Plan the guardrails, then just start with AI

There is no fixed AI recipe for your business, so a six-month AI strategy is out of date before it is finished. Plan the guardrails and the goal, put safe AI in everyone's hands, and learn from what breaks.

Finlay Ekins
Co-founder · · 5 min read
Get the next drop
Share

The best way to start with AI in a 20 to 100 person business is to start, safely, rather than to plan. There is no fixed AI recipe for your business or your industry, and the tools change every day. Plan the guardrails and the goal, then put AI in your team's hands and learn from what breaks.

That cuts against a line most of us have heard. "If I had an hour to solve a problem, I'd spend 55 minutes thinking about the problem and 5 minutes thinking about solutions." It is often credited to Einstein. Quote Investigator found no evidence he said it, and the earliest version it could trace was a 1945 leaflet called "Get It Right the First Time".

With AI, you won't get it right the first time. Nobody does.

Why doesn't a perfect AI plan work?

A perfect AI plan doesn't work because there is no fixed recipe to plan from. Every business is different, every vertical is different, and the tools you would be planning around change every week.

So the six months or a year you spend writing the perfect AI strategy is time the strategy spends going out of date. By the time it is approved, half the tools in it have changed and the other half have competitors.

This is why a startup way of operating works so well with AI. You start, you see where it breaks, you fix it, and you go again. If someone has convinced you there is one fixed way to roll out AI, they are wrong, because the thing they are describing is different every day.

Unless you make it an iterative, experimental rollout, you won't get AI right.

What happened when a property group put Claude in everyone's hands?

The clearest example I have is a Melbourne property group we work with.

When we started, five people there were using AI, mostly as a search box. Most people didn't know where to start with a tool that broad.

They didn't write a strategy. They set Claude up safely, with the admin gating access, and then put it in everyone's hands. The mindset was: let's get this in front of everyone, push it, and see how it goes.

Real AI champions rose out of that. Seven weeks after kickoff, most of the office was using it, and staff with no coding background were building their own tools. People learned skills, took it home on the weekends and taught themselves more, and pushed their own work forward. The team is now productive on Claude, and the questions have moved on to custom automations and agents of their own.

None of that would have come out of a planning document. It came out of starting small, keeping the admin gated and safe, and putting AI in the hands of everyone.

What did moving our agents onto a server teach me?

Moving my agents onto a server was painful, and it was worth it. I run my development on an always-on server now, with agents working around the clock. Getting there was not smooth.

It kept going down. Things kept breaking. One missing firewall rule alone cost us weeks of flaky connections, which I wrote up in the server guide. No plan would have found that rule. Running it and watching it fall over did.

Once it worked, it was far more valuable than the old setup. I can run things overnight, and in October I shipped more than double the work I shipped in each of the three months before, measured in merged pull requests.

You have to go through that stage and figure it out as you go. There is no version of this where you skip the breaking part.

When does planning still matter?

Planning still matters for the guardrails and the goal, just not the path.

Decide up front what you are trying to improve and what is off-limits. That part deserves real thought, because getting it wrong is how a rough experiment turns into a data or security problem. Then leave the how open, and let the experiments tell you which way to go.

How do you experiment with AI safely?

You experiment with AI safely by setting three guardrails before anyone starts, and then letting people play inside them.

  1. Approved tools and data rules. Pick the AI tools your team can use, on business accounts, and agree what data never goes in, such as client personal information. The OAIC's guidance on commercially available AI products covers the privacy side.
  2. A human check before AI acts. The AI drafts, a person approves, and nothing goes to a client or another system without that check. My own agents still check with me before they act on anything.
  3. A small pilot with one team. Start with one team and one workflow, learn from it, then widen the rollout.

That is enough to start. As your AI use grows, Australia's Guidance for AI adoption sets out the fuller set of practices, and it makes the same point: you don't need to do everything at once. When you are ready to trust an AI answer at scale, write a set of test cases first, the way Anna does with 20 known answers.

What should you do this week instead of writing an AI strategy?

This week, set up a small, safe space and start playing.

Make a folder on your computer and start using Claude to build something that gets your internal work moving faster. It doesn't have to be connected to anything. If that is safer for you, keep it disconnected. The point is to start experimenting and get curious.

That curiosity is what builds the capability. Every small thing you build teaches you what AI can and can't do in your business, and that is how you learn and grow fast. A plan can't do that for you.

Finlay Ekins

Written by

Finlay Ekins

Co-founder

Finlay Ekins is the co-founder of Hourglass AI, where he leads the technical side of the build: the agents, the integrations, and the engineering judgement about what will hold up in production. He studied mechatronics engineering in Melbourne and left it for AI when it became clear the intelligence, not the hardware, was the interesting part. He started the business that became Hourglass on his own, made $700 in its first month, and spent ten months of failing and pivoting before it worked. He writes a weekly newsletter on AI from Australia at finlayekins.com, 14 essays since June 2026, and his standing advice to clients is to try ChatGPT properly for a month before commissioning anything custom.

More from Finlay

Common questions

Where should a 20 to 100 person business start with AI?

Start with one small, safe experiment rather than a strategy document. Set up an approved AI tool on a business account, agree what data never goes into it, and have someone try it on one repetitive internal task this week. What breaks in that first week tells you more about your business than a six-month plan would.

Do you need an AI strategy before you start using AI?

You need a goal and guardrails, not a detailed strategy. Decide what you want to improve and what is off-limits, such as client data or sending anything without a human check. Then iterate on how. AI tools change every week, so a fixed plan for the path goes stale before you finish writing it.

What guardrails should be in place before staff experiment with AI?

Three to start: a short list of approved AI tools on business accounts with clear rules on what data can go in, a person checking anything before the AI acts on it, and a small pilot with one team before rolling out further. Australia's Guidance for AI adoption and the OAIC's privacy guidance are good references for the detail.

How long does it take for AI to catch on inside a team?

In our experience, AI champions catch fire in about seven weeks once everyone has access and works on real tasks. That is how it went at a Melbourne property group we work with: seven weeks after kickoff, most of the office was using it. Admins gated access so it was safe, and some staff went on to build their own skills on the weekend.

AI Strategy ServicesImplementation PlanningDigital Transformation

References

  1. 1. Quote Investigator: I Would Spend 55 Minutes Defining the Problem and then Five Minutes Solving It: No evidence Einstein said it; the earliest version found is a 1945 leaflet. Read 16 November 2026
  2. 2. National AI Centre: Guidance for AI adoption, implementation guidance: The Australian Government's fuller set of AI practices, and its advice not to do everything at once. Read 16 November 2026
  3. 3. OAIC: Guidance on privacy and the use of commercially available AI products: The privacy side of choosing approved AI tools and data rules. Read 16 November 2026
  4. 4. Hourglass AI: Your own coding-agent server: How I moved my agents onto a server. Read 16 November 2026

Talk to the founders about your team.

  • In 30 minutes, we find the 3 jobs eating the most hours in your team's week
  • We show you what we've already built for businesses like yours
  • You leave with a written plan for what to build first
Michael Batko
Michael BatkoCo-founder, CEO
Finlay Ekins
Finlay EkinsCo-founder, COO

You talk directly to the founders.

Rather read first? Get each new skill and playbook by email.

Unsubscribe in one click.