In its first week on TopDo.ai, ML LABS cut the time it spends on management work and sync overhead by roughly half. The gains held as the system grew, the opposite of what usually happens: in most tools, each added product, document and agent makes the next week of coordination a little slower than the last, until the tool itself becomes the work.
ML LABS is its own client here. Omar runs the company with AI agents, and TopDo.ai, a system ML LABS builds and operates for work that people and AI do together, is where that work is planned, done and recorded. The result comes from a few design decisions, and each one removes a job that someone would otherwise have to do by hand.
Overhead Halved In The First Week
- Management work and sync overhead fell by roughly 50% in the first week.
- The gains held as the system grew, with no slowdown from added scale.
- Status became something to look up rather than something to ask for.
Most of the overhead was never the work itself. It was keeping the plan and the work in agreement: a document updated while the task that implements it was not, a decision taken in a call that never reached the one who acts on it. The cost appears as re-work, status meetings, and decisions taken twice because the first record could not be found.
Status Answered Without A Meeting
Pick any process and ask three things without contacting anyone: how many items it holds, how long the oldest has waited, and where it is stuck. In most companies the answer is spread across people's memories, and assembling it is what the standing meeting is for.
The process mining manifesto (van der Aalst et al., 2012) sets out what a process needs before it can be measured: an event log where each event belongs to a case, names an activity, carries a timestamp and, ideally, names who did it. Spreadsheets and email leave artifacts instead, which prove something happened but cannot be counted or aged.
| Where status lived before | What the record holds instead |
|---|---|
| A shared folder of files | An identifier that travels with the item |
| An email approving something | A recorded approval, by a named actor |
| A spreadsheet column called "Status" | A state: open, closed or blocked |
| A weekly meeting that produces the count | A query that produces the count |
Coordination in real development work (Kraut & Streeter, 1995) falls back on asking people where formal records are weakest, and interrupted work (Mark, Gudith & Klocke, 2008) costs stress and effort. TopDo.ai records the work as it moves, so the answer is there.
Plan And Work Kept In One Record
Documents usually live in one product, tasks in another and decisions in a third, and each boundary is a seam where the plan and the work come apart. An AI agent cannot carry a broken link in its head: a link missing from the data does not exist.
TopDo.ai's model is three levels deep and one type wide. A workspace contains products; a product contains nodes; a node is the only content type. Each node carries a kind (item or doc) and a state that is open, closed or blocked. The specification and the task that implements it are the same shape of object, and the link between them is a real edge in the data, not a pasted URL that quietly breaks when a page is renamed.
Agent Work Proven As It Happens
An agent in TopDo.ai is an actor, not a macro. It holds an identity, a permission set and a history, and in the data it looks exactly like a person performing the same action, so the same rules, history and reversal apply to its work. The AI Risk Management Framework (NIST, 2023) reaches the same requirement through accountability and traceability.
Every change, by a person or an agent, goes through one layer that writes a hash-chained audit row in the same database transaction as the change itself. Each row carries the hash of the one before it, so no row can be rewritten or removed without breaking every hash after it. An audit written later, in a queue, can fail while the change succeeds and leave no trace of the gap; written in the same transaction, the record and the change cannot diverge.
An audit trail written after the fact is a story. One written inside the same transaction as the change it records is evidence an auditor can rely on.
Which actions an AI agent may take without a person is a separate decision, and the pattern that keeps consequential steps deterministic covers it: the model only proposes, and a rules engine, not the model, commits any step that cannot later be undone.
Gains That Hold As The System Grows
Most systems slow down as they grow because each new need arrives as a new tool or type, and each adds a seam. Coordination then grows with the number of people and agents writing across the seams, faster than the work, and an agent writes across all at once.
TopDo.ai grows along one shape instead. A new product is more nodes, a new document is another node, and a new agent is another actor under the same permissions and the same audit path. There is still one type, one state machine, one permission model and one recorded path for every change, so nothing new has to be reconciled. The interface an agent learns stays the same size however large the workspace ever grows.
That is why the first week's gain did not fade. The saving came from removing reconciliation, not from working faster inside it, and growth never adds any back.
The Same Record Runs Your Engagement
The order that made this work is data model first, automation second: an arriving agent inherits a legible system that is safe to run, rather than every seam and a write token. The context an agent needs is the operational half of the same design.
Every ML LABS engagement runs in TopDo.ai, so you see its state daily without asking, and the workspace stays yours after the engagement ends. To take one of your processes apart this way, its states and the record each step should leave, start with a first call.
References
- van der Aalst, W. M. P., et al. Process Mining Manifesto. Lecture Notes in Business Information Processing, 2012.
- Kraut, R. E., & Streeter, L. A. Coordination in Software Development. Communications of the ACM, 1995.
- Mark, G., Gudith, D., & Klocke, U. The Cost of Interrupted Work: More Speed and Stress. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2008.
- National Institute of Standards and Technology. AI Risk Management Framework. NIST, 2023.
