Human + Agent Workflow Redesign
Redesign a whole workflow step by step: what a rule should do, what an agent should do, and what stays with people, then measure the result against a baseline.
The business problem
Most AI pilots bolt a tool onto one step of a process that was designed for people. The step gets faster, the queue moves somewhere else, and the cost per outcome barely changes.
Real leverage comes from redesigning the workflow end to end: deciding who does each step, where the handoffs sit, and which approvals still make sense once preparation is no longer the bottleneck.
How it runs today
- The process grew one workaround at a time and is documented nowhere.
- Each step waits for a person, even when the step needs no judgment.
- Handoffs happen by email or chat, and work stalls between teams.
- Approvals exist because of a problem years ago, not because of current risk.
- Nobody has a baseline, so nobody can say whether a change helped.
Automate the work. Keep judgment human.
Plain automation where the logic is fixed. Agents where the input is messy. People wherever consequences or ambiguity rise.
Deterministic automation
- Steps with fixed logic: routing, calculations, lookups, notifications and status changes.
- Handoffs between systems where an API or file feed already exists.
- Enforcing the order of steps and the deadlines between them.
- Capturing the measurements for every case as it moves through the workflow.
Agents
- Preparation work: gathering inputs, reading documents and assembling the case for a decision.
- Drafting the outputs a person would otherwise write from scratch.
- Monitoring work in flight and chasing the next party before a deadline passes.
- Executing routine next steps within defined permissions, and escalating the rest.
People stay responsible for
- Judgment calls, approvals and exceptions the design reserves for people.
- Accountability for the outcome of the workflow, with a named owner.
- Relationship-heavy steps: clients, counterparties and negotiations.
- Deciding when the redesign is working, and when to roll back.
Systems and integration
- The system of record for the workflow (ERP, CRM, claims, matter or TMS)
- Document and file storage
- Email, Slack or Teams for handoffs
- Workflow, case or ticketing tools
- Reporting and BI for the baseline and results
Validation and controls
- A step-by-step map naming the actor for each step: system, automation, agent or human.
- Approval points set by risk and reversibility, not by habit.
- Agent permissions, stop conditions and escalation paths defined before go-live.
- A measured baseline before any change, and the same measures after.
- A tested rollback to the previous process.
What we measure
- Human touches per case
- End-to-end cycle time
- Errors and rework per 100 cases
- Cases completed per FTE
- Cost per completed outcome
An example workflow design
Illustrative, not a client result. It shows where each responsibility sits.
Commercial insurance submission intake
Trigger: A broker submission arrives with an application, loss runs and schedules.
An illustrative design showing how we would approach the workflow. Not a client result.
- 1AutomationCaptures the submission, creates the record and checks the broker and line of business against appetite rules.
- 2AgentExtracts insured details, exposures and loss history from the attachments and fills the submission record.
- 3AgentLists missing information and drafts the request to the broker; chases it if no reply arrives.
- 4AutomationValidates the completed record and scores it for priority using the underwriting team's rules.
- 5HumanThe underwriter reviews the prepared file, decides whether to quote and sets terms and price.
- 6SystemPolicy admin and the broker are updated, and each step's timing is logged against the baseline.
Controls
- Agent cannot decline, quote or bind
- Out-of-appetite submissions go to an underwriter, not auto-decline
- Low-confidence extractions are flagged on the file
- Full log of each step, actor and timing
What we would measure
- Underwriter hours per submission
- Time from receipt to quote-ready
- Submissions handled per underwriter
- Cost per quoted submission
Assess a workflow
Bring us one workflow that is slow, repetitive, error-prone or expensive. We'll help determine whether AI is actually the right lever.