# How it works | Lifter How it works · the mechanism # Can it do the job on Sunday, while you’re asleep? A fully supported agentic operating system If it only works while someone is watching it, it isn’t a workforce. Everything below is what happens once it is. The work, the controls, the reporting and the onboarding, in the order you meet them. Book a demo → See last night → Prefer the list to the argument? Every capability, grouped → Time Right now State - 03:00 . Mel reconciling overnight orders . Running - 06:41 . Brooke needs approval to refund $186.50 . Waiting - 07:00 . Riko delivers the daily performance brief . Queued 24/7 Never off shift 00 People who had to be there Always on and always learning · built around your business, not picked from a catalogue · work that completes with nobody present · hours the business didn’t previously have · every action on the record · credentials never reach the model · runs in your cloud or ours · 100 plus connectors live · every employee, inside guardrails you set · always on and always learning · The night shift · Tuesday into Wednesday ## What your workforce did while the business was closed This isn’t a claim about productivity. It’s an example of a night, shaped exactly as the platform reports one, between the time your team went home and the time they came back. Read the highlighted row first. Seven tasks ran with nobody present. One touched money, so it stopped, went to the person accountable, and waited two hours for an answer. The one that asked is the point. That isn’t a gap in the automation. That’s the control plane doing exactly what it’s there for. Time Activity State - 18:04 . Team goes home. Store closed. . Offline - 21:30 . Brooke triaged 14 overnight tickets, answered 11 . Done - 23:40 . Scout consolidated 17 learnings into memory . Done - 02:14 . Brooke answered a delivery enquiry, other timezone . Done - 03:00 . Mel reconciled overnight orders against target . Done - 06:30 . Lachie flagged coverage risk on 3 SKUs . Done - 06:41 . Brooke asked to refund $186.50 on a damaged order. Held. . Asked - 07:00 . Riko delivered the daily performance brief to #trading . Done - 08:47 . First person opens a laptop. . Online - 08:52 . Sam W. approves the refund. Recorded. . Closed 07 Completed with nobody present 01 Paused and asked a person first Inside an agent ## Not a prompt. Six parts your team can read and change. An agent is a self contained unit with its own identity, its own permissions and its own memory. Six tabs in the product, and six things you can inspect and correct. The last one is the difference between an agent and a chat window. 01 ### Identity A name, a job title and a written description of what it’s responsible for. Readable by anyone on the team. 02 ### Skills Written procedures for doing the job. Shared across the business, editable, and some the agent writes for itself as it learns. 03 ### Connectors The systems it’s joined to. Configured once centrally, allocated where needed, revocable instantly. 04 ### Tools Exactly which actions it may take inside each system. Scoped per tool, not per app. 05 ### Memory What it has learned about your business, in plain files your team can open and correct. 06 ### Scheduled jobs The work it does whether anyone asks or not. This is the part a chat window can’t have. Meet the agents → Control ## Trust is a dial, not a leap Every action an agent can take gets one of three stances, set per tool rather than per app. Start it narrow. Widen it as it earns the room. A prompt injection can raise a request. It can never take an action on its own. Allow Low risk, reversible, audited. Runs without asking. Drafting a note, building a dashboard, summarising a thread. Ask Pauses for a person in the channel your team already uses. The default for anything that touches a customer, costs money or changes inventory. Deny Not available to that agent at all. They don’t see it. They can’t try to use it. BR Brooke Customer experience 06:41 Asked before refunding $186.50 on a damaged order. Photo on file, customer waiting. Approve Deny Held for Sam W. in #cx-approvals Scope Riko, tool access Enabled - SYS 01 . Order management . 28 / 31 - SYS 02 . Customer support desk . 19 / 24 - SYS 03 . Email platform . 12 / 40 - SYS 04 . Finance system . 0 / 52 Scoped deliberately, per tool. The uneven numbers are the point. This is a decision someone made, not a switch someone flipped. Organisational, not personal ## The guardrails are what let you hand this to everyone Most businesses end up with AI in the hands of the three people who happen to be good at prompting. Not because the rest aren’t capable, but because nobody could safely give them access to real systems. Permissions set per person, per agent and per tool are what change that. Finance reaches what finance needs. Nobody else does. Everything is on the record either way. Everyone gets access ### People are free on every plan There’s no per seat fee, because a capability only three people can reach isn’t an organisational capability. Nobody gets more than they should ### Scoped per person and per tool The same agent behaves differently depending on who is asking. Your rules, enforced by the platform rather than by trust. It arrives as yours ### Under your brand, on your domain Your logo, your colours, your typeface, your URL. To your team it feels like a tool your company built, which is most of why they actually use it. Our product is quiet on purpose. It’s wearing your brand, not ours. This site is loud because it’s Lifter talking. The platform your team opens every morning is calm, neutral and carries your identity rather than ours. That isn’t an inconsistency, it’s the point of a white label product, and it’s why the thing you see in a demo will look like your company rather than like this page. The system ## One place where all of it lives Isolation, capability based access, shared storage and an audit trail are the primitives every operating system has been built on. Lifter builds them for a company's AI workforce rather than for its files. Layer 04 Your business Where the work gets asked for. Your people Your systems Where they already work Layer 03 Your workforce Everything your agents are, and everything they build for you. Agents Skills Memory Connections Automations Applications Layer 02 The control plane Every single action passes through here. Permissions Credentials Approvals Audit Configuration Layer 01 Models and tools Configurable by design, with a backup behind the default model, and changing the model touches nothing above it. Configurable models Failover to a backup Native connectors 100 plus tools How the platform works → Reporting · example report ## The four numbers we hold ourselves to. The platform measures its own output the way you would measure a team, so the value is visible without anyone having to build a business case for it. These four are the ones we think matter, and they’re reported from the first month. Work completed 1,284 Tasks finished by your agents this month, across every connected system. With nobody present 94 % The only figure that separates a workforce from an assistant. Everything else can be done by a chat window with someone typing into it. Paused for approval 76 Each one routed to the accountable person and recorded with its outcome. Proof the guardrails are live rather than theoretical. Outside business hours 61 % Work that happened when the office was empty. Hours the business didn’t previously have. Illustrative figures, shown to make the report concrete. We publish real ones as our foundational clients produce them. These are the numbers we want to be judged on. Most of the work completing with nobody present, and a real share of it happening when the office is empty. That’s what we’re building toward, and the platform reports it from the first month so you can judge us on your data rather than ours. The point ## Your best people, more productive than they have ever been. Almost none of the overnight work in that log was being done before. It was waiting until someone arrived, and then competing with everything else on their day. Now it arrives finished, and the people you hired because they’re good spend their week on the work that actually needs them. Effectiveness ### The week goes on judgement Nobody starts the day on catch up. The queue is cleared before your team walks in, so the hours they have go to the calls that genuinely need a person to make them. Productivity ### More gets done, and the standard holds Work runs the moment it can rather than the moment someone is free. Nothing sits in a backlog waiting for capacity, and the standard doesn’t slip when volume rises. The multiplier ### A great operator, with a workforce behind them The people already driving your business can see further and move faster than they could alone, and it compounds the longer they work together. How it starts ## Hiring, without the hiring process Adoption is a path we walk with you, not a platform we hand over. Anyone can generate an agent in an afternoon, and that’s a graduate on their first morning. We spend the weeks after that watching the work, correcting it, and widening what it’s allowed to touch as it earns the room. Day one ### Meet your team We introduce the agents built for your industry, scoped tightly to the work that matters to you first. Week one ### Onboarding Your own instance is stood up, configured and secured, carrying your brand on your domain and connected to the channel your team already uses. Week two ### On the job They go live inside real workflows, supported and monitored, while your team builds confidence through use. Day thirty ### Growing the team Adoption expands and new capability is added as you’re ready for it, always in a structured way. Common questions ## What people ask first How is this different from the AI assistant we already have? Those are personal agents, built to work for one person in one session. The moment a company needs that work to run safely, with proper access control, shared knowledge and a record of what happened, a personal agent hits the same four walls every time. Credentials, context, approvals, and work that only happens while someone watches. Which AI model does Lifter run on? Not one in particular. Lifter is agnostic to any specific LLM, and models are configurable by design. It runs on a default model with a backup behind it, so if the default has an outage, work fails over and your agents keep going. Agents, skills, memory, connections, automations and applications all sit on the platform, so none of what you’ve built depends on which model is doing the work. Where does our data live? Wherever you need it to. The runtime is self contained, with local configuration and local credential storage, and it deploys in your cloud, ours, or both, with your business kept isolated from every other one we run. Couldn't we build this ourselves? You could, and some businesses will. It takes a platform team building credential handling, sandboxed execution, durable approval routing and a proper audit trail, plus someone accountable for maintaining it as models keep changing. We run it, secure it and grow it with you. What happens to the AI work we’ve already paid for? It moves in. Your agents, your skills and the applications you already have can be brought across and keep working, with permissions, approvals and audit applied to them from the day they arrive. Next step ## See it running on your own systems. Reading about it only goes so far. Put one of your own processes in front of us and we’ll run it against systems you recognise, with your permission stances set the way you would actually set them. If it isn’t a fit we’ll say so on the call. Pricing is here. Would rather have had the list? Every capability, grouped → Enquiry ### Thanks, we’ve got it. We’ll reply within 48 hours. If it’s easier, pick a time in the calendar . Source: https://lifter.work/how-it-works