Narrowing a very broad AI product down to one real customer
Ravon Regulated AI Product
Outcome
Six industries evaluated against four questions, and legal chosen as the first market to prove the model in.
Ravon could build almost anything. What it did not have was a reason for one specific customer to buy one specific thing. I ran the discovery, compared what came back across sectors, and used the evidence to narrow the customer, the roadmap and the requirements all at once.
- Company
- Ravon Group
- Product
- On premise enterprise AI system for regulated organisations
- Role
- Product and Business Development Coordinator
- Timeline
- January 2025 to October 2025
- Status
- Legal chosen as the first market
30
discovery conversations
6
sectors compared
3
product areas prioritised
6
industries evaluated
Team and ownership
Said plainly, because it is what makes the rest believable.
I owned
- The discovery programme and what came out of it
- Ideal customer profile and qualification criteria
- Choosing the three product areas
- Requirements for citations, access control, audit and approval
- Four stage journey and the reusable CRM, proposal and content workflows
I supported
- Commercial briefs used by sales and technical teams
- Positioning against cloud platforms and internal build attempts
- Sector priority inside the target profile
I did not own
- Technical architecture and how it gets deployed
- Model selection and evaluation
- Pricing and packaging
- The decision to enter through legal first
The problem
The company had a lot of capability and no shortage of ideas. That sounds like a nice position. It is actually harder than being constrained, because every opportunity looks equally good and the roadmap turns into a list of features.
Conversations kept ending politely. People agreed the technology was clever, then did nothing, because nothing on offer pointed at a problem they were already paying to solve.
There was a bigger issue underneath. The companies most likely to need this could not use the obvious version of AI at all. Sending client files to an outside service was either banned outright or far too risky to consider.
What made it hard
- Our main competitor was doing nothing. Most people were not choosing between us and a rival system. They were choosing between us and carrying on exactly as they were.
- Three people sit in every buying conversation and want different things. An owner or managing partner signs. An IT lead can kill it. A compliance lead has to be satisfied, not sold to.
- Claims had to stay modest. When the whole pitch is that you can be trusted, one oversold feature costs more than it wins.
- I was running discovery alongside a commercial job, so the research had to be useful to sales straight away, not just feed a roadmap later.
My job
- Run the discovery and turn it into ranked requirements, use cases and briefs that sales and technical teams could pick up and use.
- Use the evidence to say who the priority customer actually is, and give the team criteria they could apply without me in the room.
- Pick the product areas, so roadmap conversations started from customer problems instead of available technology.
- Turn the objections we kept hearing into specific product requirements.
- I did not own the architecture, the deployment model or pricing. Entering through law firms was company strategy. My discovery backed it up, it did not start it.
What I found out
- Around thirty conversations across six sectors. Four of them, legal, financial services, insurance and healthcare, kept hitting the same wall and became the target.
- The pattern was not company size or budget. It was whether confidentiality was a legal or professional duty rather than a preference. Where it was a duty, the conversation changed completely.
- Under about twenty people, nobody felt the problem. You just ask across the room. Over about five hundred, the buying process took longer than we could survive.
- Three objections came back almost word for word. How do I know it is right? Who can see what? What happens when it gets something wrong?
- Curiosity and real intent look identical in a first meeting. The only reliable tell was whether they had already written an internal policy banning public AI tools. Those people had done the thinking and wanted a way through it.
What I decided
- 01
Pick one customer: regulated companies with 20 to 500 staff
Big enough that knowledge is genuinely scattered and it hurts. Small enough that a few people can decide without an eighteen month procurement process. This gave the team something they could check in a first call, and made it completely fine to say no.
- 02
Three product areas, not a platform
Knowledge chat, document generation and supervised agents. Each one mapped to work these companies already did and already paid for. Naming three moved roadmap discussions off generic AI features and onto real, repeatable problems.
- 03
Turn objections into features
Rather than answering worries about accuracy, security and accountability in meetings, I wrote them up as product. Citations back to the source, role based access, audit trails, and a person approving anything that matters. An objection you can demo away stops being an objection.
- 04
Write down who is not a fit
Companies with no regulatory pressure would compare us to a free tool on price, and we lose that comparison every time. People who wanted a pilot were a cheaper, different purchase. Saying this out loud saved more time than any qualification framework.
- 05
Four stages, each with a rule for moving on
Discovery, qualification, proposal, follow up. Each stage had a condition you had to meet before advancing, so a deal could not creep forward just because a meeting felt good. The CRM, proposal and content workflows were reusable, so the team kept using them after I handed over.
The thinking, not the screenshots.
These are the structures the work produced. That is the part that carries over to the next product.
Artifact · Industry evaluation
Six industries, four questions, one first market
The discovery fed a decision about where to start. These are the questions it was judged against, and how each industry came out.
The four questions
- 01How quickly is the industry adopting AI?
- 02How operationally prepared is it?
- 03Is there a defined budget?
- 04How long does interest take to become a live implementation?
Legal
ChosenAlready understands the value of AI, so the category does not need explaining. Document heavy, sensitive, and built around connected workflows, deadlines and approvals.
Manufacturing
NextReal need and the strongest opportunity outside legal, but much of the market is not AI ready yet. Entering now would mean educating customers first.
Finance
LaterSignificant demand and budget, but heavy regulation and real operational risk. Trust would need certifications and long term evidence first.
Healthcare
LaterSame shape as finance. Strong need, but the safety conversation is a separate sale and procurement moves slowly.
Logistics
Ruled outInterest was there, but the reachable market was smaller and the larger operators were already building internally.
Large listed enterprises
Ruled outAttractive customers, but procurement, security review and proof of concept cycles take too long to validate a new platform against.
Legal is not a permanent limit. It is the most credible and commercially viable market to prove the wider model in, then expand from a position of evidence.
Artifact · Prioritised roadmap
Three product areas, each tied to work already being paid for
Naming three moved roadmap discussions from generic AI capability to repeatable customer problems.
Enterprise knowledge chat
Ask a question in normal words. Get one answer back, with a link to the exact document and page each part came from.
Replaces
Searching four systems, or asking the one person who remembers
Document generation
Reports, contracts and proposals built from the company's own material, in its own format and language.
Replaces
Rebuilding a document that has basically been written before
Human-supervised agents
Routine work runs on a schedule. It watches first, and anything that matters waits for a person to approve it.
Replaces
Chasing deadlines and checking numbers by hand
Artifact · Objection to requirement
Four concerns, specified as product
A concern you can demonstrate away stops being an objection. One you argue about stays one.
“How do I know the answer is right?”
Source citations
Every part of an answer points at the document it came from
“Who can see what?”
Role-based access
Admins set what each person can reach. Nothing steps outside that.
“Can we prove what happened?”
Audit trails
A permanent record of everything the system did, so you can reconstruct it later
“What if it acts wrongly?”
Human approval
Anything that matters stops and waits for a named person to say yes
What changed
- Product conversations started from a real customer problem instead of a capability we happened to have.
- Qualification became something anyone on the team could do, which improved the pipeline far more than adding volume would have.
- Trust worries stopped being handled with reassurance and became things you could point at.
- The four stage process outlived me, which was the entire point of designing it as a system rather than as my personal habit.
- The discovery fed the industry evaluation that followed. Six industries were scored on AI adoption, operational readiness, budget and time to a live implementation, and legal came out first. That decision was made around the time I left.
Product outcome
Six industries evaluated against four questions, and legal chosen as the first market to prove the model in.
What I learned
Discovery is not asking people what they want. Nearly everyone described a feature, and nearly none of those features were the real constraint. The useful signal came from comparing sectors. When a law firm, an insurer and a bank describe the same obstacle in three different vocabularies, that is a requirement. When one person describes something vividly, that is usually just one person. I also learned that narrowing the customer felt like throwing revenue away and did the exact opposite. The conversations got shorter and much better.
What I would do better
I treated all thirty conversations as equally valuable and they were not. Someone who had already written a policy banning public AI was worth several merely curious people, and I only spotted that pattern near the end. Next time I would sort by demonstrated intent from day one, and go looking for evidence that I was wrong. I spent more effort confirming my theory than trying to break it. I would also test the three product areas against something a customer could actually use, rather than against their description of their own problem.