Every week another founder asks us the same question. Some flavor of: should we use AI for this? Most of the time the honest answer is "for a small part of it, yes, and for the part you think, no."
This is not a rant against AI. We have been doing language tech since before ChatGPT existed, and we still ship it to clients every quarter. It is a rant against pretending the technology is something it is not, which makes it harder to use the parts that genuinely work.
What AI is actually good at right now
The interesting things AI does for businesses today are unglamorous.
It reads long documents and summarizes them accurately enough that someone can decide whether to read the whole thing. It drafts the first version of structured text — a product description, an email reply, a brief — that a human then corrects in a quarter of the time. It classifies messy input. It transcribes calls. It extracts data from invoices in fifteen different layouts. It powers search interfaces that finally understand what a person is asking for.
These are not headline-grabbing use cases. They are unflashy, well-bounded tasks where being eighty-five percent right with a human checking is materially better than being zero percent right with no human at all.
A small change to one of these workflows can shave hours off the day for an entire support or operations team. The cumulative effect across a year is real money and real time. It is just not a magazine cover.
What AI is bad at right now
It is bad at being a single source of truth for anything that cannot be wrong. Legal advice. Compliance answers. Final numbers in a financial report. Anything where a confident-sounding lie ships to a customer.
It is bad at replacing judgment in genuinely ambiguous situations. The questions where the answer depends on knowing your business and three years of context cannot be patched by a longer prompt.
It is bad at staying the same. Models drift. Prompts that worked last month behave differently this month. The infrastructure you set up needs maintenance the same way any production system does. Vendors who hand you a chatbot and walk away are setting you up to find this out the hard way.
The vendor smell test
We get asked to review AI proposals from other vendors fairly often. There are a few patterns that show up in the bad ones.
The pitch is full of acronyms and short on outcomes. If a proposal cannot explain in one sentence what a specific person will do differently after the system ships, the proposal does not know what it is for.
The cost is impossible to verify. Token usage, model fees, hosting, retraining — these add up in ways that are easy to under-disclose at signing. Ask for a worst-case month and a typical month, in numbers, on paper.
The vendor wants to own the integration but not the result. If they will not commit to a quality bar that you can measure, you are paying for a science project.
The architecture diagram has fifteen boxes for a problem that needs three. Complexity in this space is usually a sign that the person designing it has not built it before.
A short list of things to skip
Some specific things that are mostly noise right now:
- Building a chatbot when a better-organized FAQ would do the job.
- Wrapping a public model in a thin layer and calling it your AI product.
- Fine-tuning on a tiny dataset hoping for a behavior shift that needs a different model entirely.
- Replacing a working rules-based system with an LLM "because it is the future."
Most of the time the right answer is a small, well-targeted model that does one thing, integrated into a workflow people already use, with a clear way to measure whether it works.
What we actually do
We pick the model that fits the job and we are honest when no AI is the answer. We have fine-tuned models for clients on dataset sizes from a few hundred examples to millions of them. We build the interfaces people actually touch, not just the inference layer. We host on our own hardware when data sovereignty matters, and on cloud GPUs when it does not, and we are clear about the trade-offs of each.
We do not push acronyms that solve problems you do not have. The technology is interesting enough on its own. The work is figuring out which of its capabilities map to your specific bottleneck, and then building the thing carefully.
What to do next
If you are thinking about AI for your business, the first useful step is rarely "find a vendor." It is two hours with someone who has shipped this kind of thing, going through your actual operations, and ending up with a short list of the places where a model could remove real friction.
If three of those places hold up to scrutiny, build one. Watch it run for a quarter. Then decide what to do with the other two.
The companies getting real value out of AI right now are not the ones with the loudest announcements. They are the ones that picked one specific problem and shipped a careful solution to it. Boring as it sounds, that is where the money is.