How we deploy
From workflow to working agents.
We work alongside your team to audit, redesign, deploy, and improve one workflow at a time.
Eight stages, from the first conversation to continuous operations.
- DiscoverDiagnoseRedesignSandboxShadowSupervisedGoverned productionImprove
The path
Eight stages, from the first conversation to continuous operations.
The first three stages change how the work is understood and designed. The next four change what the agents are allowed to do. The last one keeps going for as long as the system runs.
Stages describe the rollout. Levels of Agency describe how much authority an agent holds at each point. An agent in the sandbox or in shadow holds no authority to act in production, and a supervised agent can act only with approval.
Discover. Show us where work gets stuck.
We start with the business operation, not with ideas about AI. In an executive conversation with the COO, CFO, CRO, head of operations or whoever owns the function, we look for the places where work stalls: manual coordination, system switching, decisions that repeat, exceptions, approvals, reconciliation, reporting and handoffs.
Result A shared view of where the friction is, and which function to look at first.
Diagnose. Map how the work really happens.
The Agentic Operations Audit follows the work through your systems and teams: workflows, handoffs, decisions, exceptions and controls. Every step is classified as deterministic, agentic or human, and the findings become an executable plan with a business case.
Result An operation map, a classified list of opportunities, a plan with 30, 60 and 90 day and six month horizons, and a business case.
About the auditRedesign. Fix the process before you automate it.
We redraw the workflow before we build anything. Steps that exist only because of old workarounds go. What remains is designed around the boundary between deterministic, agentic and human work.
Result A redesigned workflow, and an agent system designed around it: which agents, which tools, which permissions and policies, and where people approve.
How we redesignSandbox. Prove it where nothing can go wrong.
The agents run in a controlled environment against test or historical data, with no way to change anything in production. We test the workflow, the permissions and the failure modes before anyone depends on the system.
Result Evidence that the system handles the ordinary cases and the awkward ones, and that its controls work.
Shadow. Work alongside your team and touch nothing.
The agent processes live cases next to your team. It sees the same inputs and forms its own answer, but nothing executes. We compare its answers with what people actually did, and we review every difference.
Result A measured picture of where the agent agrees with your team, where it does not, and why.
Supervised. Act, with a person approving first.
The agent starts to act, and a person approves each consequential action before it happens. Approvers see the evidence, the policy result and the agent's confidence before they decide. Every approval and every override is recorded.
Result A track record of approved actions, and a clear record of the cases people changed.
Governed production. Authority widens as the evidence builds.
Once the record supports it, high-confidence actions that comply with policy execute on their own, and uncertain or exceptional cases escalate to a person. Over time, approved parts of the workflow run within defined permissions, policies and controls.
Result Agents operating approved parts of the workflow inside limits you set, with every action on the record.
Improve. Measure, tune and extend.
We keep operating the system and measuring it. We tune what is not working, decide with you when an agent has earned more authority, and extend to the workflows next to it and then across the department.
Result A system that gets better with use, and a wider operation running on the same controls.
Redesign
Fix the process before you automate it.
Most organizations do not need another AI tool. They need to know which parts of the operation should stay deterministic, which need human judgment and which can safely become agentic. We map that boundary before we build anything.
- 01Map. Follow the work as it happens, not as the procedure describes it: elapsed time, touch time, exceptions, handoffs, owners and the systems involved.
- 02Simplify. Remove the steps that exist only because of a workaround, a missing integration or an approval nobody reads. We do not automate a broken process.
- 03Classify. Sort each remaining step as deterministic, agentic or human.
- 04Architect. Design the agent system around that boundary: which agents, which systems and tools, which permissions and policies, which approval points, and which evidence each action carries.
- 05Agentify. Build the agentic steps and deploy them into the systems you already run, starting in the sandbox.
Three kinds of work
Every step is one of three kinds.
The boundary decides who or what runs each step.
Deterministic: conventional automation
Rules and conventional automation. If a fixed rule can decide the step every time, a fixed rule runs it.
Agentic: an agent, inside controls
Contextual judgment, adaptive execution, research and orchestration. The step needs information gathered from several places, or a decision that varies with the case, inside limits we can define.
Human: a person
High-accountability, strategic, sensitive or not yet bounded judgment. People decide. Agents may gather the evidence.
Levels of Agency
Autonomy should be earned.
No workflow goes from manual to unrestricted overnight. Every agent holds a defined level of authority, and it moves up only when the evidence supports it.
Every agent’s record holds
- Current level
- Where the agent operates today.
- Target level
- Where we and the workflow owner expect it to get to.
- Requirements to progress
- What has to be true before its authority increases.
- Evidence required
- What we measure and show to prove it.
More autonomy is granted only when the evidence supports it.
An agent can hold a different level in each workflow. The level says what the agent may do, and the deployment stage says how far the rollout has come.
The agent watches the workflow and builds an understanding of it without affecting production.
- Authority
- Read access to the systems and records the workflow uses. No ability to change anything.
- People
- People do all of the work. What the agent observes goes to the team building the system.
- Controls
- Read-only access, scoped to the systems in the workflow
- No permission to write, send or approve
- Everything the agent reads is logged
- To move up
- Its understanding of the workflow is confirmed by the people who run it
- The source systems are mapped and access is agreed
- A baseline of today's performance exists
- Evidence
- Observations that line up with what people did on the same cases.
The agent proposes decisions and actions while people remain responsible for carrying them out.
- Authority
- Read access, plus the ability to draft a recommendation and put it in front of a person. It cannot act on it.
- People
- People review each recommendation, then act on it or override it.
- Controls
- Read-only access to the systems of record
- Every recommendation carries its evidence and the result of the policy check
- Each recommendation, acceptance and override is recorded
- To move up
- Recommendations are accepted without changes for the cases in scope
- The actions the agent would take are defined, scoped and permissioned
- Approvers are named
- Evidence
- The record of recommendations against outcomes, and the reasons people overrode them.
The agent prepares actions and carries them out only after a person explicitly approves.
- Authority
- Read access, and write access for the specific actions it has been cleared to prepare. Each one waits for approval.
- People
- Approvers decide on each consequential action, with the evidence in front of them.
- Controls
- Write access limited to named actions in named systems
- Approval by a named role before anything executes
- The approver sees the evidence, the policy result and the confidence
- Approvals and overrides are written to the audit record
- To move up
- Approvals for a class of cases are consistently given without changes
- Confidence scores are calibrated: high confidence goes with correct outcomes
- The workflow owner has agreed the thresholds
- Evidence
- The approval history, the reasons for overrides and the calibration results.
High-confidence actions that comply with policy execute automatically. Uncertain or exceptional cases escalate.
- Authority
- Write access to approved actions, used alone only when confidence clears the threshold and every policy check passes.
- People
- People take every escalation and review a sample of the automatic actions. They still approve anything policy reserves for them.
- Controls
- A confidence threshold set for each type of action
- Policy checks before every execution
- Anything below the threshold or outside policy escalates with the evidence attached
- Ongoing review of the actions taken automatically
- To move up
- Performance holds across the class of cases over a sustained period
- Exceptions and escalations are understood and handled
- Recovery has been tested
- Evidence
- Exception, escalation and intervention rates over time, and a policy record with no unresolved violations.
Agents independently operate approved parts of the workflow within defined permissions, policies and controls.
- Authority
- Independent operation of approved portions of the workflow, inside the permissions and policies set for them.
- People
- People own the workflow, set the limits, review performance and take every case the agent escalates.
- Controls
- The same permissions and policies as at lower levels, unchanged as autonomy grows
- Continuous monitoring of cost, latency, exceptions and policy violations
- Escalation paths to named owners
- Recovery procedures ready, and an audit record for every action
- To stay here
- Performance stays inside the limits the workflow owner set
- Policy violations are investigated and closed
- Authority is reviewed against the evidence, and it can be reduced as well as increased
- Evidence
- Sustained performance against the agreed limits, and a complete record.
Controls
Governed from the first day.
At every level, every consequential action sits inside the same eight controls. What changes as an agent earns authority is what it is allowed to do inside them.
Evidence
Material actions carry their inputs and the records that support them.
Policy
Written rules decide what is allowed, and every check is recorded.
Permission
Access is explicit, set for each system and each tool. An agent may know what to do without having permission to do it.
Confidence
The agent's confidence, measured against a threshold, decides whether it acts, recommends or escalates.
Approval
Where the risk calls for it, a person approves before anything executes.
Escalation
Cases outside policy, or below the threshold, go to a named person with the evidence attached.
Audit
Every step is written to a record: what happened, when, under which policy, and who or what did it.
Recovery
Actions are observable and built for recovery, so a wrong step can be found, contained and corrected.
Your systems
We deploy across the systems you already run on.
You do not migrate systems or add another dashboard for your team to watch. Agents work in your ERP, email, chat, spreadsheets and document stores, and approvals reach people where they already work.
ERP
CRM
Email
Slack or Teams
Banking systems
Shared drives
Spreadsheets
Document systems
Warehouse and planning systems
Databases and data warehouses
After go-live
Once it is running, we keep measuring.
We stay involved after go-live. We operate the system, measure it, and decide with you when it has earned more authority.
- Success rate
- Cases completed correctly, start to finish.
- Exception rate
- Cases that fell outside what the agent could resolve within policy.
- Escalation rate
- Cases the agent handed to a person.
- Human intervention rate
- Cases where a person changed, overrode or completed the agent's work.
- Confidence calibration
- Whether high confidence actually goes with correct outcomes.
- Policy violations
- Actions that broke a rule, and how each one was resolved.
- Recovery events
- Times something went wrong and had to be reversed or corrected.
- Cost per completed workflow
- What each finished case costs to run.
- Latency
- How long a case takes from trigger to completion.
- Business KPI impact
- Movement in the number the workflow exists to change, measured against the baseline from the audit.
Expansion
From one workflow to an agentic operation.
We begin with one constrained workflow. When it has earned its authority, we add the workflows next to it, then the rest of the department, then the connections between departments.
- One workflowA constrained problem with a clear owner and a baseline we can measure.
- Adjacent workflowsThe steps upstream and downstream, on the same connections and controls.
- The departmentMost of the function's operational work running as one agent system.
- Across departmentsAn operating layer that spans finance, revenue, operations and risk.
The governance stays the same as the scope grows: the same controls, the same record and the same rules for earning authority.
FAQ
Common questions
Discovery starts with an executive sponsor: the COO, CFO, CRO, head of operations or whoever owns the function. The audit and the redesign then work with the people who do the work day to day, and with whoever can open up the systems involved.
No. Deployments run on the software you already use. If the audit finds a broken integration, it goes into the plan. Replacing your systems of record is not part of the approach.
Only once it has earned that. Until then, people approve. Even at the highest level, an agent operates inside permissions, policies and controls, and anything outside them escalates to a person.
We agree the requirements for each step up with the workflow owner before the agent starts. Then we review the evidence together: results against the baseline, the record of overrides, confidence calibration, exceptions and policy violations.
Then the audit says so. Some work is better run by fixed rules, and some should stay with people. Knowing which is part of the job.
For every consequential action you can see the input, the evidence, the policy checks, the confidence, the approval and the action taken, all in the audit record. That is a decision trace. It is not a window into a model's private reasoning, and we do not present it as one.
It depends on the workflow, the systems and the risk. The audit ends with a sequenced plan with 30, 60 and 90 day and six month horizons, so you know what to expect before you commit to a deployment.