Your software remembers everything
None of it does the work.

Your software remembers everything and does none of the work

Open any professional practice and count the software. There is a CRM, or a vertical system that behaves like one. There is a mailbox, a calendar, a document store, an accounting package, a portal belonging to somebody else that you log into twice a week. Every one of them is excellent at the same single task: remembering what happened. Not one of them does anything about it.

This is not a defect. It is what those products are, and the distinction is old enough to have a literature. Geoffrey Moore named it in 2011, separating the systems of record that hold the transaction from the systems of engagement built around how people actually work.^[R3] Fifteen years later the record half is close to solved. Storage is cheap, the schemas are mature, and a practice of four can run a stack that a bank could not have bought in 2005.

What did not get solved is motion. A mortgage file waits eleven days for one pay stub. An insurance claim sits because the adjuster replied to a thread nobody was watching. A legal intake stalls between the signed engagement letter and the first substantive step. A placement cools because the candidate went quiet on a Thursday and Monday had its own problems. In every one of those cases the software knew. The record was accurate, the field was correct, the date was right there. Nobody was short of information. What was missing was somebody to act on it.

That person is the constraint on the whole business. Not their judgment, which is the part worth paying for, but their memory, which is not. A book of thirty open files is roughly the point at which a good professional stops being able to hold all of them at once, and every file added after that is carried by a reminder, a spreadsheet, or luck.

Why the record system was never going to fix this itself

The obvious answer is to make the system of record do more. Every vendor in the category is trying, and some of them are very good at it. There is a structural reason it only goes so far, and it has nothing to do with how well any of them build software.

A system of record sees its own records. That is the shape of the thing. But the work that stalls a file almost never happens inside one system. The pay stub is in email. The signature is in a portal. The appraisal is a PDF somebody attached to a thread. The reschedule is in a calendar. The client’s actual state of mind is in the tone of a reply that no field on any record captures.

So an agent that lives inside one of those systems starts the job already unable to see most of it. It can do excellent work on the slice it owns. It cannot chase a document that has never been in its database, because from where it stands the document does not exist.

This is the reason the interesting layer is a layer, and why it has to sit above rather than inside. Not because any record system is bad, but because the work that needs doing is defined precisely by the fact that it crosses them. Anything that can only reach one system is solving a smaller problem than the one a practice has.

What a system of action is

This section describes the end state. Some of it runs today and some of it does not, and the ledger in section seven separates the two line by line rather than leaving you to guess which is which. But the shape of the thing is the argument, so here it is whole.

InVice is a system of action over the systems of record. The phrase is doing real work, so take it literally. It does not replace the record. It reads it, works out what should happen next, and does that thing in the places where work actually happens, which is your mailbox, your calendar, and the tools you have connected.

The unit is a Mission. You give it an outcome in plain language rather than a sequence of steps: get this file to closing, get this claim to a decision, get this candidate to an offer. A Mission is durable, which is the part that matters most and is easiest to skip past. It does not run once and finish. It holds a goal across weeks, wakes on its own schedule, notices that eleven days have passed without the pay stub, and does something about it. When the situation changes it works out a different route to the same outcome rather than failing a branch somebody configured in advance.

Above Missions sit Operations. A Mission carries one client; an Operation carries a population and coordinates the Missions inside it. That is the difference between an automation that runs per record and a view of every open file at once, where the question “which of my two hundred clients is quietly stuck” has an answer that does not require a status meeting to produce.

Underneath both is continuous client context. Not a chat history, and not a summary regenerated on demand, but a persistent understanding of a relationship that is isolated to that client and accumulates. The intention is that the fortieth interaction is informed by the first thirty-nine, and that the context belonging to one client never leaks into the work of another.

What the end state looks like from the outside is narrow and specific. The files move. Nothing that needed a human decision was decided without one. And the professional spends their week on the part of the job that required them to be a professional, because the part that required only a memory is no longer theirs.

The layer sits outside on purpose

Every vendor eventually asks a buyer to move something. That request is where most of this category dies, and avoiding it is not a sales tactic here. It is the architecture.

Our position is fixed and it is stated in our own founding document: no migration of your vertical system is required. You adopt InVice as your operations workspace and your existing stack stays exactly where it is. The firm keeps its stack. InVice adds the layer that stack is missing.

The reason this is architectural rather than commercial follows from the last section. If the defining property of the work is that it crosses systems, then a layer that required you to consolidate into it would have destroyed the thing it was built to solve. A system of action that owns your record is just another record.

It also has a practical consequence that matters more than the principle. When adoption requires a migration, the trial is a project, the project needs a budget and a quarter and a person to own it, and the evaluation never starts. When adoption is connecting a mailbox and describing a process, you find out whether it works this week.

One more piece of the same idea, because it is the part buyers ask about last and should ask about first. Our pricing attaches to a workspace and a band of people, not to a seat. The published plans run from one person to fifty and above. Adding the sixth colleague to a workspace does not multiply the bill, which means the thing that decides whether a pilot widens is whether it worked, rather than what widening it costs.

Authority is the hard part

Everything above is the easy half. Software that acts on your behalf, in your name, toward your clients, raises a question that software which merely remembers never had to answer: what is it allowed to do, and what does it do when it is not sure? This section is present tense. These are controls that run today, and each one is something you can check.

Consequential sends wait for a person, by default. Not as an available setting, which every product has, but as what ships when nobody configures anything. The question worth asking any vendor is never whether approval is supported. It is what happens when no one has touched the defaults.

A connected tool is approved one at a time, and the read asks first. Looking up a record in a system you have connected requires an unexpired approval grant, on every call. Not the first call. Every one. The AI agent does not get admitted because it asked convincingly, and not because a session happens to be attended by a human.

Untrusted content carries no authority. An inbound email and an uploaded document are data, never instructions. Neither can invoke a tool or change what the AI agent is permitted to do. That is enforced in code rather than requested in a prompt, because a model instructed to ignore instructions is not a control. It is a standing architectural constraint, which means a change of model never relaxes it. We are not going to tell you that manipulation is impossible. We are telling you where the boundary is drawn and what enforces it.

The model key is yours. You connect your own provider account and the AI agent runs on it, across six providers configured in production, including a self-hosted endpoint. If a better or cheaper model appears, moving to it is a setting rather than a renegotiation. The honest caveat travels with the claim: you choose your model and where it runs, so if you connect a United States provider key, your prompts are processed in the United States.

A named human owns every Mission. Judgment, approval and accountability stay with a person, and the audit trail records who approved and when. Client data does not train a shared InVice model. No analytics or error-tracking SDK is wired into the product; the public landing pages count visits and two actions with Vercel’s cookieless analytics. Client records and vault documents are configured in a Canadian region, in Montreal.

What we will not do is tell you what any of that makes you. Whether these controls meet the obligations your practice actually carries is a conclusion about your practice, and a vendor who reaches it on your behalf has told you something about their sales process rather than their architecture. InVice is designed around PIPEDA, Quebec’s Law 25 and CASL in Canada, and the GDPR in the European Union. The security brief states what is configured, hop by hop, including the hops that are not in Canada.

The same shape almost everywhere

We started on mortgage files and that is still the first desk, because going deep on one profession is how a product learns the difference between a document slot and a milestone. But nothing in the description above is about mortgages.

A file that lives for weeks, touches several parties, waits on documents, and dies quietly when nobody chases it is not a feature of one industry. It is the shape of client work. This model repeats in insurance claims, restoration, legal defense, property management, and any document-heavy client operation where files live for weeks and touch multiple parties.

The engine is horizontal and the experience is vertical. What changes between professions is vocabulary, document templates, milestone names and the relationships that matter. What does not change is the mechanism: a durable goal, a persistent context, an action taken, a human holding the decision. That is why recruiters, insurance advisors, legal intake teams, property managers, immigration consultants and realtors each have a page here, and why they are variations rather than separate products.

It is worth saying where it stops. There has to be a file. A business where the client buys once and the relationship ends at delivery has nothing for this to move, and no amount of engine makes one appear.

Where we actually are

A vision piece that does not say what exists is a brochure. Here is the ledger, in three groups. We would rather you arrive already knowing which is which than discover it in a demo.

Running today. Missions that hold a goal, wake on their own schedule and act in your connected mailbox and calendar. Approval as the default on the send path. Per-tool approval on connected systems. Your own model key across six providers. Multiple people in one workspace across four roles, with a named owner on every Mission. Retrieval over your own documents, scoped to the Mission that asked. A public agent gateway you can put on your own website. Client records and vault documents configured in a Canadian region.

Being built. Writing updates back into the vertical systems we read from, which is the half of the loop that is not closed yet. Approvals and notifications routed to the right person at the right step, internally or through Slack, rather than to the workspace owner. Deeper population-level visibility across an Operation.

Open, and stated as open. End-to-end verification of where client data is processed at every hop, which is why this page says records and documents are configured in a Canadian region rather than making a residency claim. Two of those hops are known to run outside Canada today and are named in the security brief. Until that verification is finished and recorded, the weaker sentence is the only one we are willing to write.

Gartner expects task-specific AI agents in more than forty percent of enterprise applications by the end of this year, up from under five percent the year before.^[R1] It also expects more than forty percent of agentic projects to be cancelled by the end of 2027, and the causes it names are escalating cost, unclear value and inadequate risk controls.^[R2] Both forecasts are about the same moment. The category is arriving, and most of what arrives will not survive contact with a real practice. That is the environment we are building in, and it is the reason this section exists.

Where this is the wrong answer

If you cannot find the paragraph where a vendor loses, you were right not to trust the rest of the page. Five of the six rows below point somewhere other than us, and the first two say the honest thing, which is that plenty of practices should buy nothing at all.

The pattern across those rows is worth naming. The more your work already lives inside one platform that somebody here configures and owns, the better that platform’s own agent layer gets and the worse we look. We are not going to pretend otherwise. Our case is the opposite shape: a practice whose client work is spread across tools that were never going to consolidate, where the record is one of five systems rather than the centre of gravity, and where nobody is employed to configure software.

If the category itself is still unfamiliar, the field guide to agentic CRMs maps who is building what and where each of them is strongest. If you are weighing an agent bundled into the platform you already run, we wrote about what that arrangement costs to own. And if the description in section one was uncomfortably close to your week, the place to start is what this looks like on a real desk. We would rather be the right answer for fewer people than the plausible answer for everyone.