Every so often, a business owner in a meeting room asks the same question: so which one should we actually be using?
It isn't a wrong question. It just has an expiry date. Whichever vendor you settle on this year will most likely be replaced next year. That market reshuffles every few months anyway: prices drop, features get matched, and things that used to cost extra end up bundled into the entry-level plan a while later. That's good news, but it also makes one thing awkward. Plenty of companies bet their entire year's AI budget on a decision that is bound to expire.
The earlier pieces in this September series were all about operations: how to sign off on results, how to keep a system from rotting, where memory lives, who signs when something goes wrong. This piece pulls them together with a question that sounds more like a procurement meeting: of the money spent this year, which part was rented, and which part stayed?
01Split your AI spending into two columns
Take a sheet of paper, list every AI-related expense from this year, then split them into two columns.
The left column is what disappears when you stop paying. Subscription fees, seat counts, a platform's monthly rent. The day after you stop paying, that capability is gone from the company without a trace.
The right column is what stays when you stop paying. The process map it took you three weeks of arguing to settle; the fourteen rules sales wrote down from twenty years of quoting experience; the materials and talking points marketing organized and versioned; and that checklist defining what "done" means. Swap out the tool, and all of it is still on the table.
Most companies, once they've made the list, notice one thing: the left column is full, and the right column is nearly empty.
That's exactly the problem. Your competitors can buy anything in the left column, and it may be cheaper next month. No one can buy what's in the right column, because it was never on a shelf to begin with. It can only grow out of your own processes.
There's a harsher version you can use as a self-check. One analysis of AI-native companies describes a thought experiment called the "thirty days without AI" test: suppose that starting tomorrow, no one is allowed to use any AI for a full month. What happens to the company? If the answer is "everyone writes emails a bit slower and making slides gets more tedious, but production and delivery carry on as usual," then the company has only installed AI on its toolbar. It hasn't yet built it into the way the company produces.
The same analysis also points out the flip side. A field experiment covering more than five hundred high-growth startups found that simply giving companies more access to AI doesn't automatically turn into value at the company level; the researchers called the step where it gets stuck "not knowing where to put it." The intervention that moved the numbers was showing these companies which part of the business others had connected AI to. Companies that received it found 44% more use cases and were nearly 20% more likely to land paying customers.
Access is rented. Knowing where to put it isn't.
02Three things that stay in the company
So what does the right column actually look like? Put plainly, it's three things.
One: how the process is cut.
This is the step most easily skipped, because it costs no money, only people's time. Many companies do it in the wrong order: pick the tool first, then look for use cases, and only at the end discover the process was never sorted out. One source puts the order bluntly: sort out the business scenarios first, then standardize the process, then build up the data. Get the order backwards, and buying tools is just burning budget.
How you cut it deserves even more thought. The AI-native discussion includes a counterintuitive claim: a large share of a company's existing processes were originally there to make up for four human limitations, namely that people are expensive, people are slow, people forget, and passing information between departments is costly. Once those limits loosen, the right move is to delete a section of the old process outright, not hand it to AI to run unchanged.
That source offers a useful test: wherever the work is mostly shuttling information around, syncing context, or making up for gaps in people's memory, those are the sections to recut.
And once "how to cut it" has been thought through and written down, it no longer has anything to do with whose model you use.
Two: how the data is fed.
This has been talked to death, but most people focus on the wrong thing. Companies usually aren't short on data; there's plenty of it in the ERP, the CRM, the shared drive. What they're short on is judgment.
One source draws the distinction neatly: data records outcomes, while what makes AI answer accurately is business context, meaning the reasoning behind the rules, the limits of experience, and where a process currently stands. A shipping record can tell AI that this order took seven days, but it won't tell it that "for this kind of customer, once it goes past five days, you call them proactively." That sentence currently lives in some senior employee's head.
Getting it out of their head is something some people call knowledge distillation. Research on transformation across organizations lists it as one of four pillars, and names it as the place where the most companies die at step one: if knowledge hasn't been distilled into raw material AI can read, everything after that is empty talk. Another review of more than thirty company case studies is blunter: 90% of projects die at one of three breaking points, which are ungoverned data, chasing too many scenarios at once, and an organization that didn't keep up.
Not one of those three breaking points can be fixed by switching vendors.
Three: how results are signed off.
Earlier pieces covered writing acceptance criteria down before rollout, and who signs for the answers. Put on this procurement sheet, it takes on another meaning: acceptance criteria are one of the few company assets that "still work after you switch tools."
"Customer service must reach a 70% first-reply hit rate; overdue cases must be picked up by someone the same day; every external quote must be confirmed by the sales manager." Those three sentences apply today, and they'll still apply after you switch tools. They aren't tied to any interface, yet they decide whether you have any grounds to say "this works."
There's a test from the transformation research worth copying down: can AI reliably take over more than half of the execution work? Another source on AI-native companies defines the gate at each level as three things holding true at once: the business value is real, the risk is controllable, and accountability is traceable. The advantage of sentences like these is that you don't need to understand any technology to take them to a vendor and ask.
03The cost of mixing them up: every switch is a restart
Put the two columns side by side and the problem becomes clear.
A company that mixes them up looks like this: the process exists only in one colleague's working habits; the business rules are written in prompts inside a chat box; the acceptance criteria are "everyone thinks it's pretty good." All three are left sitting in the vendor's interface, or in one person's window.
Then one day, that vendor raises its prices, gets acquired, or simply gets replaced by something better. You switch. The next six months are as good as wasted: the prompts need retuning, the rules have to be drawn out of sales all over again, and what counts as "done" has to be argued out one more time.
This isn't hypothetical. One analysis goes further: rather than renting all of their intelligence inside someone else's black box, companies should first work out what to rent and what to own. The author condenses it into a line adapted from a crypto-world saying: not your weights, not your product.
That source is talking about building your own models, at a scale far beyond the typical small or mid-sized business. But the "rent vs. own" dividing line itself can be scaled down to any size. For a thirty-person company, what you should own is those three things above: how the process is cut, how the data is fed, how results are signed off. Model weights aren't on the list.
How much would you have to redo if you switched vendors? The less you'd redo, the more this year built up. If you'd have to start almost entirely over, what you bought this year was usage, not capability.
04So should you choose at all?
Someone will ask: so it doesn't matter which tool we pick?
No. Of course you choose a tool. That decision just needs to be put back at its proper weight: a procurement line item signed once a year, reversible, renegotiated next year. The strategy lives in the right column.
In practice, two things are worth doing deliberately.
First, decouple what the right column produces from the tool. Process maps, rules, acceptance checklists: write them in the company's own files, saved in the most ordinary formats. If anything important can only be seen inside some platform's back end, it already doesn't fully belong to you.
Second, treat the trial period as a chance to produce the right column, not to compare features. Most trial periods get spent comparing whose interface looks nicer, and that's a waste. What those thirty days should produce is first drafts of the three things; which tool to pick is the last decision you need to make.
One analysis raises a question for gauging what a company is made of, and it fits well here: when the business grows, is the first thing you add long-term headcount, or process and automation capability? There's no standard answer, but it's really asking the same thing: what are you accumulating?
So what are you buying?
Before A company's capabilities were written in the org chart and the equipment list
Now There are three more things that aren't on the balance sheet: a process that's been cut, a body of judgment that can be fed in, and a set of acceptance criteria that still hold after you switch tools
They don't show up on the books, because nobody ever invoiced for them. But on the day you switch vendors, these three are the only things in the whole company that move over intact.
Back to that question in the meeting room. "Which one should we use" doesn't need a whole afternoon; it's worth about twenty minutes. Spend the rest of the time on three other questions: How should this be cut? Who writes the judgment down? What counts as done?
For companies that can answer them, switching tools is just switching tools. For companies that can't, every switch means starting over.
Let AI move fast. Let people hold the anchor. Of the money you spend this year, at least one column should be the kind that stays.
Rooted in craft. Built for the new wild.