Help users prepare replies and find information, or give your team a way to classify and route incoming work. We build agents around your data, tools, and approval rules.
Assistants in your product that prepare drafts, summarize records, and answer questions using sources the user can access.
Operations agents
Agents that classify tickets, route requests, or reconcile records within agreed permissions. Missing information and cases outside the rules go to a person.
Voice and conversational interfaces
Voice or text interfaces for tasks where people need to ask questions and consult information without navigating forms.
Supervision built in
Approval rules, evaluation cases, and logs defined with your team, including who reviews exceptions and when an action must stop.
Proof
Frankie, Amgen’s voice assistant
We built Frankie so Amgen’s field sales teams could ask questions and consult company data by voice on iOS. The project predates today’s LLM assistants.
Sophilabs presented us with a product that we were proud to show people.
A support representative uses the assistant to prepare a reply and checks it before sending.
Input
A customer asks what their plan includes.
Proposal
The assistant prepares a reply using the plan details and shows its sources.
Review
The representative checks the reply against the cited terms and edits it if needed.
Action
The reply is sent only after the representative approves it.
An agent in your operations
An agent helps sort incoming requests according to the team’s routing rules, with exceptions sent for review.
Input
A request arrives in a shared inbox.
Proposal
The agent suggests a category and a responsible team using the routing rules.
Review
Missing details or cases the rules do not cover go to a person.
Action
The agent records or routes the request only within its permitted actions. Other actions await approval.
The workflow determines which approach fits. We agree the sources, permitted actions, and review requirements before building.
When teams come to us
01
Support demand is outgrowing the team.
People spend time sorting requests and looking up the same information. An assistant can help prepare replies for review.
02
Users need help taking the next step.
They know what they want to do but have to navigate several screens. A copilot can help them find the right action and prepare it.
03
Your pilot stalls before release.
The demo works, but exceptions, permissions and inconsistent results still need attention. We help identify what to test and fix.
How it starts
Discovery examines the workflow, data, and integrations to help choose a first use case. With a defined scope, you can start directly with implementation.
How we build it
01
Map the workflow
Identify the task, the people involved, and the decisions that need approval.
02
Connect the tools
Define data sources, permissions, and integration requirements.
03
Test the first workflow
Evaluate representative cases and failures with your reviewers before release.
04
Review the next scope
Your team decides whether to change permitted actions, using evaluation results and the consequences of an error.
What “supervised” means
We define which actions need approval, who reviews them, and how corrections feed into instructions and tests. Your team decides whether to expand autonomy for each type of action, based on evaluation results and the consequences of an error. Fewer corrections alone do not justify removing oversight.
Grounding
Define the sources it can use, how to check its answers, and what to do when information is missing.
Scope
Document allowed systems and actions, with access enforced in the application and approvals agreed with your team.
Evaluation
Test representative cases and failures before releasing changes to prompts, models, or data.
Audit trail
Record the information needed to investigate outputs and actions, with agreed access and retention rules.
Monitoring and escalation procedures define who responds when results fall outside the agreed criteria.
What an engagement looks like
We start by agreeing on a workflow and acceptance criteria. The AI Discovery Sprint helps define that scope; you can also bring a defined use case directly to implementation. We estimate the first release after reviewing data, integrations, and approval requirements. The work includes the agent, evaluation cases, operating documentation, and instructions refined during development. We agree who will run and maintain it before launch.
Explore where an agent could help and define a first scope.
Frequently Asked Questions
Will it hallucinate in front of our users?
It can produce incorrect answers, even when it uses your data. We test those failures on your cases, make sources available for checking, and define when a response needs review. Monitoring after launch helps identify problems; these controls do not eliminate the risk.
Can it act inside our existing tools?
Yes, where the tools provide the access and integrations needed. We check those requirements, define permitted actions, and agree approval rules before building. Some systems may need an integration or data preparation first.
How much autonomy should we give it?
We agree approval requirements for each type of action with your team. The level of autonomy depends on evaluation results and the consequences of an error; some actions may continue to need human approval.
Which model does it use?
We choose the model based on the task, data restrictions, quality, latency, and operating cost. It sits behind an interface your product owns. Changing it can require adjustments to prompts, tool integrations, and infrastructure, followed by evaluation on your cases before release.
Does our data leave our environment when an agent runs?
That depends on the deployment. We document which data each service receives, where it is processed and stored, and the applicable retention and training terms. If your data must stay inside your environment, that restriction guides the model and infrastructure choices.
Who maintains it after launch?
We agree ownership, monitoring, support, and handover before launch. Your team can take over with the code, evaluation cases, and operating documentation. If we continue maintaining or developing the system, the agreement defines the scope, responsibilities, and price.