# Our approach - Radexus > Nearly every enterprise AI project that fails, fails for the same reason. Not the model. Not the idea. The plumbing underneath it. *Source: https://radexus.com/approach/* --- Approach # The model was never the hard part. Nearly every enterprise AI project that fails, fails for the same reason. Not the model. Not the idea. The plumbing underneath it. The pilot works because somebody hand-cleaned the data for the demo. Production does not, because in production the same customer exists five times, finance and sales disagree about what open pipeline means, nobody can trace where a number came from, and no one is allowed to write anything back into the system that matters. **That is a plumbing problem.** It is dull, unglamorous work. It is also the work. Two people in this room would give you different numbers for open pipeline. Both would be right. *01*Why deployments fail ## Five things go wrong, in roughly this order, in almost every company. - 01 **The same customer has four names**They sit in your ERP, your CRM, the dealer portal and the service app under four spellings. Every report built on top is quietly wrong, and everyone defending their own report is quietly right. - 02 **Nobody wrote down what the words mean**Open pipeline. Stalled quote. Active dealer. Everyone uses them and no two people mean the same thing. Software will not settle that argument. It will speed it up. - 03 **Nobody can say where a number came from**It appears on a screen and no one can trace which system it came from or when. The first time it looks odd, it stops being believed, and belief does not come back. - 04 **Nothing is allowed to write back**The output lands in a spreadsheet nobody opens, because letting software touch the ERP is unthinkable. Insight that cannot become an action is entertainment. - 05 **Nothing is checked after go-live**No thresholds, no tests, no gates. Quality drifts because the world moved, not because the code changed, and the first person to notice is someone senior losing confidence. **None of these are model problems.** A better model built on the same plumbing gives you a more articulate wrong answer, faster. *02*How we work ## Number first. Build the memory once. Ship one thing. Check everything. **Agree the number before anything is built** With the CFO in the room, not the functional head. Written definition, written exclusions, a named owner. If it cannot be computed from data you already keep, it is not a number, it is a hope. **Build the memory once** Resolve the customers, write down the definitions, put a source on every field. It is the expensive part and the part everything else reads, which is why the second play always costs a fraction of the first. **Ship one play, not a platform** One thing, live, against the agreed number, with the pass mark written before it is built. It clears or it does not, and both outcomes get reported to you. **Let it write back, safely** Dry run first, then writes approved by a named person, with a full audit trail and a rollback window. Insight that cannot act is not worth building. **Gate every release** Tests with pass marks, several of them scored by your team rather than ours. Below the mark, nothing ships. There is no override. **Hand it over, then keep it true** You own the system and the memory in it. We stay on to keep the number true as your business changes, which is a different job from maintaining software. *03*What we are not ## We are not an agentic AI company. There is a large industry selling agent frameworks, orchestration layers and platforms you adopt, configure and then staff. We are not in it, and the distinction matters more than it sounds. Whether a system uses an agent loop, a state machine or a scheduled job is an engineering decision, made per play, by us, in your build. It is not a product, and it is certainly not something you should have to have an opinion about. **The hard part was never the loop.** It is knowing that these four records are one company, that this definition of open pipeline is the agreed one, that this claim resolves to a real source, that this knowledge is eight months old and should no longer be trusted, and that this write needs a named person to approve it before it touches your ERP. Buy an agent platform and you have bought a way to run steps. You still have all five plumbing failures, nothing accumulates, and now you also have a framework to maintain and people to hire who know it. - **Not an agentic AI platform**Nothing here for your team to adopt, configure or staff. We use whatever control flow a play needs and you never see it. - **Not a software vendor**No seat count, no per-user pricing, no renewal that can switch your system off. - **Not a consultancy**We do not bill bodies by the month and we do not deliver decks. The output is a working system. - **Not a systems integrator**We do not migrate you onto anything. Your stack stays exactly where it is and we read it. *04*The engagement ## Seven stages. Seven exits. Each stage has a gate, and you can stop at any of them keeping whatever has been produced so far. Nothing here requires a decision you cannot reverse. 01 ### First meeting We tell you what we think your problem is, from the pattern we see in your sector, then run a play live on companies you name. You provide four customer names. Nothing else. You pick one problem, or we stop here. No fee*60 minutes* 02 ### Discovery Read access to your systems under NDA. We map how revenue actually moves, check what your data can support, get finance and commercial to agree the definitions in writing, and put a number on the leak using your figures rather than ours. If the number is not worth several times the work, we say so and you keep the map. Chargeable*2 weeks, credited* 03 ### Pilot scoping We pick one play, define the metric, and write the sign-off criteria before anything is built. You name an owner for the metric and an approver for any write. Signed criteria, or no pilot. No fee*1 week* 04 ### Pilot delivery We build the slice of memory the play needs, ship the play, and attach its evaluation gate. Your team marks results right or wrong weekly, and those marks become the first golden dataset. It clears the metric or it does not. Both are reported. Chargeable*5 to 8 weeks* 05 ### Full scoping We price the rollout from what the pilot proved, not from what anyone hoped. You decide the scope and the second play. Priced against measured value. No fee*1 week* 06 ### Build and handover Production deployment in your cloud or on premises, write-back gated by named approvers, definitions versioned. The memory is built once, so every play after the first costs less. Nothing promotes without clearing its evaluation suite. Chargeable*10 to 14 weeks* 07 ### Assurance A named engineer inside your team maintaining the plays, and someone who owns the number in your quarterly review. Definitions re-versioned, confidence re-calibrated, gates watched, models re-routed as the market moves. Annual, cancellable, and the system keeps running if you cancel. Annual*Held against the metric* **Why the discovery fee exists.** It is the qualification. A company that will not spend two weeks finding out will not spend six months building, and finding that out in week two costs both sides far less than finding it out in month six. The fee is credited against the rollout, so it costs you nothing if you go ahead. ## Bring four customer names. We do the rest before we meet. Sixty minutes, no fee, and a real result on the screen rather than a proposal in your inbox. [Start a discovery](https://radexus.com/start/)