Workflow
Most AI automation projects fail because they automate tasks without redesigning the workflow.
Drafting a document or summarizing a ticket is useful. It does not change a process that still waits on handoffs, queues, and someone to update the system of record.
2.8×
more likely that AI high performers report fundamental workflow redesign.
McKinsey, The State of AI in 20251
How I see it
Automate the workflow, not the leftover tasks
Workflow automation starts with the completed result, then asks which steps still need a person. The rest can be a rule, a model, or an agent. The design error is to sprinkle AI onto every existing step and call the process automated.
The valuable work is subtraction. If a handoff exists because two teams could not share a queue, fix the queue. If a review exists because the packet was incomplete, complete the packet. If a person copies fields between systems, integrate the systems or let software write back.
AI belongs where the step requires interpretation: reading a document, classifying an exception, drafting a next action, or assembling context from more than one source. Rules still belong where the logic is stable.
A production automation has a baseline, an owner, a definition of done, and a way to see when the new path is worse than the old one. Without those, you have a demo that will be abandoned when the first ugly case arrives.
How the engagement runs
What you actually buy.
A service is a sequence with an artifact at each step. It is not a transformation program you purchase as a bundle.
Map the unofficial path
Inbox, spreadsheet, chase, and the person who unsticks the case — not the SOP.
Cost per case and the real handoffs are visible.
A clean process map nobody recognizes.
Decide how far to change the work
Improve the current path, redesign it, or give one outcome to an agent.
The path matches the economics and the risk.
Automating leftover tasks on the old path.
Put people on the control points
Judgment, authority, and risk acceptance stay human. The rest can move.
Named steps are retired after the new path holds.
Every step still needs a person, so nothing got cheaper.
Do not buy this for
These engagements fail for predictable reasons.
Automating the official path only
The exceptions are the work. If the new path cannot handle them, people will keep the old path.
Leaving the system of record to a human
If someone still has to paste the result into the CRM or ERP, you automated a side task.
Calling a copilot a workflow change
A writing assistant can help a person. It does not remove the queue.
Before we start
Questions that decide whether to engage.
What is the completed result, in one sentence?
If you cannot name it, you are automating activity.Which steps would disappear if the packet arrived complete and the system updated itself?
Those steps are the redesign, not the model.Where must a human still decide, approve, or accept risk?
Design those control points on purpose. Do not leave them implicit.How will you know the new path is worse than the old one?
If there is no comparison, you cannot operate the automation.Who owns the workflow after the project team leaves?
If the answer is IT and the business, nobody owns it.
How we size it
Automation valuecases per year × minutes removed per case × loaded cost, plus error cost avoided, minus the cost to run and supervise the new path
If supervision and exception handling consume the savings, the design is not finished.
This is the wrong engagement if
Do not start this engagement yet.
- The process is rare, highly political, or depends on relationships that should stay human.
- There is no system of record and no owner who can retire old steps.
- A checklist or a better form would remove most of the cost.
- The company wants a bot on the current process and will not change the process.
A useful next step
Bring one expensive workflow. Get guided into production.
We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.
Discuss an AI opportunity