There are two ways to build an AI product. The first is to start with what AI can do and find a commercial problem that fits. The second is to start with what a commercial decision actually requires and ask where AI can genuinely help with that.
Most products being sold into estate agency right now were built the first way. A language model exists that can read text and produce outputs. Seller enquiries are text. Therefore: AI seller qualification. The logic is superficially coherent. The product ships. It does something that looks like what was promised.
But the starting point shapes everything that follows. A system built around what AI can do will optimise for AI performance. A system built around what the decision requires will optimise for decision quality. These are not the same objective, and they do not produce the same system.
I want to be honest about why this distinction matters to us, because it is not an abstract engineering preference. It is the reason Pulsero is built the way it is, and it is the reason the alternative approach produces systems that underperform in commercial settings even when the underlying AI is capable.
What the decision requires
The decision an estate agency makes when a seller enquiry arrives is not a simple one. On the surface it looks like a routing problem: which sellers should we contact first, and in what way. But the decision that needs to support that routing is considerably more demanding.
It requires an accurate reading of what the seller has communicated, including things they have said indirectly or incompletely. It requires placing that reading in the context of what actually predicts instruction likelihood, which is not the same as what sounds most urgent or most motivated in the text. It requires consistent application of that contextual logic across every enquiry that arrives, not just the ones that are easy to read. And it requires producing an output that the agent receiving it can evaluate, act on, and push back on when their own knowledge of the situation suggests a different reading.
That is a demanding set of requirements. Reading the enquiry accurately is one part of it. But accurate reading is the beginning of the process, not the end. What follows ”” the contextual judgement, the consistent application of logic, the production of auditable reasoning ”” those are different requirements, and they are not naturally satisfied by the same system that handles the reading.
Where AI genuinely helps
AI is genuinely good at the reading part. This is not a minor capability. Seller enquiries arrive in every form imaginable ”” terse portal messages, discursive emails, walk-in conversations reconstructed from memory, forwarded WhatsApp threads. The range of expression, the variation in what is stated directly versus implied, the informal language and incomplete sentences that characterise real human communication: these are exactly the conditions under which large language models outperform every previous approach to text understanding.
A capable language model reading a seller enquiry can surface context that a human would identify but that no keyword system or structured form could capture. It handles the messiness of real communication without needing that communication to conform to a particular format. That is valuable, and it is genuinely difficult to replicate without AI.
The question is what you do with that capability. If you ask AI to read the enquiry and also produce the final classification, you are using a single system for two tasks that have different requirements and fail in different ways. The reading capability is real. But the classification decision needs things that a language model does not natively provide: stable logic that does not drift between enquiries, calibrated uncertainty rather than uniform confidence, and reasoning that can be examined and corrected when it is wrong.
Where it does not
The failure mode that results from asking AI to do too much is not dramatic. The system does not obviously break. It produces outputs that look reasonable, expressed with apparent confidence, at a speed no human process could match. The problem is in what those outputs are actually worth.
A language model making classification decisions will produce different outputs on similar enquiries as the model is updated. It will express the same confidence on a well-evidenced classification as on a poorly-evidenced one. It will produce reasoning that sounds coherent but may not reflect the actual basis for the output. And when it is wrong, the error will often not be visible until its consequences have already arrived.
These are not problems that can be solved by making the model more capable or the prompt more specific. They are structural characteristics of using a probabilistic system for a task that requires deterministic properties. No amount of fine-tuning changes the fact that a language model is optimised for producing plausible outputs, not for applying a defined logic consistently and transparently across every case it encounters.
The agents using these systems sense this even when they cannot articulate it. They learn not to fully trust the output because the output behaves in ways they cannot predict or explain. They use the tool less. The capability that was purchased goes partially to waste because the system was not built with the requirements of the decision in mind.
The question that matters
The most useful question when evaluating any AI system making commercial recommendations is not how capable the underlying model is. It is what the system is asking AI to do, and whether that matches what AI is actually good at.
A system that uses AI for everything it can plausibly handle has optimised for the presence of AI rather than the quality of the output. Those are not the same objective. Building from the decision outward — asking what each part of the process requires, then matching the right mechanism to each requirement — produces a harder system to build and a more reliable one to use.
What this means in practice
Every output Pulsero produces shows what the enquiry said, what it suggested, and why the system reached the conclusion it did. The agent can agree with it, push back on it, or override it with information the system did not have. It is an informed position to be evaluated, not a verdict to be accepted.
That is only possible because the reading capability and the classification logic are handled separately, by parts of the system suited to each. The technology exists to collapse those into one step. The question is always whether doing so serves the decision or just serves the technology. For classification with real commercial consequences, the answer is consistent: keep them separate, make the reasoning visible, build something people trust rather than tolerate.