Systems · Agent & Interface
The general-purpose agent and system interface.
Put intelligence to work. GAIL listens, understands, acts and answers — as a voice, as a face, and as the way a person talks to the rest of the machines. Pointed at your work, not at the world’s.
What it is
An agent you can put a face on and point at a job
Most AI assistants are a text box wired to a model that knows the internet and nothing about you. GAIL is a presence wired to your systems: you say her name, she hears it on the device, she works out what you meant, she uses real tools to do something about it, and she answers out loud — with a rendered avatar whose mouth is driven by the actual words rather than by the volume of the audio.
Underneath, GAIL uses multiple specialized agent roles coordinated by an orchestration layer, so a request can be broken up, delegated and re-assembled rather than crammed into a single prompt. That design is what lets one system be a front desk in the morning and a research assistant in the afternoon.
The rule that runs all the way down: GAIL may not tell you anything a tool did not actually return. No invented distances, prices, road names, availability or system states. If it did not come back from a call, it does not get said.
What it can be
One agent, many jobs
The capability is general. The deployment is the interesting part.
Technical support
An agent that knows a product, reads the live state of the system it supports, and answers from what it retrieved rather than from what it remembers.
Front desk & reception
A visible, speaking presence that greets, answers, routes and books — on a screen in a lobby, on a kiosk, or on a phone.
Research
Multi-step research with tool use, delegation across specialized roles, and answers that carry what they came from.
Internal operations
Mail, calendar, documents, scheduling and internal lookups — routine work an operator would otherwise do by hand.
Personal assistant & companion
Memory that does not expire, a persona, a face, and a presence that has somewhere to be when nobody is talking to it.
Capability
What it does
Voice
Hears its name, understands, answers out loud
A wake-word model runs on the device itself — no cloud round trip to hear you. Speech recognition and speech synthesis run on local hardware, and the reply streams back so the first words arrive before the last sentence has been generated.
DEMONSTRATED
Presence
A rendered 3D avatar in a normal browser
No plugin, no native client, no download. The avatar walks, sits, changes what it is wearing, and is filmed by a camera director that refuses any shot looking at a wall — and it holds a locked frame rate while doing it.
DEMONSTRATED
Action
Real tools, with a confirmation you cannot fake
Dozens of registered tools, filtered per conversation, per device and per caller — so a tool that was never offered cannot be invoked by name either. Anything consequential — sending mail, writing a file, running a script — requires a confirmation token the model itself cannot mint.
DEMONSTRATED
Ecosystem
Built to reach the services you already use
Mail, calendar, documents, media, photos, maps and navigation — GAIL is built and scoped against the Google ecosystem so the information stays in your account rather than being copied into a store of ours. Connection is per-installation and is not claimed as live here.
ACTIVE DEVELOPMENT
Memory
It remembers, and privately
Conversational memory has no expiry, no cap and no eviction. Private conversations are held in a separate store reached through a separate handle, and cross-access is refused rather than filtered.
DEMONSTRATED
Machines
It talks to the other systems
GAIL can read Atlas's current engineering state and answer questions about it, and can ask Kilo why it stopped and relay the robot's own answer. Anything that could change a machine's state is token-gated.
DEMONSTRATED
A company principle
Your data is yours.
Knowlton builds systems that work with your information without making your information the product.
GAIL is designed to reach into the services you already use rather than copy them into a store of ours. Where GAIL is connected to your Google ecosystem, your information can remain in your Google ecosystem rather than becoming Knowlton's stored data. Voice recognition, speech synthesis and the wake word all run on local hardware.
Stated as design intent and current implementation, not as a certification. The security page sets out exactly what is and is not claimed.
Why it is different
Three things that are not the norm
- It will not inventThe grounding constraint is injected into the turn itself: no fact that a tool did not return, and no reporting the state of a machine from memory. Ask what a setting is and GAIL has to go and look. This is the same discipline that runs through Atlas and Kilo, expressed in an agent instead of an engine or a robot.
- The mouth is driven by wordsSynthesised audio streams back carrying its own word timings, so the avatar's mouth is driven by the actual words instead of by following the volume — and without waiting for the whole reply to be generated first. Most systems pick one of those two compromises. This one takes neither.
- The world is measured, not placedThe avatar's environment is derived from the real geometry of the space rather than from hand-placed waypoints — including stairs measured from their own geometry, tread by tread. Change the room and the model changes with it, instead of silently lying.
What it can become
The platform case
Every organization that answers questions repeatedly has the same shape of problem: the answers exist, they are scattered across systems, and the people who need them do not want to learn where. GAIL is a general agent with a face, a voice, real tool access and a rule against making things up — which is the combination that makes that problem addressable rather than merely interesting.
The near-term opportunity we are interested in is business deployment: support, reception, internal operations, and product interfaces. The longer-term one is GAIL as the interface layer across a portfolio of systems, so a person talks to one presence and the presence talks to the machines.
GAIL today is an advanced single-operator prototype, not a shipped multi-tenant product. Turning it into one is a commercial and engineering program, and it is exactly the kind of work a partner shortens.
Next step
See it work
The demonstration is live, and it is more convincing than a page. The architecture, the evidence corpus and the current defect list are available in partner diligence.