Operations
21%
of enterprises have a mature model for what an agent may do without a person.
Deloitte, State of AI in the Enterprise 20261Most delayed decisions are incomplete files, not hard judgment.
Approvals sit in inboxes because the packet is missing, the policy is tribal, or nobody owns the exception. A model that scores the case does not help if a person still has to rebuild the file.
How I see it
Automate the routing, not the judgment you cannot write down
Decision work looks like expertise. A large share of the calendar is routing: which queue, which policy, which missing document, which person can actually say yes. That routing is automatable once the policy can be written down.
The design error is to treat every decision as too important for software, or every decision as safe to unsupervised. Some fields can be assembled and some actions can be taken inside a threshold. Some cannot. Name the boundary.
Credit exceptions, claims routing, order holds, discount approvals, and guest-recovery cases all follow the same shape. The completed result is a decision with a reason, written back to the system of record, with a human on the cases that change money, risk, or a customer commitment.
Do not start with an enterprise decisioning platform. Start with one approval type that already has volume, a policy someone can state, and an owner who will retire the inbox path when the new one works.
Where the cost sits
Decision costvolume of one approval type × hours of wait and reconstruction × loaded cost, plus error cost when the decision is late or wrong
Use the labor and wait terms if error cost is thin. Do not invent a risk model to justify the project.
Is this your problem?
Five signals the cost is already real.
Which approval type sits the longest, and why?
If the delay is a missing packet, automate assembly. If it is true judgment, keep the person.Can someone write the policy for the routine cases this week?
If not, you have a management problem, not a model problem.What action is allowed without a person, and what is not?
If you cannot name the boundary, do not put software on the decision.Does the decision write back to the system of record?
A recommendation in chat is not an operating change.Is there an owner who will retire the old approval path?
Two paths means the new one will be used for clean cases only.
What done looks like
The result, not the category.
A solution is finished when a named result no longer waits on reconstruction.
Complete the packet
Assemble the missing invoice, approval, ID, or exception note so the case is ready to act.
A person no longer rebuilds the file from email and shared drives.
The system produces a summary someone still has to reconstruct.
Write it back to the system of record
The result lands in the ERP, HRIS, or operations system. A side spreadsheet is not done.
The official system is current without a paste step.
Someone still re-keys the answer.
Retire the unofficial queue
Inbox, chase, and the person who knows which system is lying stop being the path.
Exceptions have an owner and the old side path is closed.
The bot is live and the spreadsheet is still the truth.
What usually fails
How solution shopping wastes a quarter.
Scoring cases that nobody works differently
A prettier priority list that uses the same inbox is not automation.
Leaving the policy in people's heads
If the agent cannot see the current rule, it will invent one. Write the policy down first.
Removing the control point to get a demo number
Unsupervised action on money or risk will be rolled back the first time it is wrong.
This is not the problem if
Do not start with a category purchase.
- The policy cannot be written down and nobody will try.
- Leadership wants unsupervised decisions on money from day one.
- The volume is a handful of unique cases a month.
- A clearer RACI or a shorter form would remove most of the delay.
A useful next step
Bring one expensive problem. 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