Start a conversation
HomeApproach
Our method

With-AIHow we design and deliver work that people do with AI.

Kollabio is built on one shift: from a person doing the work to a person getting work done with AI. With-AI is the method behind it. Human-centered design, AI engineering, Agile, and DevSecOps run as one loop, so none of them is a separate workstream. Inside it, Earned Autonomy sets how much the AI does on each task.

The methodWith-AIFour stages, repeated every increment, with all four practices at each one.The rule inside itEarned AutonomyAI takes on more of a task only when the evidence supports it.
Three commitments
01

People keep the decisions they're accountable for.

AI takes on tasks. Judgment that carries legal or mission accountability stays with a named person.

02

We test AI tasks before people rely on them.

Each task gets an evaluation suite with known answers before it reaches a user, and the suite grows with every correction.

03

Security and oversight ship in the first increment.

Logging, audit trails, and review screens are part of the first release, not a hardening phase at the end.

The loop

Four stages, repeated every increment.

All four practices contribute at every stage. What we learn in production feeds the next round of discovery.

Earned Autonomy

How an AI task earns its autonomy.

Each task starts at a level set in the Design stage and moves up only when evaluation results and override rates support it. Program owners and governance boards get a shared way to say how much the AI is doing, and pilot to production becomes a measured path.

Person AI Share of the work, illustrative

A task moves up a level only when evaluation results and override rates support it.

Decisions that carry legal or mission accountability stay at level 2 or below.

Today's example: the workforce-tool data assistant runs at level 1 in a sandbox.

Stage 01

Discover the work

Who does the work today, which tasks AI could take, and which decisions people must keep?

We start in the field with the people doing the work: interviews, on-site observation, and technical spikes on the hardest integration points. We map every task, then sort it into what people keep, what AI could take, and what we don't know yet.

  • Human-centered designField research with the people doing the work
  • AI engineeringData and feasibility check on what a model can do reliably
  • AgileBacklog built from the work map, ranked by hours saved
  • DevSecOpsData access, ATO path, and security constraints identified up front

Output

A work map with a person lane and an AI lane.

Work map
Illustrative
PersonReceive requestDecide eligibility AICheck recordsDraft summaryDraft notice

Each task sits in one lane. Decisions that carry legal or mission accountability, marked with a shield, stay in the person lane.

Handoff spec
Illustrative
TaskDraft case summary
AutonomyLevel 2 · Draft and approve
Person checksEvery summary, against the source record
If unsureBelow the confidence threshold, the person gets the record with no draft
LoggedDraft, edits, approver, and time

One spec per task. The review screen, the fallback rule, and the audit trail are designed together.

Stage 02Sets the starting Earned Autonomy level

Design the handoff

For each task, what does the AI do, what does the person check, and what happens when the AI is unsure?

For each task, we decide how much the AI does and how the person stays in control. That means a starting Earned Autonomy level, a review screen that shows the source next to the AI output, and a rule for what happens when the AI is unsure.

  • Human-centered designReview screens and plain-language explanations of AI output
  • AI engineeringConfidence thresholds and fallback rules for each task
  • AgilePick the thinnest slice that proves value
  • DevSecOpsLogging and audit trail designed with the handoff

Output

A starting Earned Autonomy level per task, review screens, and fallback rules.

Stage 03

Engineer with evidence

Can we show it works, safely, before people rely on it?

We build in two-week increments, with designers and engineers on the same team. Each AI task gets an evaluation suite with known answers before it reaches users, and that suite runs in the pipeline with the security and accessibility checks.

  • Human-centered designUsability tests on each increment, including Section 508
  • AI engineeringEvaluation suites with known answers and red-team tests
  • AgileTwo-week increments with metrics leadership can read
  • DevSecOpsSecure CI/CD, with the evaluation suite running in the pipeline

Output

A working increment with evaluation results.

Evaluation run
Illustrative
Known-answer suitePass
Red-team: prompt injectionPass
Red-team: data leakagePass
Section 508 checksPass
Usability test: review screenFix before release

The suite runs in the pipeline on every build, next to the security scans.

Monthly learning review
Illustrative
TaskOverridesDecision
Draft case summaryPropose level 3
Check recordsHold at level 2
Draft noticeRedesign review screen

Overrides feed the test set. A task moves up only when the evidence supports it.

Stage 04Decides which tasks have earned the next level

Evolve in production

Where do people override the AI, and what should change next?

After launch, we watch where people override or work around the AI. Each correction goes into the test set. A monthly review decides what to redesign and which tasks have earned the next level.

  • Human-centered designStudy where people work around the AI, then redesign
  • AI engineeringMonitor quality and drift; add each correction to the test set
  • AgileMonthly review decides which tasks have earned a new level
  • DevSecOpsContinuous monitoring, incident response, and rollback

Output

Monitoring, an override log, and a monthly learning review.

Six practices

Six practices, scheduled together at every stage.

Each practice contributes to every stage of the With-AI loop, so a pilot reaches its governance review with tests, an audit trail, and usability findings already in place.

Practice / stage
01
Discover the work
02
Design the handoff
03
Engineer with evidence
04
Evolve in production

Human-centered design

Field research with the people doing the work

Review screens and plain-language explanations of AI output

Usability tests on each increment, including Section 508

Study where people work around the AI, then redesign

Domain-driven design

Map the domain with the people who run it: terms, rules, and boundaries

Bounded contexts and a shared vocabulary for each workflow

Services and APIs drawn along domain boundaries

Update the model as rules and policy change

Continuous Agile delivery

Backlog built from the work map, ranked by hours saved

Pick the thinnest slice that proves value

Two-week increments with metrics leadership can read

Monthly review decides which tasks have earned a new level

Cloud and DevSecOps

Hosting, environments, and data access identified up front

Logging and audit trail designed with the handoff

Secure CI/CD, with the evaluation suite running in the pipeline

Monitoring, incident response, and rollback

Security and compliance

Security categorization, ATO path, and control requirements identified up front

Access controls, privacy, and audit requirements designed into the handoff

Controls implemented and documented as evidence for the authorization package

Continuous monitoring, annual assessments, and findings tracked to closure

Data architecture

Inventory of data sources, owners, and known quality issues

Data models and access rules for each workflow

Migrations and pipelines tested against known records

Monitor data quality and fix problems at the source

Where it applies

All four capabilities run on With-AI.

Ready to start

Let's design the first two weeks.

Most engagements start with a two-week Discover stage. Tell us about the program and we'll propose a focused scope.

Start a conversation See our capabilities