The shape exists because the expensive failures in this work are decided early. A system built against the wrong workflow is a good build of the wrong thing, and no amount of engineering later recovers it.
Stage 1: the audit
A structured review of how work moves through your organization: which systems are involved, where the handoffs are, where your best people spend time on coordination, and which workflows would repay automating.
It produces a ranked list with the expected return on each candidate, and an explicit note on what we would leave alone. The report is yours to keep and act on with anyone, including your own team or a cheaper builder. What separates the kinds of firm selling this work is worth reading before you commit to any of them.
What we read during it:
- Your history, not your process documents
- Six months of resolved tickets, closed invoices or won deals. The documents describe the intended path. The history shows the exceptions, and the exceptions are what drive the build.
- The systems and how they connect
- What holds the record of truth, what writes to it, and which handoffs are a person copying between two screens.
- Where policy lives
- Whether the rule for an exception is written down or held in someone's judgment. This one difference routinely moves a build by weeks, and it is the second largest driver of what an engagement costs.
Stage 2: the build
We build the agent system end to end: the agent itself, the integrations with the tools you already run, the human approval checkpoints, and the monitoring. All four are part of the build, so there is no later phase where the system finally becomes safe to use.
You see a working version every week. Weekly review exists so disagreements surface while they are cheap, and what your team pushes back on in week 3 is the highest-value signal in the project.
Stage 3: deployment
Autonomy is granted per workflow against a bar your team sets, so a bad week stays contained to one workflow.
Shadow
The agent reads live work and drafts what it would do. Nothing sends. You get a real accuracy number on your own data before committing to anything.
Supervised
The agent acts on every run and a person approves before it lands. Approval takes seconds against writing from scratch, so throughput improves here even with a human in every loop.
Autonomous, by workflow
Turned on for the workflow that scored highest in shadow mode, while the others stay supervised. There is no single switch for the whole system.
Standing review
A weekly sample read by a person, and a monthly revisit of where the approval boundaries sit. An agent left alone for 6 months drifts.
Stage 4: operating it
Models change, workflows evolve, edge cases surface that no audit would have found. Operating covers monitoring, those changes, and the review of approval boundaries. It is quoted separately from the build so you can also choose to take it in-house once the pattern is established.
Most clients expand into adjacent workflows from here, which is scoped on its own terms rather than assumed. The agent patterns are where most of those start.
What your team owns throughout: a few hours a week during the audit, a weekly demo during the build, and one named person who can approve a policy decision inside a day. Committee approval is the most common reason a 6 week build becomes a 10 week build. More on that in the pricing drivers.
Keep reading
Buying AI
How AI agent pricing works
Why AI agent work is scoped after an audit, the 6 things that move the number, what you can do to lower it, and what to budget for after launch.
Agents
What our agents do
The agent patterns that hold up in production: front desk, revenue operations, support, and back office. What each takes over and where each breaks.
Last updated August 12, 2026. Questions this page did not answer go to team@agentintegrator.io, or browse the rest of the guides.
