Large companies have spent $30–40 billion on generative AI. MIT NANDA found that 95% of the organizations it studied saw no measurable P&L return from GenAI. Its reason: the tools had no memory and no customization. They never learned how the business worked.
Change nothing. Transform everything.
Most “AI transformations” fail because they ask your people to work differently. We build around how your company already works.
Keep the workflow. Ditch the busywork.
Thirty minutes, no slides.
If there’s nothing here worth doing,
we’ll say so.
50M+
Downloads a month
48K+
GitHub stars on our tools
30K+
Operators read us weekly
92%
Client retention since 2023*
* at Automaze, our sister company – 25 people who build the software – since 2023.
What breaks
The answer exists.
Nobody can see the whole thing.
Somebody asks it in a meeting, and the room goes quiet.
Not because anyone is bad at their job. The answer does exist. It is just spread out: some of it in email, some in the system you run the work in, some in meeting notes, and some in four people’s heads. There is no single place where the pieces add up.
You see it later, as a write-off, a bill for work done twice, or a client who noticed before you did. You almost never see it as a problem someone is tracking.
Every write-off started as a promise nobody wrote down.
88% of companies now use AI. Only 6% qualify as McKinsey’s AI high performers. The companies getting value and the companies getting none are running the same models.
Gallup asked employees who have AI at work but do not use it why not. Forty-six percent said they prefer to work the way they already do.
Of the employees who do use AI, 61% are just chatting with a bot. The two most common uses – writing and search – are also among the lowest-rated for productivity in Gallup’s study.
Sources · MIT NANDA, The GenAI Divide · McKinsey, The state of AI · Gallup AI indicator
The adoption tax
The most expensive part of AI
is getting people to actually use it.
- Retraining.
- New processes.
- Adoption programs.
- Resistance.
- Change management.
Nobody puts it in the proposal. Everybody pays it – in the months before anything works, and in the people who quietly go back to the old way.
We remove most of that line item. There’s no new workflow to adopt.
How we work
We learn the company
before we automate it.
On the left, the systems it watches. On the right, the models it calls to do the work.
01
Observe
We watch how the work moves, not how the org chart says it does.
Channels you approve, meetings, and the systems your team already runs on. The engagement is paid, time-boxed, and useful even if we stop there.
02
Understand
Observation is not understanding. We turn what we see into a structured model of how the operation actually works.
Who decides what. Where work stalls. What depends on whom. What has been tried before, and why it did or didn’t work.
03
Build
Where the path is stable, we build fixed rails. Where judgment is required, we use agents.
Most of what looks like custom software doesn’t need to be. And when the right answer is “you don’t need us for this,” we say so.
04
Run
It goes into production at the autonomy level you choose.
Not a pilot with an expiry date, and not a prototype your team inherits as a maintenance problem. What survives the engagement keeps running.
05
Compound
The system improves as it works on your own operational data, while the model of the company keeps evolving with it.
You keep something you own, not something you depend on us to operate.
Meet Acropolis
Predictability starts with boundaries.
Acropolis separates observation, organizational memory, and execution into distinct layers with explicit responsibilities.
That separation is intentional. It makes behavior easier to reason about, trace, test, and govern – and prevents one part of the system from quietly accumulating responsibilities it was never designed to own.
If a layer’s behavior starts drifting beyond its defined role, we treat that as an architectural problem, not something to patch with another prompt.
Most AI systems grow by accumulation: another agent, another workflow, another integration, another layer. Eventually it still works, and nobody can explain why.
Acropolis is built from the opposite direction: clear boundaries first, then only the complexity the company actually requires.
Observation
Read-only by design
Connects to the approved systems, conversations, and meetings required to understand how the organization operates.
It can observe, structure, and pass information forward. It cannot initiate work, contact people, or change operational state.
Organizational memory
Company operational model
Holds the company’s current state rather than a pile of documents, along with the history of how it got there.
It is separate from both the raw source systems and the execution layer, so what the company knows is not coupled to whichever agent happens to be acting on it.
The model belongs to you and remains exportable.
Execution
Action happens at the edge
Execution is isolated from observation and memory. It receives the context and authority required for a task, then operates through Slack, Teams, and the systems your people already use.
Every action is bounded by permissions, visible while it runs, and attributable afterwards.
Control
Your org chart. Your rules.
Access inherits the roles and permissions you already have. If someone can’t see the finance channel today, nothing we install changes that.
Leadership gets a single view of what’s running, what acted, what escalated, and where each autonomy setting sits.
Autonomy is explicit
- Watch Reads and reports what it finds.
-
Suggest
Prepares the action;
humans execute it. -
Gated
Prepares the action;
humans approve it. - Notify Acts, then reports what it did.
- Auto Acts within its defined permissions.
You set the level. You change it when the evidence earns it.
What you keep
Infrastructure. Not a slide deck.
Your operational model
A live picture of how the company actually works – who owns what, who decides what, what depends on whom, and what changed.
It belongs to you and remains exportable. You are not locked into a proprietary file format or a consulting deliverable that goes stale.
Production systems
What we build runs in production, at the autonomy levels you choose.
Not a pilot environment. Not a prototype your team has to rebuild after the engagement ends.
An asset that compounds
The system continues learning from your own operational data as the company changes.
Every month it runs, it knows more about the company than it did the month before.
What the engagement installs is a resident AI: your operational model, running in production, on infrastructure you own. We also leave the team with the context to operate it: a short orientation for users and a working session with leadership on governance, oversight, and autonomy.
Where it runs
It runs inside your company
This is what we mean by resident AI. It runs on your infrastructure, inside the tools your people already use, across the whole operation rather than the part someone pasted in, and what it learns stays yours.
Acropolis runs on infrastructure you control – your cloud, your network, your security perimeter. There is no VarOps-hosted service and no data plane on our side. Your operational knowledge stays where it already lives, and the brain we build is yours to own and export.
Start small
Pick one function.
Prove it in production.
IT and internal support
The small problem that used to cost forty minutes and a support ticket gets solved where it happened. Anything that cannot be fixed reaches your IT team with the cause already found. Every employee feels it in week one.
Finance operations
Chasing payments, matching records, the reports nobody has time for. High volume, clear rules, mostly rails.
People operations
Onboarding, access requests, policy questions, approvals, and the repetitive coordination around people joining, moving, or leaving. The rules are known; the work is just scattered across systems and people.
Client operations
Follow-ups, handoffs, status updates, document chasing, and the work that falls between sales, delivery, and account management. High coordination cost, lots of context, and usually no single system that owns the process.
Who this is for
We’d rather you know now than in month four
This works when
- You’re a $50–500M business
- A senior owner can make decisions without a six-month cycle
- You’ve tried AI, but it hasn’t materially changed how work gets done
- You don’t want to build a large internal transformation team
- The operational friction is real enough that leadership can name it
This doesn’t work when
- You want a pilot, report, or proof of concept
- You’re mainly looking for training, workshops, or AI literacy
- The decision requires a long committee or procurement process before anything can run
- You already have the mandate and team to build this internally
- There is no specific operational problem leadership wants solved
Why VarOps
We build the systems we talk about.
We wrote the books
Company-Scale Agentic AI is the operator’s guide to a company that runs on intelligence. Production-Grade Agentic AI is what engineers open when the method needs explaining. We didn’t read about AI-native operations. We wrote down how they’re built.
We open-source what works
MUXI, Agent Formations, FAISSx, OneLLM, Proof, yfinance, and many more – are running in production at companies we have never met. We do not just talk about software that has to work at scale.
We deliver
Automaze, our sister company, has kept 92% of its clients since 2023. The risk in any transformation is that someone describes the problem well and then builds the wrong thing. That number is our answer.
We publish our numbers
1 in 2,000 scored answers contains a fabrication. How Acropolis performs, and how we measured it, is written down where anyone can check – including the parts where something else does better. See the benchmarks →
Before the call
The questions you should ask.
What is resident AI?
AI that lives inside the company that owns it. Four things have to be true.
It runs where the work already happens, so nobody has to adopt a new tool. It can see the whole operation, not just the part someone remembered to paste in. It runs on your own infrastructure, with no vendor-hosted service in the middle. And what it learns belongs to you and can be exported.
Most company AI fails at least one of those. Ours is built to pass all four, and you can hold us to it.
Learn more on the resident AI page.
Is it reading our private messages?
The part that watches sits only in the channels you approve. It sees exactly what your permissions already say a person in that role can see – if someone can’t open the finance channel today, nothing we install changes that. And it is one-way: it never posts, never replies, and never messages anyone.
Where does our data actually live?
Inside your infrastructure. Acropolis runs in your environment, and nothing is sent to us automatically – no telemetry pipeline, no external control plane. Where our team needs access to support the system, we log in through your own access controls, the same way any service provider would.
The full architecture is documented on our Acropolis page.
How is our data handled and kept private?
We agree the scope in writing before anything connects: which systems, which channels, which mailboxes, and what is excluded. It reports patterns, not people. What it learns lives in your company brain, which you own and can export at any time. And we tell your people what it can see before it sees anything, because a system nobody agreed to is a system people work around.
Do we own the company brain?
Yes. You own it and you can export it. What you export is the understanding itself: who owns what, what was decided, what changed and why.
Do we always end up needing custom software built?
No, and we’ll say so. Most of what looks like it needs custom software doesn’t – a lot of it is rails on systems you already run. When the answer is “you don’t need us for this part,” that’s the answer you get.
How long does an engagement take?
From the start, the first loop takes thirty days: answering real questions from your own evidence, and running at least one approved action in production. After that there is no end date. You are paying for a machine that stays current, not for a project that gets delivered and left.
What’s the difference between VarOps and Automaze?
VarOps does the watching, the understanding, and the machines. Automaze is our sister company – 25 people who build software, 92% client retention since 2023 – and it is who builds when a transformation needs hands. Same people behind both, different jobs.
How much does it cost?
There is no list price, because the scope comes from what we find in the first step, not from a package. The shape is always the same: a fixed fee to build, then a monthly fee to keep the machine running. The first step is paid and has a fixed end date, so you learn what a build would cost before you commit to one. And you get a range on the first call, not three meetings later.
Why is there a monthly fee?
Because a brain nobody feeds goes stale, and a stale brain is worse than no brain: it is confidently wrong. The monthly fee keeps the understanding current and the machines running. It is not support, and it is not hosting.
Next step
Let’s see if there’s something worth transforming.
Give us a little context first. Then thirty minutes: you tell us what needs to change, and we’ll tell you straight whether there’s a transformation worth doing – including when there isn’t.
Nothing to prepare.