By the Onsites AI team · Last updated · 4-minute read
Some desks arrive at self-hosting the way people arrive at homeownership: not because the apartment is failing, but because the requirements changed — a compliance office, a data-residency clause in the biggest contract yet, an AI policy that says the model must be ours, or a 40-seat bill whose arithmetic finally favors a license. The good news is that the move is deliberately unheroic: it is the same product changing address, and a well-planned migration has one week of parallel running and zero lost history. Here is the decision checklist, the sequence, and the organizational surprises worth front-loading.
The honest triggers are the ones no cloud product can contract away: a hard residency or sovereignty requirement; regulated workloads where the processor question must end with "there is no processor"; an air-gapped or offline network the desk must live inside; or economics that genuinely favor licensing — large steady seat counts where $15 per seat annualized beats metered cloud pricing, or heavy constant AI usage where owned compute beats per-action credits. The honest anti-triggers deserve equal billing: a three-seat team with light AI usage pays zero on the cloud and should stay there; self-hosting's minimum is a 10-seat annual license plus an operations owner (someone who runs the Docker deployment, the backup drills and the update calendar). If the move is "we want more control" without a named person to hold it, the cloud desk is already doing that control's job for free.
Day one of the plan is not a migration; it is a deployment. Stand the self-hosted instance up on your LAN or tenancy and give it a domain your team can reach (the deployment doc walks compose, storage and the license trial start — the trial is the full product, 60 days, nothing up front). Week one: connect the self-hosted instance to your own AI endpoint (BYO model) and verify features the pilot team will use. Week two: the history moves — exports from the cloud workspace flow in; conversations, CRM records, catalogs and documents land where the new desk will serve them, and spot-check the samples against the questionnaire-grade standards you presumably migrated for. Week three is the cutover itself: point the inbound channels — mail routing, widget on the production site, any connected messaging accounts — at the self-hosted URL, and run one overlap week where old and new desks both receive the same inbound mail while the team verifies parity on live traffic. Week four: decommission — archive the cloud workspace's final export, and cancel what is no longer used.
Technical cutover is the small half. On the cloud, three jobs were invisible: backups (paid tier), updates, and capacity. Self-hosted assigns each a name and a calendar — backup scripts on a schedule plus the quarterly drill, the monthly release window with staging, storage growth watched like any database. AI governance shifts too: the endpoint is yours, so the review habits and model policy are documented inside your own policy stack, typically alongside the compliance questionnaire that triggered the move. Budget conversations simplify in a different key: instead of metered seats-credits-storage, expect the license line plus hardware plus the operations hour — a trade the checklist article arithmetic makes concrete before you commit.
Take a ten-person team with a fast-growing attachment library to see when the license genuinely wins. On the cloud: three seats free, seven paid at $15 each — $105/month — plus a 50 GB storage pack at $50/month, plus AI credits that a copilot-heavy team burns at, say, 1,500–2,500 credits a month ($20 pack, often two). That lands near $180–200/month, plus the metered line item that grows with every attached PDF and exported CSV. On self-hosted: the 10-seat annual license floors at $150/month equivalent — and the two volumetric meters disappear by design, because storage is your hardware and AI is your model. The crossover is real but conditional: light-attachment, light-AI teams stay cheaper on the cloud; attachment-heavy, AI-heavy, seat-stable teams cross over, and the bigger the storage line becomes, the harder the license wins. Run your own numbers with the actual calculator before deciding — the point of this section is the shape of the crossover, not a prophecy.
A punch list that keeps week three boring: identity — same seat map as the cloud desk (admins first); data — history exports loaded and spot-checked against the original thread counts; channels — mail routing flipped and verified end-to-end (most real incidents live here: SPF/DKIM on your own domain now matter operationally, not just on paper); widget — production site snippet repointed; AI — endpoint configured with fallback, one test draft per feature class; runbooks — backup, restore and update procedures exist as files before week three, not after the first incident; rollback — the old cloud workspace stays warm (read-only) through the overlap week, because the cheapest rollback plan is the one you have not yet paid for. Teams that skip the last item buy their education at retail prices: every migration story that went badly traces to one of five skipped lines, and the overlap week is where each is found for free.
Doors open both ways, worth knowing before the first one closes. The same export discipline carries your data out of the self-hosted desk back to the cloud (or anywhere else), because exports are yours by design on every tier; the cloud workspace re-imports history the same way the self-hosted desk did. Nothing is a one-way trap on either side — which is the quieter, more important argument for trying self-hosting during the trial window at all: a 60-day no-payment pilot that ends with "we learned we prefer the cloud" is a cheap experiment, not a failure — the experiment is bounded, the data is portable, and the desk keeps serving customers on either side of it.
Why would a customer move from the cloud to self-hosted?
Usually one of four reasons: a hard data-residency or compliance requirement, regulated workloads where no third-party processor is allowed, the need to run inside an air-gapped network, or economics at scale — large steady seat counts and heavy AI usage where licensing plus owned compute beats metered pricing.
Can we run cloud and self-hosted in parallel during migration?
Yes, and you should: deploy self-hosted in week one, import history in week two, flip inbound channels in week three, and keep both receiving mail for one overlap week before decommissioning the cloud. One week of overlap catches every parity issue without customer-visible downtime.
What history can be moved from cloud to self-hosted?
The exports carry conversations, CRM records, product catalogs, quotes, orders, invoices and documents — enough to work from day one with full context. Spot-check thread counts against the original after import before cutover.
What organizational changes come with self-hosting?
Backups, updates and capacity become named responsibilities with calendars: nightly backup scripts plus quarterly restore drills, a monthly staging-tested upgrade window, storage watched, and an AI model policy documented inside your own compliance stack.
Is the move reversible?
Yes: the same export discipline that brought data in carries it back out — to the cloud or any other environment. Self-hosted starts as a 60-day trial with no payment up front, so a pilot that ends in preferring the cloud costs only the learning.