Enterprise software is built org-down. Systems demand things from the people doing the work, such as data entry, status updates, and adoption of the next platform, so that the organization can have records, reports, and now “AI readiness.” Every generation of software has raised this tax, and every generation of workers has paid it.
We build the other way: person-up. Start where work actually happens, with the individual. Capture reality there. Derive everything else: the records, the reports, the organization’s intelligence.
We call this first-person AI. It works from where the person sits: it sees what they see, holds what they hold, acts with their access, and works on their behalf. Its opposite is third-person AI: data extracted, consolidated, and watched from outside.
Systems should observe work, not tax it. Everything below follows from that sentence.
Work is one job, not two.
Somewhere along the way, knowledge work split in half: doing the work, then recording, tracking, and reporting the work. The second half was always overhead. We tolerated it because records required human hands. They no longer do. AI’s first assignment is to take the second job entirely: watch the work happen and produce the records, the updates, the reports.
Not everything called reporting is mechanical. A forecast commit is a judgment, not an observation. So the division of labor is precise: AI drafts from what it observed; the person confirms what carries their name. But the typing, the logging, the reconstructing of last week from memory: that ends now, and no one will miss it.
Records are observations, not testimony.
Today’s systems of record hold what people say happened: stage guesses, hurried notes, fields filled under duress on a Friday afternoon. Captured-at-source data holds what did happen: the email as sent, the meeting as spoken, the commitment as made. CRM hygiene has been an unsolved problem for twenty-five years because it rests on human data entry, and no amount of training, nagging, or mandatory fields has fixed it. First-person AI does not improve data entry. It abolishes it. Full fidelity is what you get in exchange: a record made continuously, at the moment of interaction, by something that never tires of writing things down.
The system of record moves to where work happens.
Follow the first two principles to their conclusion. When reality is captured upstream, at the email, the call, the meeting, the CRM stops being the source of truth and becomes a downstream report. So does every system whose value rests on being the place people type things into. The record of your business lives where your business happens. Everything else is a view.
Intelligence is a layer, not a feature.
Every vendor in your stack now sells AI: an assistant in the CRM, another in the marketing platform, another in the meeting recorder, another in the support desk. The same frontier intelligence, bought four or five times over, each copy behind its own credit meter, each seeing a fragment of your customer. Buy intelligence once, from the model provider you choose, and bring it to everything you use. What you invest in making it good (playbooks, prompts, context) accumulates in one place that belongs to you instead of scattering behind vendor walls. We have made this argument at length in The Case Against Embedded AI. The short version: most teams have this default backwards, and the burden of proof belongs on the feature.
Access begins with the person.
The standard enterprise AI project extracts everything, consolidates it in one place, rebuilds permissions on top, and only then lets anyone ask a question. That rebuild is where AI projects go to die: months of governance work standing between an organization and its first useful answer. First-person AI inverts the order: the AI works with the user’s own scope, under the user’s own credentials. The permission model you already have is the access layer. Nothing needs rebuilding, and one person gets value on day one without a platform project, which is also how tools actually spread through organizations.
Some questions are larger than one person’s scope; territory plans and whole-pipeline analytics need aggregation. Aggregate later, from consenting individuals, once individual value is flowing, rather than letting the org-level fifth of the value block the individual four-fifths. And an agent holding a person’s credentials must be engineered with care, but its failures are bounded by what that one person could already see. A global data lake with hand-rebuilt permissions fails for everyone at once.
The organization owns its data because its people captured it.
First-person describes vantage point and scope, not possession. Capture happens at the individual, inside a workspace the organization keeps, and the graph outlives the employee. This is how organizations come to own their data in the only sense that matters: full fidelity, captured first-hand at the moment of interaction, held in their own store, not left inside SaaS products to be negotiated back later through APIs that meter it, charge for it, or quietly close.
The old fear, that the rep leaves and the relationships leave with them, is the reason CRMs were invented. First-person capture answers it better than the CRM ever did, because the record was made completely and continuously rather than typed in reluctantly.
The human remains the author.
The industry is racing toward agents that act autonomously in your name. We take the opposite bet, and the difference lies in which loop the person belongs in. Observation needs no approval: records, tracking, and reports derived from what happened should flow without anyone stamping each one. Otherwise the second job returns, dressed as a review queue. Action needs an author: anything that commits you (an email in your name, a forecast you will be held to, a promise to a customer) leaves by your hand. First-person AI removes people from the loop of paperwork and keeps them in the loop of commitment. This also bounds the machine: an agent that cannot commit without you can only ever fail as far as you could.
The watcher works for the watched.
AI that observes work is one design decision away from software that surveils workers, and what separates them is architecture, not intent. The capture belongs to the person. What flows upward flows because they release it, not because someone above them taps it. Reports are derived for the worker first and shared by the worker’s hand. Consent is not a settings page; it is the structure.
This principle binds us before it binds anyone else. Any vendor, ourselves included, who builds passive capture that reports to the boss first has built a panopticon, and should be called one.
The stack we are describing is smaller than the one you have: one intelligence layer you chose, connected by each person’s own hand, writing records no one had to type, into a store the organization keeps. Person-up. Derived, not demanded. Owned, not rented.
We believed this enough to build Actioner on it: no model inside, first-person by design. Read us accordingly. And hold us to principles 7 and 8; that is what they are for.
