Give teams useful AI with a clear data boundary.

Why the chat window is not the hard part

Any team can install an AI tool in an afternoon. The hard part shows up in month two: someone pasted a customer contract into a cloud model, a departing employee still has an API key, nobody can say which model produced a document, and the one laptop with the indexed files needs a rebuild. Private AI for business is the set of decisions that prevent those outcomes, and most of them are organisational decisions that software can enforce but not make for you.

Three deployment shapes

Choose the smallest shape that meets your control requirements; you can grow into the next one without changing the desktop experience.

The controls administrators actually need

Vendors list dozens of enterprise features. In practice, four categories decide whether a deployment is safe to run: which destinations content may go to, who can access what and how that ends, what evidence exists afterwards, and how you recover. The table shows where each is enforced.

Enterprise controls ship across releases and some are still being completed. Write the controls you need into the pilot’s acceptance criteria and verify them on the version you deploy rather than relying on a feature list.

A 30-day pilot plan

Week one: pick one team, one repeatable workflow, and a set of representative test documents that contain no production secrets. Decide which models are acceptable and whether inference belongs on devices or a shared server. Install Glyph from the download page, verifying checksums, and, if you chose a server, stand up Glyph Private with Docker Compose.

Weeks two and three: run the real workflow. Measure answer quality against what the team does today, latency on the actual hardware, and the failure cases — wrong answers, slow answers, tasks that needed a cloud model. Exercise the operational path once: revoke a user, restore a backup, collect a support bundle.

Week four: compare results to the exit criteria you wrote in week one. Record the limitations and name the person who can accept them. Expand seats only when the process is repeatable without an engineer in the room. The readiness checklist is the owner-and-evidence version of this plan.

Cost and the hardware question

Local AI shifts cost from per-token fees to hardware and operations. A 16–32 GB laptop runs a capable mid-sized model with no marginal cost per request; a shared GPU host serves a larger model to a whole team. Cloud models remain the right tool for some tasks and can be permitted narrowly. Glyph pricing is per seat with a free tier for local chat; see pricing for individual plans and contact sales for organisational terms.

Support without surrendering data

Customer-controlled deployments keep infrastructure and data access with the customer. When something goes wrong, Fortaify works from redacted diagnostic bundles, logs that contain operational metadata rather than prompt or document text, and documented procedures. There is no vendor master password and no standing vendor account inside your environment — and that is a property to look for in any vendor, not only this one.

Questions to ask any private AI vendor

Use these in every evaluation, including of Glyph. Good answers are specific, name defaults, and point at documentation.

• Where does inference run by default, and what exactly changes when a user picks a hosted model?
• What leaves the device when the product is idle — telemetry, sync, licence checks — and can each be disabled?
• How is a departing employee’s access revoked, and how quickly does it take effect on their device?
• What evidence exists after the fact: which user, which model, which destination, when?
• What does support need from us, and does any vendor credential exist inside our environment?
• Which compliance claims are certified, which are implemented controls, and which are planned? Fortaify’s Trust Center answers this for Glyph.