Fix your foundations before automating.
Most of the AI work being sold right now is automation first. It is the wrong order, and it costs the businesses that buy it more than they think. Here is what we believe instead, and why.
Automation does not fix a broken process. It makes the broken process faster.
Walk into almost any agency pitching AI and the conversation starts in the same place. Which tasks can we automate. Which agents can we ship. How many hours can we take off your plate by the end of the quarter. It sounds like progress. It photographs well in a deck.
The problem is what happens underneath. When you automate a process that does not work, you do not remove the problem. You encode it. Every workaround, every undocumented exception, every handoff that only succeeds because a person quietly fixes it in the moment gets baked into a system that now runs without that person watching. The mess is still there. It just moves faster and answers to no one.
Then growth arrives, and the cracks open. The automation that looked tidy at ten transactions a day starts producing wrong answers at a thousand. Nobody can say why, because the logic was never understood in the first place. It was inherited. You end up slower than before, more fragile than before, and paying down tech debt you chose to take on.
Plenty of good businesses are getting conned into automating processes that never worked in the first place.
That is the part that bothers us. The client did not ask for tech debt. They asked for help, trusted the people selling it, and walked away with a more complicated version of the thing that was already costing them money. The incentive to sell automation early is obvious. It is shiny, it bills well, and the consequences show up later, usually after the invoice has cleared.
Run it by hand first. Many times. Then automate what already works.
Good engineering teams do not ship to production on the first idea. They pilot. They run the thing in the small, watch where it breaks, and only widen the road once the route is known. We do the same with process. Before a single agent gets built, the process gets run manually, over and over, until the way it actually behaves is no longer a guess.
Think a hundred reps, not three. The point is to make every mistake by hand, where mistakes are cheap and visible, instead of discovering them later inside an automated system where they are expensive and silent. By the time we are ready to automate, there is nothing left to learn about the process. The unknowns have all been spent.
Only then does it get built. And even then the job is not finished, because a process worth automating is a process worth measuring. We watch the outcome, compare it to the manual baseline we already established, and keep tuning. The automation earns its place by being better than the hand, on numbers, not on vibes.
Diagnose
Map the processes, the data, and the systems. Find the real constraint, not the one that is easiest to talk about. Most of the time the bottleneck is somewhere nobody is looking.
Experiment
Run the process by hand. Do it again. Keep doing it until the failure modes are obvious and the fix is proven. Roughly a hundred reps before a line of automation gets written.
Automate
Build automation only for the processes that already work. Then measure the outcome, watch it in production, and keep improving it. Automation is the last step, never the first.
Jeremy spent two decades building Finder into a business doing more than $100M in revenue at a $650M valuation, bootstrapped, well before any of this AI tooling existed. The business was built the slow way, by hand, learning each process before scaling it. The lesson was simple and expensive to learn. Systems built on broken foundations do not survive growth. They crack the moment the load goes up, and they crack at the worst possible time. None of that was built with AI. It was built by understanding the work first.
AI capability should be taught, not rented back to you by the seat.
There is a whole model emerging where someone wraps a thin layer around Claude or an MCP server, charges you per agent per month, and quietly makes your business dependent on them forever. That is an AI tax. It is rent on capability you could own, dressed up as a product.
We would rather hand you the skill. Show your team how the tools work, build the things that need building with you in the room, and leave you able to run and extend them with or without us. A capable team beats a vendor lock-in every time. If the work we do makes you need us less, that is the work done properly.
You do not gamble the business on automation that has never been proven. You do not inherit a black box you cannot explain to your own board. The processes that get automated are the ones that already worked by hand, measured against a baseline you watched us establish.
The result is an edge that compounds instead of a liability that grows. Strong foundations first, automation on top, capability that stays in your team when we leave. That is the whole argument. It is also the only way we have seen it hold up under real growth.
See where automation pays off, and where it would just lock in the mess.
A focused 1:1 audit of your business to find the processes worth proving by hand, and the ones worth automating once they do.