Connected systems
Connected systems are the quiet layer: the automations, integrations, and AI workflows that stop work from being copied by hand between tools. I design them so they save effort without taking control away from the people who rely on them.
- Typical work
- Automation, integrations, AI-assisted steps
- Designed for
- Clear rules, visible results, safe failure
- Human in the loop
- Where a decision carries real weight
- Handed over
- Documented, so you can maintain it
The short version
Repetition is where good work leaks away.
Somebody retypes a customer’s details into a second system. Somebody reformats a report every Friday. Somebody waits on an email that would have been sent automatically. The task is small each time, and the sum of it is a working day spent moving information rather than improving it.
Automation is not glamorous, and that is the point. It removes the friction that people have stopped noticing, and it gives the same result every time instead of a slightly different one.
Where the time goes
These are the patterns I look for first, because they tend to be hiding in plain sight.
- Copying between tools. The same record typed into two systems, then corrected later because they drifted apart.
- Chasing status. A person waiting to find out where something is, instead of being told.
- Rebuilding the same report. The numbers are ready; the packaging is done again every week.
- Reformatting for different audiences. The same information rewritten for a customer, a colleague, and a filing.
- Checking by hand. A human repeating a lookup that a rule could do reliably.
What I build
Workflow automation
When something should happen, it happens: a new enquiry becomes a task and a calendar entry, a completed form updates a record and starts the next step, an overdue item notifies the person who can act on it. The rules are written down in plain language so they can be argued with.
Integrations
Two tools that should know about each other do. A form writes to the system of record, a payment updates a status, a calendar event creates the project. The point is that the data is entered once and stays consistent.
AI-assisted steps
Where language is the bottleneck, a model can help: summarising a thread, drafting a reply from a set of facts, extracting the important parts of a document, sorting a request into the right category. The output goes to a person when being wrong would cost something.
Notifications and hand-offs
Good automation tells you what happened. A workflow that finishes silently is a workflow nobody trusts, so each step reports its result in plain language, including the ones that failed.
Keeping people in control
This is the part I will not compromise on. A system that runs on its own should still be understandable by the people depending on it.
- Visible steps. Each stage is named, and a person can see where a piece of work is and what it is waiting for.
- Reversible actions. Where an action can be undone, it is. Where it cannot, a person confirms it first.
- Failures that speak up. A broken step stops the workflow and tells the people who can fix it, with the context needed to understand it.
- People where it matters. Decisions with legal, financial, or reputational consequences keep a human review step, even if the rest of the workflow is automatic.
- No hidden dependencies. If a workflow depends on a third-party service, that dependency is documented, along with what happens when the service is unavailable.
The result should feel boring in the best way. The work arrives where it should, in the format it should be in, and nothing surprising happens.
Where it fits
A good fit
- Repetitive work that crosses two or more tools
- A team growing faster than its manual processes
- Reporting that is prepared on a schedule every week
- Requests that need sorting before a person can act
Not the right shape
- Work that is genuinely one-off and irregular
- Automation of decisions that should stay human
- Processes nobody has agreed on yet
- Systems that need guarantees no platform can give
Questions about connected systems
Where should AI be used in a workflow?
On the steps where a person currently retypes, reformats, or summarises the same kind of thing, and where being slightly wrong is easy to spot. I keep AI away from decisions that carry legal, financial, or reputational weight unless a person reviews them.
What happens when an automated step fails?
It stops and says so. Failures are reported to the people who can fix them, with enough context to understand what happened, and a retry that is safe to run. Nothing carries on quietly with bad data.
Will this lock us into a platform?
I aim not to. Workflows are described in ordinary language, the data flows through documented steps, and where a platform does the work, the alternative is written down. Leaving should be a decision, not an extraction.
How much of this can run without me?
After handover, the workflows run on their own and you maintain them. I document how each one behaves, what it depends on, and what to check when something changes upstream.
Related work
Automation needs somewhere to deliver its results. Dashboards and tools give the output a place to land, web experiences handle the parts customers see, and search and growth keep published information current without manual effort.
What keeps getting done by hand?
Tell me the task you have stopped noticing. That is usually the one worth automating first.