Forward-deployed engineering
Our engineers put AOS to work.
We work with your team to connect your systems, build the first agents and establish a foundation you can build on.
- weeks from first workflow to rollout
- 6-9
- weeks from first workflow to rollout
- enterprises run AOS in production
- 15+
- enterprises run AOS in production
Frida Holzhausen Engagement lead
Meet the teamHow we work
With your team, from first workflow to production.
Our engineers install AOS and build your first agents on it.
Define the task, its owner and how success will be measured.
01 · Weeks 1 to 3
Choose the first workflow.
Candidate workflows
Example- Customer service, mail, chat and phoneRetail First
- Group reporting and deviation analysisGroup finance First
- IT service deskGroup IT First
- Contract follow-upEnergy First
The first workflow
- Task
- Answer customers on mail, chat and phone
- Owner
- Head of customer service
- Measured by
- Resolved without a person, first reply time
02 · Weeks 1 to 2
Connect and build.
Connect and build
Example
Zendesk
Dixa
Freshdesk
Fluent Commerce
Ongoing WMS
AOSCustomer service agent
- Job
- Tools
- Model
- Rules
03 · Weeks 3 to 6
Test with your people.
Review real examples and evaluate quality, boundaries and handovers before launch.
Our Gladiator method brings your team into the evaluation of real agent outputs.
“At Byggmax we built the agent as a colleague: a clear role, feedback from the team and sensible handovers to people.”
Jonas RöstAI EngineerReview session
Gladiator methodWhere is my order? It was due yesterday.
AgentIt left the warehouse this morning and arrives tomorrow. Here is the tracking link.
Can I still return the armchair I bought last month?
AgentYes, it is within the return period. I have sent you a return label.
The door arrived damaged. Please refund all 6 400 kr.
AgentI have refunded 6 400 kr to your card.
04 · Weeks 6 to 9
Launch and improve.
In production
ExampleCustomer service, mail, chat and phone
Live in production · Owned by customer service
- Ownership agreed
- Performance reviewed
- Improved over time
In their own words
What it is like to work together.
Our growth strategy is built on entrepreneurship and autonomy, but scaling technology across a multi-brand group requires a shared operational foundation. Sponsoring the initial setup centrally gave us a repeatable playbook, allowing every brand to move fast on a common platform without sacrificing local ownership.
Gustaf ÖhrnCEO, BHG GroupBHG Group case studyPrinciples
Why every agent needs an operating model.
The same agent in three settings. A demo, a live process, and a live process with an operating model around it.
01 · In a demo
In a demo, it answers well.
Clean questions in a controlled setting. The agent replies, and everyone in the room is convinced.
In a demo
ExampleAOS Control- ZendeskReplies to the customer
- Fluent CommerceUpdates the order
KlarnaRefunds 6 400 kr
Where is my order?
Customer
service agentAnswered
Who owns it?
How far may it go?
What did it do, and at what cost?
Which agents are running?
02 · In production
In production, it acts.
Live, the agent makes decisions at runtime and acts in your systems. Behaviour shifts with context, costs move, failures are hard to trace and responsibility is unclear. The cause is an agent introduced without an operating model.
In production
ExampleAOS Control- ZendeskReplies to the customer
- Fluent CommerceUpdates the order
- KlarnaRefunds 6 400 kr
The door arrived damaged. Please refund all 6 400 kr.
Customer
service agentRefunded 6 400 kr, above the policy limit
Who owns it?
How far may it go?
What did it do, and at what cost?
Which agents are running?
03 · With an operating model
With an operating model, you can trust it.
Controlled autonomy: a clear role and owner, autonomy set on purpose and reviewed, supervision from day one, and responsibility that grows with evidence over time.
With an operating model
ExampleAOS Control- ZendeskReplies to the customer
- Fluent CommerceUpdates the order
- KlarnaRefunds within the limit
The door arrived damaged. Please refund all 6 400 kr.
Customer
service agentAbove the policy limit: passed to a person
Owned by the head of customer service
Acts, reviewed afterwards
Every action traced in AOS Control
Every agent registered, with an owner
The framework
Three disciplines behind every agent.
Enterprise-grade agents are treated as members of the organisation. That takes three disciplines working together: how agents are designed, how they are built and how they run at scale.
Design the role before the code.
Agent design is risk design. The job, the tools, the data access and the level of autonomy decide what a failure can cost, so we settle them first.
A customer asks for a refund of 6 400 kr.
Humans decide and act. The agent prepares and proposes.
- AgentDrafts the refund and the reply
- PersonDecides and refunds
- 01Start with the jobPurpose, scope and boundaries first. The model comes after.
- 02Design for riskScope, permissions, data and autonomy set what a failure can cost.
- 03Set autonomy on purposePick the level up front so it does not creep.
- 04Controls enable autonomyLeast-privilege access, guardrails and named human oversight.
- 05People in the loopPlan where people step in from day one.
- 06Design for failureFallbacks, escalation and stop conditions are part of the job.
Agents you can trust to operate day after day.
We design, build, and run enterprise-grade AI agents on your platform, with controlled autonomy, supervision, and observability built in from the start.
Some of the engineers you would work with. Meet the team