AI Process Redesign: Your Workflows Were Built Around Human Limits. Agents Don’t Have Them.
Walk through the operations floor of any lender and count the handoffs on a single loan file. Data entry. Credit check. First review. Second review above a dollar threshold. Documentation check. Settlement queue. Six or seven pairs of hands, each with its own form, its own queue and its own service target.
Ask why the process looks like that and you’ll get “that’s how it’s always been done.” Push a little and a better answer comes out. It looks like that because no one person could do all of it. Not well, not quickly, and not four hundred times a day.
That’s the part most AI programs skip. They take the workflow as given, find the slowest step, and put an agent on it. The step gets faster. The workflow keeps its shape. Eighteen months later the CFO is looking at a fifteen per cent improvement in one queue and asking why a technology that was meant to change everything hasn’t.
Every Process Is a Record of a Constraint
Batch approvals exist because approvers had to be gathered in one place. Handoffs exist because one person’s knowledge ended where another’s began. The four-eyes check exists because a single reviewer couldn’t hold a whole file in their head and stay objective about it. Queues exist because work arrived faster than people could clear it. Month-end close exists because you couldn’t reconcile continuously, so you stopped, gathered and counted.
None of that is bad design. It’s good design for the constraints of the time. A well-run operations team is a machine for moving work through people who each know part of the picture, and the process is the map of who knows what.
Remove the constraint and the process doesn’t become efficient. It becomes odd. An agent reads the entire file in seconds and then sits in a queue built to protect a human reviewer’s afternoon.
Why AI Doesn’t Force the Redesign the PC Did
Bill Gates published a long essay on the AI transition last week. Most of the coverage went to his proposal to tax AI tokens and robots. The more useful passage for anyone running a business is his account of why the personal computer took twenty years to change how people worked and why AI won’t. Software had to be written. Prices had to fall. People had to learn the tools.
AI runs on devices people already own, takes instructions in plain English and can learn from the same material you hand a new starter. His summary is that it doesn’t require people to adapt to it. It adapts to them.
For adoption, that’s good news. For process design, it’s a trap. Because AI fits so easily into the work as it stands, the path of least resistance is to leave the work as it stands. Copilot drafts in five minutes the memo that used to take an hour, so the memo still gets written, still goes to the same three people and still waits for the same Thursday meeting. The technology has adapted itself to a process designed for a world that didn’t have it.
The PC forced redesign because you couldn’t use it without changing how you worked. Typing pools didn’t survive word processors. AI forces nothing. Redesign has to be a decision someone makes.
“Don’t Automate a Broken Process” Is Only Half the Rule
Most leaders have heard some version of that advice. It’s sound as far as it goes, and it lets far too many processes off the hook. A process can be documented, compliant, well-run and hitting every target, and still be the wrong shape for an AI agent. Broken isn’t the test. The test is whether each step exists for a reason that still holds.
Run every step in the chain against one question: would this exist if a single tireless worker with perfect recall could see the whole file at once?
Some Steps Should Stay Human. Choose Them on Purpose.
This is where a second idea from the Gates essay is worth borrowing. He argues that some work should be designated “Human Reserved,” kept for people even when machines can do it, and that wherever the line falls it should be drawn deliberately rather than by default. He’s making a national policy argument. It works just as well inside a process map.
In financial services, certain steps should stay with a person for reasons that have nothing to do with capability. Telling a customer their loan has been declined. Deciding a hardship application. Signing off on a decision that a regulator may later ask a named executive to explain. FAR puts accountability on individuals, and an accountable person needs to have actually decided something, not waved through an AI agent’s recommendation at the speed the AI agent produced it.
Other steps should stay human because they build humans. The junior analyst’s first pass on a credit file is slow, and a senior reviewer could do it in a third of the time. It also happens to be how senior reviewers are made. Gates cites Stanford research showing employment falling among young workers in AI-exposed occupations while their older colleagues were unaffected. That’s a labour-market problem. Inside an operations team it’s a succession problem. Automate every entry-level step and in five years nobody in the building can do the senior review.
The distinction that matters is between a step you kept because nobody questioned it and a step you kept because you chose to. The first is a workaround. The second is a control, and you can explain it to a board or a regulator.
What Redesign Looks Like in Practice
It doesn’t look like a transformation program. It looks like one end-to-end process, mapped as it actually runs rather than as the procedure manual describes it. Every step gets a label:
- Capacity: it exists because people couldn’t get through the volume
- Knowledge boundary: it exists because the next person knew something this one didn’t
- Control: it exists to catch, check or authorise something
- Development: it exists because someone learns by doing it
Capacity and knowledge-boundary steps are candidates to collapse. Controls need a harder look: some still need a person, some only ever needed a record, and an agent that logs every action and decision often gives you a better record than a signature on a form. Development steps are the ones to protect, and to protect deliberately.
Then design from the outcome backwards, assuming one agent can hold the whole file, and place people where you’ve decided they belong. The result almost always has fewer steps, fewer queues, and more explicit decision points than the process it replaced. It’s also easier to govern, because you can point to each human step and say why it’s there.
On the Microsoft stack the tooling for this is already mature. Dynamics 365 agents can take a case from intake to a ready-to-decide state, Copilot Studio lets you build the ones Microsoft hasn’t, and Agent 365 gives you the identity and audit layer to run them under the same controls as your people. None of that is the hard part. The hard part is the decision to draw the map before you deploy anything.
The Typing Pool Problem
The organisations that got the most from the PC weren’t the ones that put a machine on every desk fastest. They were the ones that noticed the typing pool no longer needed to exist. A typing pool, for anyone who never saw one, was a room of typists who produced the letters and memos managers dictated to them. When word processors reached individual desks, the room emptied inside a decade. That was easy to spot, because the PC couldn’t work inside the old process.
AI can. It will run inside your current workflows without complaint, and it will make each of them a little faster. The equivalent of the typing pool won’t announce itself. It will be a process that runs well, hits its targets, and shouldn’t exist.
If you want a second pair of eyes on which of yours that is, get in touch with the 365 Mechanix team. We work with financial services organisations across Australia and New Zealand on exactly this kind of redesign, and we’re happy to start with a whiteboard rather than a proposal.
FAQs
Why do most AI projects deliver smaller gains than expected?
Because they automate a step inside a process built for human limits. The step gets faster, but the queues, handoffs and approvals around it stay, and they were the real source of elapsed time. The gain is real but bounded by the shape of the workflow.
What’s the difference between automating a process and redesigning it?
Automation keeps the steps and speeds one or more of them up. Redesign asks which steps would exist if one agent could see the whole file at once, removes the ones that wouldn’t, and places people deliberately at the points where judgement, accountability or development require them.
Which process steps should stay with a person?
Steps where accountability has to sit with a named individual, steps that involve a significant customer outcome such as declining credit or deciding hardship, and steps that develop the judgement your senior people will need later. The important thing is that each one is a conscious choice rather than a leftover.
What did Bill Gates’ AI essay say that’s relevant to business processes?
Two things. First, that AI adapts to people rather than requiring people to adapt to it, which is why it slots into existing workflows without forcing them to change. Second, that some work should be “Human Reserved” by deliberate choice. Both ideas apply directly to how organisations decide which steps in a process to automate and which to keep.
Where should we start?
One end-to-end process with a measurable customer or financial outcome. Map it as it actually runs, label each step by why it exists, and redesign from the outcome backwards. Hardship, onboarding, claims and month-end close are common first candidates in financial services.
How does 365 Mechanix help?
We work with organisations across ANZ to redesign processes around AI agents on the Microsoft stack, from the mapping exercise through to build and governance on Dynamics 365, Copilot Studio and Agent 365. If you want to work through which of your processes are candidates, get in touch.