Most enterprise AI conversations start at the wrong end. They start with generation — draft this document, summarise that thread, produce a specification — as if the hard part of enterprise delivery were writing things down. It is not. The hard part is knowing what is worth writing down in the first place, and being able to defend that judgement to the people who will fund it. That is why we built Fungy.ai as a discovery layer rather than another generator.
The layer nobody owns
Demand in a large organisation almost always arrives as a sentence. A line in an email, a comment in a steering meeting, a request forwarded from a business unit with the words “can we look at this?” attached. Between that sentence and something a delivery team can responsibly act on sits a stretch of work with no clear owner: clarifying what is actually being asked, finding the stakeholders who hold the missing context, sizing what it would take, and turning all of it into requirements someone can review. Everyone touches this stretch. The PMO chases it. Business analysts rebuild it through follow-up calls. IT inherits whatever survives. But no single tool is built for it, so it happens in inboxes and spreadsheets, and it takes weeks.
When I look at where enterprise projects go wrong, the failures rarely originate in the code. They originate here, upstream, in the part of the lifecycle that everyone treats as pre-work rather than as work. So that is where we pointed the engineering.
Why a discovery layer, not a chatbot
A chatbot answers the question you happen to ask. A discovery layer is opinionated about the questions worth asking. That distinction drove almost every design decision we made.
Fungy is built around the shape of pre-delivery discovery, not around a single prompt box. It guides a requester from a vague pain to a structured need, then it enriches that need with the stakeholder questions the requester did not think to ask. It sizes the initiative along value, effort and risk so that two competing requests can actually be compared. Only then does it generate a requirements package — business, functional and non-functional — for a human to review and hand off. The point is that it strengthens every step of discovery, not only the first interview. The first interview is the easy part; the twelfth follow-up, three weeks later, when the original context has evaporated, is where the value leaks out.
A generator gives you a document. A discovery layer gives you a decision you can stand behind.
— Dr. Michalis Loutsaris
Evidence is the output, not text
The single most important framing for us is that the output of discovery is not a document — it is a review-ready evidence base. Text is a by-product. What a decision-maker actually needs is a common basis for judgement: what is being asked, why it matters strategically, which KPI it is meant to move, who the stakeholders are, what it would take to deliver, and where it is exposed. When all of that is assembled in one place and structured consistently, prioritisation stops being a debate about who argues most persuasively and becomes a comparison of like with like.
Engineering to that goal changes what “done” means. A generated specification that reads well but rests on unexamined assumptions is not done. An evidence base that surfaces what is ready, what is missing, and what needs intervention is done — even when the honest answer is “this initiative is not ready to fund yet.” Making the gaps visible is a feature, not a failure.
What we deliberately did not build
It would have been tempting to let the system make the call. It does not, by design. Fungy improves the quality of the inputs to a decision; it does not make the governance decision itself. That guardrail is deliberate and it is load-bearing. Enterprises do not need software that quietly decides what gets funded — they need software that makes the human decision faster, better-evidenced and easier to defend afterwards. The moment a tool starts substituting its own judgement for the organisation’s, it stops being trustworthy in exactly the settings where trust matters most.
The same instinct shapes how the AI itself is run. Standardised prompts, monitoring and output validation are not bureaucratic overhead bolted on at the end; they are what let a non-deterministic model sit inside a governed enterprise process at all. Innovation with professional review and control is not a slogan for us — it is the only version of this product an IT organisation should be willing to put upstream of its delivery pipeline.
Building Fungy as a discovery layer was, in the end, a bet about where enterprise AI creates durable value. Not in producing more text faster, but in turning an unclear idea into one review-ready evidence base — so that by the time a request reaches delivery, the hard thinking has already been done, and it can be trusted.
Key takeaways
- Enterprise projects tend to fail upstream, in the ownerless stretch between a request and something delivery can act on.
- A discovery layer is opinionated about the questions worth asking; a generator only answers the one you typed.
- The real output is a review-ready evidence base — and surfacing gaps is a feature, not a failure.
- The AI improves the inputs to a decision; humans still own the decision. That guardrail is the point.