Part 2 of 2. The first post explained what superintelligence means for a business and why trust becomes the scarce thing. Read Part 1 →
In the first post I argued that when intelligence is cheap and every company rents the same agents, three things decide who wins. Whether you can control what acts in your name, whether you can supervise it, and whether it has the context to act well. Most of the public conversation about superintelligence is about those same three things at the scale of civilisation. I want to bring them down to the scale of one company, because that is where they get solved or not, and I want to start with something small.
Here is a situation that has happened in every company I have worked with. A promising lead mentions on a call that he is away until the third of the month and would like to reconnect when he is back. Everyone on the call nods. Nobody sets a reminder. Three weeks later someone remembers, but by then the lead has gone quiet or gone elsewhere.
Nothing about that was a knowledge problem. The company knew the return date. It was said out loud, on a recorded call, and repeated in a follow-up email. The problem was that the knowledge lived in one person's memory and one person's inbox, and neither of those is a system.
Multiply that by everything else. A delivery lead posts in a channel that some extra dashboard work might push the launch date. An investor asks about new opportunities in the last line of an email about something else. A project team agrees a plan in a meeting and then drifts from it over the next month, one small decision at a time. A supplier mentions a delay in passing. A customer raises the same complaint for the third time, to three different people.
None of this is hidden. All of it is evidence. The company's intelligence already exists. It is just fragmented.
Most leadership teams try to solve this with dashboards. The trouble is that a dashboard can only show you what made it into the CRM or the project tool, and that is usually the smallest and most out-of-date slice of what the company actually knows. The real record of the business is in the calls, the emails, the meetings, and the chat threads. That is where decisions get made and where risks first show up, and almost none of it is captured anywhere useful.
This was expensive before AI. In a business run partly by AI agents it becomes dangerous, for the reason I gave in the first post. An agent can only act on the evidence it can reach. If it can reach the CRM and nothing else, it will make confident decisions on stale, partial information, at speed, all day.
From scattered evidence to connected intelligence
We built Argus at Soligence to fix the fragmentation problem, and I want to explain it in plain language, because the idea matters more than the software. I also want to be clear about what it does today and where it is going, since the difference matters.
Argus captures the raw evidence of everyday work as it happens, which means the calls, emails, meetings, and messages. It then reads that evidence for the things that matter to running a business, namely facts, decisions, action items, risks, and open questions. Each one is pulled out, linked to the people and accounts it concerns, and tied back to the original call or email so anyone can open the source and see for themselves.
We call the result structured intelligence, which is a slightly grand name for a simple thing. Instead of six partial views of a relationship, there is one connected view of who said what, what was agreed, what is at risk, and what is still unresolved. It is the difference between a company that has data and a company that has understanding.
Knowing is only the start. The point is to put that understanding to work. So when the lead says he is back on the third, Argus connects the conversation to the date and suggests reaching out that morning. When you are raising capital and Maya has backed similar deals and recently asked about new opportunities, Argus connects the dots and suggests a call. When a project starts drifting from the plan, Argus compares the work underway with the decisions that were actually taken and suggests a review to get the team back on plan.
The recommendation shows up where the team already works, in Teams or Slack, with the context behind it. If you want to check, you open the evidence and watch the call or read the email. You never have to take the software's word for it.
That traceability sounds like a detail, but it is the basis for everything that comes next. Once AI agents start taking actions in your name, the ability to trace every action back to the evidence behind it is how you will be able to trust them, supervise them, and defend them to a customer or a regulator.
Supervising the agents that act for you
Which brings me to supervision. Nobody would hire fifty new account managers, give them no manager, no shared system, and no record of what they promised, and then let them loose on the customer base. That is roughly what a lot of companies are about to do with AI agents, because the agents are cheap and the tooling to manage them does not come in the box.
Managing them properly means having a layer above the agents that does three jobs. The first is what Argus does today. The second and third are where we are taking it, and they are only possible because the first is in place.
A shared record
The first job is to be the record the agents act from. Every agent that touches a customer, whether you built it, your software vendor supplied it, or a partner runs it, should read the current state of the relationship from one place and write what it did back to that same place. If that record already holds every decision, commitment, risk, and open question with the evidence attached, the agents are working from the truth instead of from whichever system they happened to be pointed at.
There is a strategic consequence here that I think most people miss. If the record lives in one place, it does not matter which AI model runs the agent. You can switch models every quarter as better ones come out and lose nothing, because the relationship lives in the record. Whoever owns that record owns the customer.
A place to supervise
The second job is to be the place people supervise. This does not mean reading every action. It means seeing actions ranked by how much they matter, with the ones that matter most routed to a person before they happen. A discount above a certain level, a change of scope, a contractual commitment, or an email to an unhappy client waits for someone to approve it. Everything else runs, but every action is logged with what did it, when, on what authority, with the reasoning and the underlying evidence attached.
You also want the system watching for things that look wrong, such as unusual volumes of outreach, commitments the delivery team cannot actually meet, or a client whose tone turns cold right after an agent contacted them, and a single place from which any agent action can be paused, reversed, or escalated. This is the feature set that turns "we use AI" from a risk statement into a compliance statement, and it is what a board member, an auditor, or a regulated customer will ask to see.
A system that learns
The third job is to be the thing the company learns from. Every renewal won or lost, every escalation, every delivery slip and what followed it gets recorded against the decisions that led there. Over time the system gets better at the judgments that matter for your business specifically, such as which accounts to prioritise, which early signals tend to precede churn, and which intervention usually works. The labs will keep making general intelligence cheaper and better. Nobody can rent the judgment that comes from years of your own relationships, properly recorded.
The way we run Soligence reflects this. Our lab tries things on our own operations first, what works gets packaged into Argus, our advisors take it to clients, and the outcomes come back into the lab. Argus is where those outcomes get recorded, which is why we run our own pipeline, delivery, and renewals through it with our own agents. If it is going to break, it should break on us first.
What to do this quarter
If you accept the argument, here is what I would do this quarter. None of it requires you to bet on a timeline. If self-improving AI arrives quickly, these steps put you ahead of it. If it arrives slowly, they pay for themselves anyway, because a company that can find its own knowledge and show its own work runs better with or without agents.
Find out where your company's knowledge actually lives. Pick one important customer and try to reconstruct the last ninety days of that relationship. What was promised, by whom, when, and where is the evidence? In most companies this takes a day and ends with a list of calls nobody wrote up, emails only one person saw, and decisions that exist in memory. That list is your fragmentation problem, made visible.
Start capturing the evidence. Calls, meetings, and email are where the business is actually run. If they are not recorded and connected, everything built on top of them is guesswork. This is the step most companies skip because it feels like plumbing, and it is the plumbing that everything else depends on.
Decide what needs a human before it happens. Write down, in plain language, the actions in a customer relationship that should never be taken without a person approving them, such as discounts above a threshold, scope changes, anything contractual, and certain kinds of communication with certain kinds of client. This becomes the policy your control layer enforces, and you can write it before you have a single agent running.
Pick one workflow and run it end to end with an agent, under supervision. Renewals are a good candidate because the evidence is rich and the stakes are clear. You will learn more from one supervised workflow than from a year of reading about AI.
Keep the record separate from the model. Whatever you build, make sure the relationship record lives somewhere that does not belong to the AI vendor. Models are going to change every few months. Your record of what you told your customers should not change with them.
Assign someone the job of judgment. Not a data scientist. Someone senior who decides what to automate, what to hold back, and what to stop, with the standing to say no.
The advantage you can build
I will end where the first post started. The fear version of this story says AI makes your business irrelevant. The hype version says you just need the latest model. Both are wrong in the same way. What a superintelligent model will never have is your evidence, your relationships, and your judgment about them, properly recorded.
The bigger point is this. Everyone calling for a slowdown is really asking for the same three things, control over what AI does, human review of the decisions that matter, and enough context that the decisions are good ones. Nobody is going to deliver those for the world in the abstract. They get delivered one company at a time, in the systems those companies run their relationships on. Argus is in its early days, and I am not going to pretend it does all of this yet. But it is built for that job, starting with the part that has to come first, which is turning what a company knows into evidence it can trust. A company with its knowledge in order is in a stronger position in a world of superintelligence, not a weaker one.
If you'd like to discuss Argus and see whether it's a fit for your company, let's talk.
Watch Argus
Previously: What superintelligence actually means for your business