6. Where Enterprise AI Creates Value
The last five pieces have been about how these systems work. This one is about where they pay, and it is the piece that matters most if you are deciding what to do on Monday.
The prevailing assumption is that AI value is spread fairly evenly, waiting to be picked up by whichever function adopts a tool first. Our experience says the opposite. The returns pool in particular places, and those places share a structure. Where the structure is present, the compression available is large. Where it is absent, you get a modest improvement and a subscription.
The three conditions
The first is that decisions are assembled from documents and tacit judgement together. Neither a database nor a model on its own can reproduce them, which is precisely why two decades of enterprise software left this work untouched. The knowledge required lives partly in files and partly in the heads of people who have never written it down.
The second is that senior bandwidth is the constraint. The cycle moves only as fast as a few experienced people can attend to it. Where a hundred juniors could do the work, automation saves cost. Where three seniors are the throttle on every quotation, evaluation, or filing, removing the wait changes the speed of the whole business.
The third is that knowledge loss or a missed obligation is genuinely expensive. The exposure may be regulatory, contractual, or the quiet attrition of the people who carry the operation in their heads. Where the stakes are low, an organisation will tolerate the status quo indefinitely and pay very little to change it.
One condition on its own gives you a useful tool. All three together, and the arithmetic changes entirely. We set this out at more length in our piece on where the compression actually is, including why manufacturing, financial services, eCommerce, and sustainability reporting meet all three so plainly.
How to find it in your own operation
The practical version of the test is a question you can ask in a corridor. Where does work sit and wait for a small number of experienced people, and what does it cost you while it waits.
The answers tend to be unglamorous. A quotation held up because the person who knows the substitution rules is in a meeting. A compliance renewal that nobody noticed until a customer's audit. An evaluation that starts from a blank sheet because the last four evaluations were never written down. This is where the value is, and it is almost never where the excitement is.
Be suspicious of the workflows that look most impressive in a demonstration. They are usually the ones with the cleanest data and the lowest stakes, which is why they demonstrate well and why they change so little.
What you can safely ignore
You will be told a great many things matter that do not matter to you.
The hardware conversation is one. GPUs, TPUs, inference optimisation, quantisation, throughput. All of it is real, all of it is genuinely important to the people building the infrastructure, and none of it should influence a decision you take about a workflow in your plant. It affects the price of what you buy, and the price is quoted to you.
Model names are another, for the reasons set out in the second piece of this series. So is the volume of parameters, the size of the context window, and most of what appears on a comparison chart. These are inputs to an engineering decision that your vendor should be making on your behalf and should be prepared to justify.
The questions that deserve your attention are the ones about mandate, adoption, measurement, and ownership. They are less interesting and they determine everything.
What we are still working out
We do not know how durable the three conditions are as a filter. They describe what we have seen, and what we have seen is a particular slice of manufacturing, financial services, eCommerce, and reporting in India and Europe. It is entirely possible that other patterns exist which we have not encountered, and that our confidence in the filter reflects the limits of our sample rather than the shape of the world.
We are also alert to the possibility that the conditions loosen. Every argument in this series about why point solutions have failed in these domains rests on the difficulty of combining documents with tacit judgement. If that combination gets substantially easier, the constraint moves, and so should our thinking. We would rather notice that early than defend a framework past its usefulness.
If you disagree with any of this, we would be glad to hear it. The series is offered as our current understanding rather than as a settled account.
—
Mitochondria is an agentic AI product company based in Amsterdam and Pune. ISO 27001:2022 certified. Classified within the limited and minimal risk tiers of the EU AI Act, with controls aligned to the GDPR, UK GDPR and India's DPDP Act.