Adding AI and Automation to an Existing Business
Start with the work a machine should already be doing.
In plain English
Finding the work in your business that a computer should already be doing, and then doing it. In almost every case it is the boring internal work, not anything your customers ever see.
The technical version: Most of the early wins involve no model at all: they are integration, data plumbing and reporting. Keeping the automation separate from the AI matters, because calling deterministic work AI is how budget goes to the wrong project and how the result gets judged against the wrong expectation. The order that works is internal tooling first, one trustworthy source of truth per thing that matters second, and anything customer-facing last.
What you could actually do with this
Nobody retypes anything from one system into another, ever again.
That is where most of your errors come from, so quality improves at the same time as speed.
The weekly numbers arrive already assembled, instead of somebody spending Monday assembling them.
You get the decision earlier in the week, which is usually worth more than the hours saved.
Quotes, proposals and invoices come out of the systems that already hold the information.
They go out the same day instead of the same week, and they stop containing the last client's name.
The thing only one person knows how to do is written down and runs without them.
Your business stops having a single point of failure who is not really allowed to take a holiday.
What it is worth
Ask your team which part of their week they would delete if they could. They will answer straight away and they will already know the hours, because it is the part of the job they complain about. That list is the project and the business case at the same time.
What I actually deliver
The technical detail, for whoever on your side has to check it.
- A ranked list of opportunities with rough cost, rough saving and honest risk against each
- A clear first project, chosen because it pays back fast and teaches the organisation how to run this kind of work
- What not to automate, and why, which is usually the more valuable half
- The build itself, or a plan your team can execute without me
- Measurement, so "it saved time" becomes hours and dollars
How it works
- 01
Follow the manual work
Where are people retyping, re-checking, copying between systems, or assembling the same report every week? That is the map, and it is usually nowhere near where leadership expects.
- 02
Separate the automation from the AI
Most of the early wins need no model at all. Calling that work AI is how budget gets spent on the wrong thing, and how projects get judged against the wrong expectation.
- 03
Fix the data that everything else needs
One reliable source of truth per thing that matters. Skipping this is why the second and third project cost three times what the first one did.
- 04
Ship the first project narrow
One team, one workflow, measured. A narrow win creates internal appetite; a broad rollout creates internal resistance.
- 05
Then the ambitious one
With clean data, a team that trusts the approach, and a measured result to point at, the harder project becomes fundable.
When this is the wrong call
Do not start with a customer-facing chatbot. It is the most visible surface you have, the hardest to evaluate, and the most expensive place to be wrong, which is why so many of them get quietly switched off. Do not automate a process nobody has documented: you will encode the chaos and make it permanent. And do not buy a platform before you know the first project, because you will then choose projects that suit the platform.
Where I have done this
IC-1 Solutions
Onboarding, matching, compliance and reporting moved off spreadsheets into software that ran unattended. Zero to $50M revenue and 200+ staff in two years, then an 8-figure exit serving 500+ clients.
Further reading
- The AI Layoff Panic, Checked Against the Data Big tech keeps predicting mass white-collar job loss. Four years after ChatGPT, here is what the controlled studies and employment data actually show, including the one signal that is real.
- Should You Let AI Agents Touch Production? A CTO's Framework A CTO's decision framework for giving AI agents write access to production: blast radius, reversibility, permission scoping, sandboxing, audit logs, and kill switches.
- HIPAA Compliance for Health Tech Startups: A Builder's Guide What HIPAA actually requires of a software company: PHI, the Security Rule's safeguards, BAAs, and a builder's checklist to get compliant without over-engineering.
Questions
We are not a technology company. Is this still for us? +
Usually it is more valuable, because the manual work has never been looked at by anyone whose job it was to remove it. The staffing company I scaled to $50M was not a technology company either; the platform is what let it compete with firms that had one.
Will this replace our staff? +
In practice it removes the part of their job they already dislike: retyping, re-checking and assembling reports. That is also where the errors are, so it tends to improve quality and morale at the same time. Projects sold internally as headcount reduction meet resistance from the people whose knowledge you need to succeed.
How fast do we see something real? +
The review is about two weeks. A first internal automation is usually weeks after that, because it deliberately targets something small and measurable rather than the most impressive thing available.
What if our data is a mess? +
Then that is the first project, and it is worth knowing early rather than discovering it three months into something more ambitious. An expensive model reasoning over untrustworthy data produces confident wrong answers.