments into tasks.

The dashboard puts the whole portfolio on one screen: every project, its status, its last update and when it could go live. It also shows the team's tasks and the project requests that sales sends us. I can assign projects there, look up a contact, and message the agent.

Reports work the same way. I open the dashboard, see what the agent drafted and approve it, and it goes out in seconds. If something needs correcting, I ask the agent, and it is done in a couple of minutes.

It works like an app on my phone, so a missed email shows up as a notification, even if I am in the middle of a chat.

Mine was the first dashboard in the company. Leadership liked it and asked other teams to build theirs.

The hardest part was getting people to change how they work, and to put time into training the agent. I had to keep chasing the team to make sure they were not keeping information from it.

Two things helped. I told the team that if they wanted to reach me or escalate something, they should send it to the agent, and it would escalate to me. And I introduced an AI productivity index, which the agent calculates from how people work with it. I publish the ranking every week. Everyone wants a good place on the table, so usage went up.

My team has a digital co-worker. We have information at our fingertips and most of our processes are automated, so the team can take on more and produce better work.

We now oversee more than 30 projects with it, and the things that used to slip, such as a forgotten report or a misplaced document, mostly do not. Many projects have been flagged early, with a note like "this is missing, please send it" or "if this does not happen, the project will be late."

Next is automating the whole team's processes.

My advice for anyone starting is simple. Understand your own process end to end before you automate it. If you are not an expert in the manual work, automation will not make you one. Your pain points are your own working life, and they show you where the agent helps most. Then let it take the admin so you can be better at the job.