Most business processes are shaped around what a person can follow. Open-weight models are cheap enough to test a better approach.
Most business processes have more steps than the work needs. They have those extra steps because a person has to follow them, and following means understanding what came before and what comes next. So the work gets cut into pieces small enough for someone to hold in their head, with a handover at every step. That was the right design when people did all of it. Whether it still is has become a live question, and the answer changes what AI is for in your business. The gain sits in how much of the process only ever existed to keep a person oriented.
Take a 30-person building supplies wholesaler. Turning a customer enquiry into a priced quote moved through five hands, and each handover existed so the next person could see what the last one had done. Start to finish it took about three days. They put AI on it, every step got quicker, but quotes still took two days, because there were still five steps.
Why throwing tokens at the problem still worked
The standard move for a while was to give the model as much room as it wanted. Long prompts, big context, let it think as hard as it likes. People call it token maxing, and it caught on because it eventually got results, even though most of what it spent was waste. The waste is measurable. Opslyft analysed 84 production AWS Bedrock deployments in early 2026 and found the median bill came in at 2.8 times what the engineering team had forecast at kickoff. Once those teams put basic discipline in place, sending simple calls to smaller models, caching repeated prompts, and sizing capacity to real demand, the median cost per answer fell from 41 cents to 7 cents. That is an 83% reduction without quality loss. The answer was reachable the whole time. It just didn’t need most of the money.
Money isn’t the only thing that gets worse when a model has unlimited room. Small decisions the AI makes really snowball, the model settles on an assumption early, treats it as correct, and builds everything after it on that footing. Best case, you catch it and rebuild. Worst case, the assumption ends up inside a shipped product or a decision the business has already acted on.
The steps that only exist for people
Here’s the part most businesses haven’t sat with yet. Your team needs to understand the inputs, and they need to understand the result. They do not need to understand every step in between. That sounds like a small change. It isn’t. Walk into almost any business and you’ll find some roundabout way of doing a job, and the reason is nearly always that someone had to be able to follow it. Once that requirement lifts, the shape of the work is open to question for the first time in years. That is what process optimisation is really about now, and it’s a different exercise from automating the steps you already have.
So the wholesaler stopped asking AI to do their five steps. They wrote down what actually happens, then asked the model a different question. Same result, same accuracy, nothing skipped and no corners cut, is there a better way to get there? What came back wasn’t five faster steps. It was a different shape, with the pricing and stock checks running alongside each other instead of in sequence, which is the sort of coordination AI agents handle well, and one person reading the finished quote rather than each stage of it.
There’s a limit worth identifying. The model has no sense of which of your steps are load bearing. AI really doesn’t have the intuition to know which assumptions are good or bad. A check that exists because of a bad debt in 2019, or because the Privacy Act 1988 obliges you to handle customer information a certain way, looks exactly like padding from the outside. Someone who knows the business has to sit with the redesign and say which steps stay. That isn’t a sign-off at the end, either. The point is that review has to stay live through the work, because assumptions get more expensive to reverse the longer they sit.
Testing is cheap enough to do properly now
None of this works if every experiment costs real money. That is what has changed significantly over the last couple of weeks. Frontier level open-weight models like Kimi K3 and GLM 5.2 arrived in the middle of 2026 and brought serious capability within reach of a normal budget. On Artificial Analysis’s independent measurement, Kimi K3 scores 57 on their Intelligence Index against GLM 5.2’s 51, which puts K3 fourth overall, behind Claude Fable 5 and GPT-5.6 Sol and roughly level with Claude Opus 4.8. GLM 5.2 costs about half as much per completed task, 47 cents against 94 cents, and runs around five times faster.
Numbers like that let you run 40 versions of a process in a week instead of arguing about two of them in a meeting. That is the real gain, and it’s easy to miss. We heard a number of people wave Kimi K3 away recently because its API price sit around the same point as one of Anthropic’s models, as though price parity settled it. It doesn’t. A per-token rate tells you very little on its own, because a model that reasons better can finish the same job in fewer calls. What matters is the cost of a finished piece of work and how good that work is, which is exactly why Artificial Analysis reports cost per task rather than cost per token.
Be honest about the trade-offs. K3 answers more questions correctly than GLM 5.2 but makes up more answers when it doesn’t know something, scoring 49% against GLM’s 72% on Artificial Analysis’s non-hallucination measure. GLM 5.2 ships under a plain MIT licence, while Kimi K3’s adds commercial conditions above certain revenue and user thresholds. There’s no shortcut on the choice either, so run your own test on a 50 to 100 prompt slice of your real work before committing.
What this means for you
The wholesaler landed on same-day quotes, with a person still approving price and margin before anything reaches a customer. What got them there wasn’t a better model. It was being able to test enough versions of the work to find a genuinely better one, and being willing to question steps nobody had questioned in years.
Start with one process you know is slower than it should be, and write down what actually happens rather than what the procedure says. Then ask what it would look like if nobody in the middle had to understand it. On savings, take Opslyft’s advice to its own customers and let the percentage be an output rather than a promise. You’ll know after the first month of measuring. If you want help working out which process to start with, take our AI Roadmap Interview. You’ll talk through where your team loses the most time, and get a plan built around your business, not a generic one.
