What can AI do for my business? Start by listing what repeats.

On October 7, Anthropic released Claude Haiku 5.5, a small, fast AI model. Anthropic says it now costs about 75 percent less to run, on average, and it is aimed at high-volume work like summarizing, sorting and classifying (Anthropic announcement). That is Anthropic's own figure, and it says the savings vary with request size.

I do not think the model matters much to you. What matters is the direction: routine AI work keeps getting cheaper. Cost is no longer a good reason to wait. The better reason to wait is that most owners cannot say what they would point it at.

So when someone asks me "what can AI do for my business," my answer is always the same. It can do the work that repeats. The bottleneck is not the technology. It is that nobody has written down what repeats.

A common bottleneck: a vague goal with no target

"We should use AI" is not a plan. It sends people shopping for tools, and tools without a target get used twice and forgotten. A cheaper model does not fix that. It just makes it cheaper to be unfocused.

How to spot what repeats

Give yourself ten minutes and a piece of paper. Walk through last week and write down every task that felt familiar. Then keep only the ones that pass four checks:

It happens weekly or more. A once-a-year task is not worth automating.
The steps are mostly the same each time. The details change; the shape does not.
It costs you something. Time from people who should be doing other work, late replies, missed follow-ups or mistakes.
You can say where it starts and stops. A task with an edge can be handed off. A fog cannot.

Common items on the list: chasing follow-ups, retyping the same data into a second tool, answering the same customer question, sorting incoming email or requests, and pulling numbers together for a weekly report. Several of those are exactly the sorting and summarizing jobs that small, cheap models are built for.

How it gets automated, and stays fixed

Pick the one item that costs you the most. Then:

Write one paragraph about it. What triggers it, what happens, who is involved, what done looks like.
Check what you already pay for. Many tools you own can already do part of it. Use them first.
Build only the piece that is yours. If the process is particular to your business, a small automation or piece of software closes the gap.
Give it an owner and make failures visible. A fix that fails quietly is worse than no fix.
Keep your data portable. Models and vendors change fast. Your setup should be easy to move.

Notice what is missing: a model choice. That comes last, and it is the easiest part to swap. The paragraph you wrote is the part that lasts.

Why I do it this way

I have built and run my own products, including a school fundraising platform and an AI business knowledge tool, so I know the difference between a demo and something that holds up on a normal Tuesday. Every project gets a fixed quote before any work starts. And I reply to every message myself.

What "fixed for the long term" means

It means the fix does not depend on a specific model, a specific vendor or you remembering to do it. When the next cheaper model arrives, you change one setting, not your whole process. That is the difference between a bottleneck fixed once and a tool you keep re-buying.

What I would do on Monday

Do the list. Five tasks, cross out anything less than weekly, pick the one that costs you most, write the paragraph. If it still looks like a good candidate after a day, it is.

Show me yours

Bring the list. On the free call you describe your week, I look for the repeat work, and you leave with one place to start and what fixing it for the long term would involve. No obligation, no jargon.

Show me yours. Text (512) 337-9530, email josh@joshdlange.com, or book the free AI assessment.

Source: Anthropic, Claude Haiku 5.5, October 7, 2026.