AI you can actually
let near the business
Most AI tools try to catch a bad answer after the AI has written it. That's a guardrail you can drive straight through. We decide what the AI is allowed to do before it does anything — and that decision isn't made of words it can argue with.
Catching mistakes late is not the same as preventing them
Nearly every AI safety feature you've seen works like a spellchecker. The model writes whatever it writes, then something reads it and tries to spot trouble. Every "jailbreak" story you've read is someone getting past that reader.
Ours works more like a job description. Before a unit does anything, it has to fit inside what it was hired to do. The AI can change what it asks for. It can't change what it's allowed to have.
You don't write the rules. Your business already did
There's no settings page where someone types what the AI may do — that's exactly how these things go wrong. Instead we build the boundary out of things your organisation already has:
The job — the role it's standing in for, and who it reports to.The places — which sites, systems and information that role actually touches.Your rules — the limits you already apply to people doing that job, including what needs a signature.
Those three become one boundary. If it wouldn't be a reasonable thing for someone in that seat to do, the unit can't do it either.
People and machines, on the same map
Most companies end up with two separate systems: one deciding what staff can do, another deciding what the software can do. They drift apart, and the gap between them is where the trouble lives. We put both on one map.
Where a person stands.
You already know your job title. What's harder to see is the shape of your work right now — which calls are actually yours to make today, who you're answerable to on this particular thing, and whose trust you're carrying into it.
Places the peopleWhere a unit stands.
The same idea for something that isn't a person: exactly where it sits, what it may touch, and how far it's currently trusted to go. That position isn't fixed — it can move as a unit builds a track record, and it travels with the unit if it works across more than one part of the business.
Places the machinesThe point is that they're the same kind of thing. When an approval comes up, nothing looks in one place for the person's authority and somewhere else for the machine's. It reads both positions off the same map, judges the combination against the same rules, and writes both to the same record. One answer to "who's allowed to do what" — covering everyone, human or not.
Roles and tick-boxes. You either have access or you don't. It can't express "trusted here but not there yet", it doesn't move as someone proves themselves, and it stops at the edge of whichever system it lives in.
Not on/off — a place on a map, with degrees of trust. It can shift as someone or something earns it. Two of them can be combined when people and units work together. And it carries across the whole business, not just one tool.
MERV places machines. VERV places people. One map, one set of rules, one honest answer to who is allowed to do what.
Someone's name is on it
Every unit belongs to a real person. Not a team, not an account — a name. If it can't work out who that is, it won't switch on at all.
And when something's genuinely a judgement call, it doesn't guess and it doesn't quietly proceed. It stops and asks them. The AI never gets the last word on anything the rules don't clearly settle.
How that's different
| ChatGPT | Claude | Gemini | A MachineForce Unit | |
|---|---|---|---|---|
| What it covers | Everything | Everything | Everything | One clear job |
| Its limits | Written instructions | Written instructions | Written instructions | Built from your business |
| When it checks | After it answers | After it answers | After it answers | Before it acts |
| Managing it | One seat at a time | One seat at a time | One seat at a time | Across the whole team |
| Who's responsible | Nobody named | Nobody named | Nobody named | A person, by name |
| Paper trail | Chat history | Chat history | Chat history | Sealed and verifiable |
| Which AI | Theirs only | Theirs only | Theirs only | Whichever suits the job |
| Saying no | It decides to decline | It decides to decline | It decides to decline | It isn't the AI's call |
To be clear: those are excellent models, and we run all three. The difference isn't the AI — it's everything around it. A unit can be built on any of them.
Don't take our word for it
Talk to one of our units, then try to talk it into something it shouldn't do.
Try one free →