SIG9
HOW IT WORKS

How an install runs, from workflow map to handover

An install starts with a map of one workflow and the numbers on it: how often it runs, whose hours it takes, and what a slip costs. I build the simplest system that removes the bottleneck, measure it against where it started, and hand it over running in accounts you own, with a guide to every step.

I start with the workflow

Most people who sell automation start with the tool. I start with the workflow: what sets it off, who touches it, which data moves, and where it waits on a person.

Automations break when the process is unclear, the data is messy, or the job needs real logic. Mapping the work first brings those to the surface while they're still cheap to fix.

The method, step by step

  1. 01
    Map the workflowI trace one workflow from the moment it starts to the moment it's done, and mark every place where work repeats, stalls, or moves only because someone remembered it.
  2. 02
    Put numbers on itI count how often it runs, whose hours it takes, and what it costs when it slips. Those numbers decide what gets built first and how much it's worth building.
  3. 03
    Build the simplest system that removes the bottleneckOne system around the step that holds the rest up, with fixed rules where the rules are firm, AI where a step needs judgment, and a person's approval wherever a mistake is expensive.
  4. 04
    Measure it against where it startedI check the number the system was meant to move, counted the same way it was counted on the map.
  5. 05
    Take the next oneThen I move to the next bottleneck on the same map, so each system after the first starts from work that's already done.

Every step gets sorted

On the map, each step gets one of five treatments. Plenty of steps should stay with a person, and I'd rather say so on the map than build something you end up switching off.

  • Automate it: the step repeats and its inputs are clean, so software can do it from start to finish.
  • AI-assist it: the step needs judgment on messy text, like working out what a lead is asking for.
  • Keep it rule-based: a fixed rule gets the right answer every time, like stopping junk mail, so AI would only add cost.
  • Route it to a person: a wrong answer costs more than the minutes saved, so software prepares the work and a person decides.
  • Leave it manual: it happens too rarely, or changes too often, to be worth building.

Honest status labels

Every system I build carries one of four labels, and it moves up a rung only once it has done what the next label says.

A ladder of four labels, from built at the bottom, through logic-tested and connector-tested, up to production, where a system runs for a client on the client's own accounts.
LabelWhat it means
BuiltAll of its parts exist in full. The label itself says nothing about live runs.
Logic-testedThe logic ran on sample data, with stand-ins for the outside services.
Connector-testedEvery outside service it uses was exercised for real on test accounts I own, including failures and reruns.
ProductionRunning for a client, on the client's own accounts. Nothing I have built is there yet.

No system I've built runs in someone else's business yet. I don't count my own business toward the top rung, so nothing I show you carries the production label. When I show you a system, I tell you which rung it has reached and what it hasn't proven yet.

Standards I build toward

These are the standards I build toward. Each system page says which ones that system has shown and which it hasn't yet.

  • When a step fails, an alert names the system, the step and the run, and goes to a named person.
  • Every run leaves a record I can check, so “what happened to that lead?” has an answer.
  • The same submission arriving twice does nothing the second time.
  • A person approves every reply the AI drafts for a lead before it goes out.
  • Emails your customers receive never carry internal reference codes.

What you keep at handover

  • The workflow map, with the numbers we put on it.
  • The installed system, running in accounts you own.
  • Your data, which stays yours.
  • Handover docs: a setup guide, and a plain guide to what each step does and why.

Questions

Do you use AI for everything?

No. AI goes where a step needs judgment on messy text, like reading what a lead wants. A step that a fixed rule can handle stays rule-based, which spends nothing on AI and gives the same answer every time.

What does connector-tested mean?

Every outside service the system touches, like an email account or a chat app, was exercised for real on test accounts I own. That includes forced failures and reruns. It's the rung below production: the parts work against the real services, and the system doesn't run for a client yet.

What do I have at the end?

A workflow map with its numbers, the system running in accounts you own, your data, and handover docs: a setup guide and a plain guide to each step.

What happens when something fails?

The standard is an alert that names the system, the step and the run, so whoever looks after the system can open the record of that run and see where it stopped. Each system page says where that is proven and where it isn't yet. A reply the AI drafts for a lead waits for a person to approve it.

Mashrur Rahman · Founder, SignalNinePUBLISHED · UPDATED