This page documents one of our own automations — the system that manages Elry’s AI automation projects. It is our operation, not a client engagement. We publish it because “we run our own company on the systems we sell” should be checkable, not taken on faith.
The process, before
Project management in a small delivery firm is a second job that eats the first. Estimating new work meant re-deriving effort from memory of past builds. Requirements lived in call notes and inboxes. Progress reports were written on Friday afternoons from whatever the week left behind. Scheduling meant holding every dependency in one person’s head. Research, testing evidence, and sign-off records were assembled by hand, per project, every time. All of it was done by the same two senior people who do the actual building — the process was consuming the most expensive hours in the company.
We didn’t formally measure the baseline before automating — we felt it in senior hours. Clients get that measurement done properly; the assessment discipline we sell exists partly because we know what skipping it costs.
What runs the projects now
The system has three parts, and the split between them is the point.
Our project management software is the system of record. Every project is a task tree created from a fixed template the moment an engagement starts: the assessment’s seven deliverables, the build phases, the acceptance criteria, the client dependencies with their dates. Nothing about that creation step is AI — it is fixed, tested software, the same structure every time.
The CRM is the front door. An enquiry — including the Describe-the-process form on this site — lands as a lead with its four answers attached. Stage moves in the pipeline drive what happens in the project system: a lead that becomes an assessment spins up the project structure; a proposal that is accepted schedules the build. The two systems stay consistent because the connection is deterministic, not because someone remembers to copy things across.
Cavalry is the workforce. Cavalry is our name for a roster of named AI agents that work inside our project management software, each with a defined role, its own knowledge base, and its own task queue. Across a project they:
- suggest the schedule and plan client workshop sessions;
- capture, validate, and store requirements;
- produce post-session findings and learnings documents;
- generate solutions plans;
- draft estimates against the record of past builds;
- read project state on schedule and draft the progress report.
Each agent logs what it did and why, in the open, where we can read it.
The boundary is the one we sell: the predictable steps — project creation, scheduling, state tracking, report assembly, evidence collection — run as fixed, tested software, the same result every time. AI does only the work that genuinely needs judgment — drafting an estimate, interpreting a requirement, shaping a solutions plan — and its output passes checks before anything leaves the system.
The human checkpoints
Four things never leave the building without a person:
- Estimates and pricing. An agent drafts; Terry decides. The draft is advice, the number is a human commitment.
- Anything client-facing. Progress reports and proposals are approved by the person accountable for the project before they are sent.
- The “don’t automate this” call. When the numbers say a process isn’t worth automating, a person makes that recommendation, in writing.
- Sign-off. Acceptance is a human decision against the criteria agreed before build. The system assembles the evidence; it does not grade its own homework.
When an agent hits something outside its brief — an ambiguous requirement, a dependency slipping, a decision above its authority — it does not guess. The exception routes to a named person as a decision request, with the context attached.
Built, and run since
Built in three weeks, June to July 2026, and in production since. It is now the required operating system for every new Elry automation engagement, including the work this site describes.
What it doesn’t do
Host client requirements-gathering sessions. Humans run the room. The system suggests the schedule, plans the workshop, and takes over once there is something captured — it does not replace the conversation where the real process shows itself.
This is Elry’s own operation, documented. No client data appears in it, and nothing is claimed beyond what was actually measured.