Nobody picks a private AI platform against nothing. They pick it against five other things they could reasonably do instead.
Every organization weighing a governed AI capability is already doing one of six things. This section takes each of them seriously, says plainly when it is the better choice, and shows the small number of questions on which the answers genuinely differ.
A vendor comparison usually pretends the alternative is an empty room. It never is. By the time anyone books a call there is a default already running, a subscription already paid for, an internal build somebody has costed, or a proposal from a firm with a logo on it. The question is not whether to spend. It is which of these six paths puts a governed capability in front of your people, and what each asks of you in return.
Bought per seat, running on the vendor's systems, live the afternoon you sign. The fastest path to ordinary productivity. The data path is the vendor's, described in their documentation rather than drawn on your architecture.
Already installed, already invoiced, already approved by somebody. It reads what your people can already read, which is both the whole of its value and the whole of its risk.
Real control, and with the right people a legitimate answer. The cost is not the first demo. It is identity, policy enforcement, metering, isolation and release tooling, maintained against a market that moves every few months.
Institutional trust, procurement familiarity, breadth, and a name your board recognises. What arrives at the end varies: sometimes a running system, sometimes recommendations and a plan to build one.
The strongest technical alternative, and the only one where the weights sit inside your perimeter. Provisioning, patching and inference optimization become a standing function you staff.
The most common option, and the only one nobody writes a business case for. It is not zero AI. It is AI without an account, a policy, a record or an owner, running in browser tabs on personal subscriptions.
Every option on this page can hold a conversation, summarise a document and draft an email. Comparing those capabilities is how a decision takes six months and lands nowhere. These are the questions where the answers diverge, and where changing your mind later is expensive.
| Path | Data at rest | Scopeable for review | Account and exit | Time to governed use | What it costs to run | Who maintains it |
|---|---|---|---|---|---|---|
| orgOS on a cloud you own | ✓Your account | ✓Architecture and posture | ✓Yours, both | △Weeks. We quote no fixed number of days | Your cloud bill, plus $500 a month | Us, from pinned releases |
| A SaaS AI assistant | ✕Vendor systems | △Their documentation | ✕Vendor holds it | ✓Same day | Per seat, per month | The vendor |
| A bundled suite copilot | ✕Suite vendor systems | △Your sharing settings | ✕Bundled with the suite | ✓Already installed | Per seat, above the suite | The vendor |
| Building it yourself | ✓Your account | △Whatever you document | ✓Yours, both | ✕Quarters | Staff time, plus the cloud bill | Your platform team, indefinitely |
| A consultancy programme | △Depends what is built | △A report, then a build | △Depends on the contract | ✕Quarters | Programme fees, then support | Whoever is retained next |
| Self-hosting open models | ✓Your account, weights too | ✓Entirely yours to draw | ✓Yours, both | ✕Months, plus a new function | Accelerators, running or reserved | Your own inference team |
| Doing nothing | ✕Wherever a tab sent it | ✕Nothing to scope | ✕No account exists | ✕Ungoverned already | Invisible, and carried anyway | Nobody |
The last two columns carry no verdict mark because they are descriptions, not judgments. Inference is billed to you by the cloud provider at list price.
Three columns in that table go to somebody else. Nothing beats an installed suite copilot on time to first use, nothing beats a self-hosted open model on keeping the weights inside your perimeter, and nothing beats a large firm on institutional familiarity. We win on the data boundary, on operator control, and on what you still hold if we part ways.
Each page below makes the fair case for the alternative first, including the conditions under which we would tell you to take it, and only then says where the difference lies. None of them names a company or a product. If a page reads as a hit piece, it has failed.
Read the page for the option you are closest to taking, not the one you like least. Then take the six questions into whichever meeting is next and ask them of us as hard as you ask them of anyone else. If the answers on this site do not survive that, the pages are wrong and we want to know. The posture questionnaire is published for the same reason: you should see the questions before the call, not during it.
Most organizations should. A bundled copilot for ordinary drafting and a governed deployment for anything sensitive is a sensible split, not a contradiction, and the two do not compete for the same work. What does not work is running both without deciding which content belongs where, because then the classification lives in each person's head.
Weeks. We quote no fixed number of days, because the pace is set by decisions you make rather than by us: the posture tier, the retention election, the first group of people to be put in front of it. The engagement path sets it out gate by gate, and the gates are the thing to hold us to.
It is a real risk and it has its own page. The mitigation is structural rather than reassuring: the deployment is in your account, you hold the credentials and the billing relationship, and infrastructure code and runbooks are deliverables. Read continuity before you decide.
No named references and no measured outcomes. What exists instead is a live demo of the running product, a published controls matrix with its standing exceptions visible, and an assessment whose deliverable you keep whether or not you proceed.
Every path on this page is already being taken by somebody in your market. The one being taken silently, by your own people on their own accounts, is the one with no owner.
A bounded engagement that maps your network topology and data flow, and tells you honestly which of these six paths fits. You keep the deliverable whether or not you proceed.