ITMedic · Green IT Consulting
Sixteen AI agents run the back office of this consulting business. Each one has a written job description — a background, a specialism, decision rules, and a boundary it cannot cross. One of them manages the other fifteen. And there is a single screen where the owner watches all of it happen.
This is not a pilot and it is not a demo. It is how the business actually ran today.
The problem
It is not the work. Owners are usually excellent at the work. What is missing is the layer above it — the one that notices the invoice never went out, that the quote is 28 days old, that the only recurring client leaves in six days, that the bank swept the account twice and nobody asked why.
In a large company that layer is a management team. In a small one it is the owner, at night, on a laptop, after the actual job is done. It gets skipped in the weeks it matters most, and it disappears entirely the week the owner is ill.
A single point of failure is not a staffing problem. It is a design problem — and it is the one we set out to fix.
What changed
An AI tool answers a question. An agent holds a job. The difference is not the model — it is everything written down around it.
Each of our agents carries a real charter: the professional background it reasons from, the specialism it owns, the sources it is allowed to trust, the decision rules it applies, what it must escalate, and what it may never do. That document is the agent. Swap the model underneath and the role survives.
Our cash agent reasons like a controller — direct-method forecasting, bank as arbiter, committed versus expected held apart. Not "a chatbot given a spreadsheet."
Bill within three days. Never a third email on a silent thread — phone instead. Named, numbered, applied whether or not anyone feels like it.
Draft, never send. Propose a price, never set one. Every irreversible act stops at a queue and waits for a human.
Every ruling the owner makes is written back. Decide once that an old receivable is not collectible, and no agent ever raises it again.
The roster
Real names, because you talk to a colleague differently than you talk to a tool. Each reports into Atlas, who ranks the week.
Command

Ranks every signal in the business, sets the operating order, and decides what reaches the owner at all.

The daily brief, the approval queue, and the continuity runbook that lets the business run without its owner.
Governance & intelligence

Thirteen-week forecasts on a cash basis, bank reconciliation, and the unglamorous question of what is actually in the account.

Account briefs, ideal-customer profiles, and verifying that the person you are writing to is still in the seat.

Solicitations and competitive reads, with a bias toward telling you not to bid.

Which motions returned anything, and which ones only felt productive.

Tiers, rate cards, and value anchors — priced, never approved. That signature stays human.

The dates that quietly end things if nobody watches them. Timing in the open; substance stays private.
Business & relationships

The call list, the pipeline of record, and the rule that a third unanswered email is a phone call.

Which client has gone quiet, and which sentence in an inbox is the first sign of a cancellation.

Grants, lending and funded revenue, ranked by speed-to-cash rather than headline value.

Turns a worry into one page counsel can answer in minutes — and stops work that should not ship yet.
Operations & engineering

The wiring between systems, and the scheduled runs that happen whether anyone is watching.

Campaign shape, audience, and what a channel is actually returning.

The unglamorous half — lists, templates, deliverability, the things that quietly break.

Posts, newsletters and client letters in one voice, drafted for approval.
The manager
Fifteen agents each reporting their own findings is not a management team — it is fifteen more inboxes. So one agent sits above them.
Atlas reads everything the others produce, ranks it against a fixed signal hierarchy, and decides what the owner sees. It sets the operating order on Monday and writes a decision memo on Friday. It is the only agent that can mark something top priority, and it is capped: no more than three at a time. Everything else waits, visibly, in order.
That cap is the design. An assistant that surfaces forty things has moved the problem, not solved it. The job of a manager is to not bring you thirty-seven of them.
The screen
Every agent writes its state to one shared ledger, and that ledger renders to a single dashboard: what each agent is doing right now, what it is blocked on, what it plans next, what changed since you last looked, and the short list of decisions waiting on you.
Not a report about last month. A live picture of the whole operation, including the parts you are not personally touching.
Click through a working dashboard
Populated with data from a fictional company, so you can click everything without looking at anybody's books.
The wiring
Agents that work in a sandbox produce sandbox answers. Ours read and write the systems the company actually runs on.
| System | What the agents do with it | Status |
|---|---|---|
| Microsoft 365 Outlook, Calendar, SharePoint, To Do, Teams | Read the real inbox and calendar, write drafts into the owner's Drafts folder, file work to SharePoint, push decisions to To Do on the phone and watch. | Live |
| M365 admin center | The agent roster registers as a connected platform, so these agents appear in the same governed inventory as any Copilot Studio agent — named, attributed, visible to IT. | Live |
| QuickBooks Online | Invoices, A/R and A/P ageing, the balance sheet and P&L behind every forecast. | Live |
| Bank statements | Reconciled against the books, because the bank is the arbiter and the ledger is an opinion. | Live |
| Power Automate | Carries agent items into To Do, Planner and Teams, and carries your completions back. | In build |
| Dynamics 365 Sales | Destined to hold the pipeline of record. Today that pipeline is a maintained file, and we would rather say so than imply otherwise. | Planned |
The M365 admin-center piece is the one that surprises people. Two ecosystems usually written up as competitors, cooperating — with the governance trail intact. That matters the moment anyone in a regulated business asks what is touching their data.
The guardrails
This is the part we would want to hear first, so it goes here rather than in a footnote.
These are enforced in the system, not written in a policy nobody reads. That distinction is the whole difference between an agent workforce you can defend to an auditor and one you cannot.
Why we are telling you
This was not a product plan. It was a small consulting business with one owner, a major surgery on the calendar, and a recurring client leaving — needing to keep running regardless. We built the thing we needed, in production, on our own books, with our own money at stake.
That is the only credential we would trust if the positions were reversed. Anybody can demonstrate an agent. Rather fewer are willing to show you the dashboard of the company they run on it.
If you are a business owner who recognises the missing layer — the invoice that never went out, the quote nobody chased, the week that fell apart because you were ill — we should talk. Not about AI. About which three things in your business are currently held together by you remembering them.