At 11pm, your newest hire answers like your best one.
A general AI assistant knows the internet. It does not know how your team prices an exception, who gets escalated to, or what never gets said to a client, because none of that was ever written anywhere it could read. A Decisionlore agent knows all three, because it is built only on what your own people captured and verified.
Decisionlore agents are not yet available to customers. There is no agent to switch on in a workspace today, no agent seat to buy, and the watch, draft and autopilot ladder described below is the governance model we intend to hold ourselves to rather than a control that exists in the product yet. We have not published a rollout timeline. When we do, we will publish it separately. What the screenshots on this page show is the part that is live now: answers grounded in your own documents, naming the source they came from, and refusing and escalating when your knowledge does not cover the question.
An agent is one job, done the way your company does it.
Three shapes, to make it concrete. These are illustrations of the kind of job an agent takes on. They are not descriptions of anybody's account.
“Can we do this, for this client, at that price?”
The agent answers the way your company actually decides it, not the way a pricing table would. It draws on the exceptions your own people have already made and verified, tells the person asking which of those it used, and flags the cases where your company has historically said no.
“Who do I escalate this to, and what happened the last time?”
A new hire on a Friday evening gets the name, the reason that is the right name, and what happened the last few times a case like this one came up. That is the knowledge that normally lives with whoever happens to be reachable, which on a Friday evening is nobody.
“Draft the reply.”
The agent writes the first version in your house voice, using the phrasing your company uses and avoiding the things your company never says to a client. A person on your team reads it and sends it. The agent does not send anything.
None of the three is picked off a shelf. Which one your company gets, and whether your memory can ground it well yet, is decided by what is already in there. That is the next section.
Your own memory tells us which agent you need, so nobody is guessing.
It starts with a systematic audit of what your company's memory already holds, run with the access you grant for it and nothing more. That audit surfaces the real gaps: the questions your staff keep asking each other, the process only one person genuinely knows, the judgment call nobody ever wrote down. Those gaps are where an agent earns its keep, and they are different in every company.
Agents are proposed from what the audit finds, not picked off a shelf of templates that would fit anybody. And if your memory does not yet hold enough to ground a particular agent well, you get told that before you buy it, and the gap becomes the next thing worth capturing.
You can check every answer against the card it came from.
The reason to trust an answer is not that the software sounds confident. It is that you can follow the answer back to a card one of your own people verified, and that the software stops when it cannot.
Every answer names the cards it came from.
Each answer draws on your company's verified knowledge cards and names which ones it drew on, so anyone who wants to check the source can open it. Nothing is asserted from thin air.
When it does not know, it says so and stops.
When your memory does not hold the answer, the agent stops and says so instead of producing something plausible. A confident wrong answer is worse than no answer, here the same as everywhere else on this platform.
It only sees what your people verified and gave it.
An agent can reach exactly what a person at your company has verified and granted it, and nothing past that boundary. Cards still marked captured, waiting on a human to check them, are not in an agent's reach at all.
You decide when it is allowed to act, and you decide with evidence.
Every agent climbs the same three rungs, in order, and your company decides each promotion with a record of what the agent actually got right in front of you. Nothing moves up because we think it is ready.
Watch
You see how it would have answered, at no risk.
Your people do the work exactly as they do today, and the agent shows what it would have said. You get a free read on whether it is any good before it touches anything. Every agent starts here, with no exceptions.
Draft
You get the work already done, and you decide if it goes out.
The agent prepares the reply, the summary, the first pass, and hands it over finished. Your person reads it and approves it. This is where most of the value shows up, and nothing leaves your company without a human name on it.
Autopilot
You hand over one narrow task you have already watched it get right.
One specific task, chosen by you, after you have watched the agent do it correctly enough times to be sure. Autopilot is something your company grants for a named job, with the evidence in hand, and it is something your company can take back.
Your key governs the agents too.
Agents work only with the access your company granted, and whatever they produce stays your company's, sealed with the same key as every knowledge card underneath them. There is no separate door for agents to walk through. An agent gets better for one reason: the memory it draws on grew, or your people corrected something. Every improvement traces back to a real card or a real correction, and you can go and look at which.
About the agents layer.
How do we know it is not answering from something nobody checked?
Agents answer only from knowledge cards a person at your company has reviewed and verified, and only with the access your company has granted. Unverified cards, marked captured, are not in an agent's reach at all.
What stops it bluffing when it does not know the answer?
It says so and stops, rather than guessing. That is the same honest behaviour the memory layer follows when a document does not cover a question, carried through into the agent.
How much control do we have over what an agent is allowed to do?
All of it. Every agent starts in watch, then draft, and only reaches autopilot for the narrow tasks it has proven itself on, one promotion at a time, decided by your company with evidence behind each step.
Will it change how it behaves without us knowing?
It improves for two reasons only: the memory it draws from grew, or your people verified or corrected something. Every improvement traces back to a real card or a real correction, not to something learning unsupervised in the background.
Find out which agent your company can already ground.
The gap analysis looks at what your memory holds today, with the access you grant for it, and comes back with three things: the agent your company could ground well right now, the one it could not ground yet, and what would have to be captured first. If the honest answer is not yet, that is the answer you get.
Decisionlore agents are not yet available to customers. There is no agent to switch on in a workspace today and no agent seat to buy, and we have not published a rollout timeline. What the screenshots on this page show is the part that is live now: answers grounded in your own documents, naming the source they came from, and refusing and escalating when your knowledge does not cover the question. The gap analysis is about that memory layer.