Four working solutions. Open any one.
Four working AI solutions with approved sources, clear boundaries and controlled workflows.
All four run right now, in your browser.
Nothing here is a recording or a scripted walkthrough. All four run the architecture described further down this page, and each will refuse you if you push it somewhere it should not go.
Explore with AI
Describe your situation in your own words. A Veren page is composed around it while you watch — the sections chosen for your question, not a template filled in.
- Decides which sections the answer needs: workflow, architecture, comparison, evidence
- Draws only on case studies we have published, and drops a quote it cannot match to its source
- Keeps the thread, so a follow-up refines the same page instead of starting a new one
- Serves a labelled example when the model is unavailable, and never passes it off as yours
Two minutes
That a site can answer the question you actually arrived with, instead of making you hunt through pages written for someone else.
In beta. Newly live, and still being watched under real questions.
Explore Veren with AIVoice assistant
Nine situations drawn from medical information desks, clinic admission and clinical trial enquiries. Speak or type, interrupt whenever you like, and watch the panel name the document behind every sentence.
- Answers only from approved documents, shown on screen as it speaks
- Stops when someone describes a real case, explains why, and names a person to contact
- Refuses clinical assessment, eligibility and prognosis, and says so plainly
- Discards audio, and sends a transcript only if you ask for one
Three minutes
That an assistant can be strict about its boundaries and still be useful inside them.
Blueprint studio
Choose a recurring workflow, answer a few questions about your systems and constraints, and watch the architecture assemble itself: sources, review steps, escalation paths and the systems it touches.
- Builds the workflow structure deterministically, then writes the language over it
- Marks every step that reads an approved source, needs a named reviewer, or escalates
- Falls back to the deterministic blueprint when the model is unavailable, and says so
- Gives you the result to keep, whether or not you talk to us
Five minutes
That the architecture of a workflow is decided before the model is chosen.
A working prototype. It is in the lab because it is still being built, not because it is finished.
Open the blueprint studioDeck builder
Drop in a document. It tells you what it found, asks three questions, then shows you the plan: what becomes a slide, what collapses to a line, and what it is leaving out, before it builds anything.
- Plans the deck before it writes it, and puts the plan on screen to be argued with
- Drops anything it cannot trace back to a section of your document, and lists what it dropped
- Says when a document is a reference rather than a presentation, and offers a real choice instead of a bad deck
- Exports a real PowerPoint file with editable text, native tables and the speaker notes
Three minutes
That the difference between a toy and a tool is what it decides to leave out.
A working prototype. What you drop is read in memory and never written to disk.
Open the deck builderAn AI solution is judged by what it refuses, not by what it can say.
Six layers, and what each one makes impossible.
This is the architecture underneath all four solutions above, and the one we build for clients. Every layer addresses a specific failure, which matters more than the technology name.
Sources
An indexed library of approved documents, each with its reference and version, addressed section by section. It is versioned alongside the code, so a change to an answer is a change with a history.
The system answering from something nobody approved. If a fact is not in the library, there is no path by which it reaches a caller.
Retrieval
An answer resolves to the document, its reference, its version, the heading and the quoted clause. The demo shows that on screen while it speaks, because it is available at the moment the sentence is formed.
An answer you cannot trace. There is no such thing here as a statement whose origin has to be reconstructed afterwards.
The model
Named in one line of configuration and swappable. The instructions are frozen so they cache and can be reviewed as a unit, and the customer's own data enters the request as data, described as data.
Lock-in to one vendor, and text inside a document quietly becoming an instruction to the system.
Safeguards
Detectors that run in code before the model sees the input, with the same rules repeated to the model afterwards. They are deliberately eager, they have their own test suite, and they are reviewed as safety code rather than as copy.
A case class the system must never handle reaching the model at all. The model can add a safeguard during a turn. It cannot remove one.
Approval and escalation
A named person reviews before anything leaves. Uncertain or restricted cases hand over to a named person and role, carrying the context gathered so far rather than starting the caller again.
Anything reaching a customer, a controlled document or a record without a person's name against it.
Limits and fallback
Every cap in one file: session length, per-visitor and per-day limits, a daily spend cap with a kill switch. When one is reached, or the model is unreachable, a deterministic assistant answers from the same sources.
A runaway bill, and a blank error page when the model is down. Every path has a defined answer that is not an apology.
We are not a notified body, an auditor or a law firm. Where a project needs a validation plan, a data protection impact assessment or a legal opinion, we say so and name who has to produce it.
What decides a turn, and in what order.
Most demonstrations put the model in charge and add rules to its instructions. That is the wrong way round because instructions are advisory, and a model under pressure will negotiate with them. In both solutions above, the model is the third authority in a turn, not the first.
The order below is enforced in code and is the same for voice and for text, because the turn logic exists once and the transports call into it. A rule cannot apply in one channel and go missing in another.
The wall clock
The session cap is checked before anything else happens. A session that has run out closes with its summary rather than continuing quietly.
The safeguards in code
Pattern detectors run against the raw input before a single token reaches the model. A stop here produces an authored response and a named handover, and the model is never consulted.
The model, with its tools
Only now does the model see the turn, with the document library and the systems it may act in. It may add a safeguard the detectors missed. It has no route to lift one.
The scripted assistant
If the model is unreachable or a spend cap has been reached, a deterministic assistant answers from the same library by lexical search. The visitor gets a worse answer, never no answer.
Six habits that survive contact with a client project.
These are the rules used to build the solutions above. They make each one safe to hand to a new user without supervision.
One engine, every channel
The turn logic is transport-agnostic. Voice, text and anything added later call the same code, so a safeguard exists once and cannot drift between channels.
Structure deterministic, language generated
The model writes the words. It never decides the shape. A workflow's steps, lanes and escalation paths are composed in code, which is why the output is reviewable.
Validated as it streams
Each object is checked the moment it closes, and the whole document again at the end. A malformed response degrades to the deterministic version rather than reaching the screen.
Every cap in one file
No limit is invented at a call site. Session length, per-visitor limits and the daily spend cap live together, so what the system will refuse to do is readable in one place.
Safety code is tested code
The detectors have their own suite and are reviewed like safety code. A false positive costs one polite sentence, and a false negative accepts something the system must never accept.
Degradation is designed
Every path has a defined answer when the model is unavailable, decided during the build rather than discovered in production.
Bring the workflow you would not trust to a demo.
The first conversation is about one process, what it touches, and who signs. If it is the wrong thing to automate, that is a useful answer and you get it on the call.
Find the right AI starting point.
Align the business case, governance and first implementation—without starting a large programme.