How to Build a Next-Generation Workflow with Generative AI

01Map the work before selecting a model

The first step is not choosing a tool. Break the current process into inputs, processing, decisions, approvals, and outputs, and note where work waits, gets reworked, varies in quality, or depends on individual expertise. Size each step so that one line can say who looks at what and decides what.

Good candidates are frequent tasks whose input and expected output can be described and whose results a person can check. Ambiguous executive decisions, identity checks, and high-impact external responses are not fully automated from day one. Let AI prepare drafts or summarize the issues, and let an accountable owner make the final call.

Where that line falls is what separates results. In a PwC Japan survey conducted 12–19 February 2026 among 932 managers and above at companies with over 50 billion yen in revenue who were involved in generative AI adoption decisions, 87% reported active use or active promotion of generative AI. Yet the working time actually saved at companies that had deployed it averaged 75 minutes per day. At the far end of the range, Panasonic Connect cut 788,000 working hours in fiscal 2025, or 3.4% of its total. What separates those outcomes is not whether AI was adopted but how far the design was willing to hand work over.

Run each task through the checklist below. Questions 4 and 5 set the ceiling on delegation: if no one can judge whether the result is right, leave the task out; if a mistake cannot be reversed, stop at a draft. For how to run the inventory itself, see our guide to business process inventory.

• Establish a baseline for cycle time and rework
• Identify the source of truth and its owner
• Separate proposal, review, approval, and execution

02Split the workflow into testable AI tasks

When one long prompt tries to complete the whole job, you cannot see where it failed. Split the work into units whose results can be checked, such as classification, extraction, matching, drafting, and format checks. Each unit can then have its own pass criteria and tests, and fixes stay contained.

For each task, define the input format, expected output, prohibited behavior, and fallback. The fallback is the person or step that takes over when AI cannot decide. A task without a fallback stalls the moment an exception appears, and the backlog often grows unnoticed.

Keep writes to external systems behind a separate permission boundary from reading and drafting. Customer communication, payments, and access changes should only receive approved information. For the control questions to settle before using AI in accounting, see our checklist on AI and internal control.

Use one sheet per task. Make it a rule that no task goes live with the reviewer or fallback left blank. You can also have AI draft the breakdown and the sheets with the prompt that follows.

You are responsible for workflow design. I want to split the workflow below into work that generative AI can handle and work that a person must finalize.

[Workflow description]
- Workflow name: [name]
- Current steps: [bulleted steps]
- Data and systems used: [names]
- Approximate monthly volume: [count]
- Impact of a mistake: [one or two sentences]

[Request]
1. Break the workflow into inputs, processing, decisions, approvals, and outputs. For each, write one line stating who looks at what and decides what.
2. Answer each of these questions with "yes", "no", or "unknown", plus a one-line reason: Does it happen many times a month? Does the input arrive in a consistent shape? Can the desired output be described in words? Can the owner tell whether a result is right? Can a mistake be reversed?
3. Split the candidate work into tasks whose results can be checked, and produce a table with input, expected output, pass criteria, prohibited behavior, and fallback for each task.
4. Where information is missing, do not fill the gap with guesses. List it as "questions that need confirmation".

03Worked example: from invoice receipt to recording payables

The following example is fictional. It applies the checklist and the task definition sheet to an accounting workflow: receiving supplier invoices and recording them as accounts payable. It does not describe any real company or report any survey data.

Run through the checklist, this workflow gets "yes" on questions 1 to 4. Question 5, whether a mistake can be reversed, is "no" because correcting a posted entry takes real effort. So AI only drafts, and an AP clerk finalizes the entry. The work is split into four tasks: classification, extraction, matching, and drafting the journal entry. The sheet below covers matching.

None of the four tasks has write permission. Draft entries are not posted straight to the accounting system; they are posted when the clerk approves them on a review screen. Logging how many invoices showed discrepancies and how many draft entries the clerk changed gives you pilot metrics with no extra work.

04Connect data and decisions through AtoO

Deltalyze calls the shift from adding an AI feature to redesigning the operating model around AI AtoO: AI to Operating Model. Retrieval, decision, execution, and recording are designed as one flow, so the basis for every decision can be traced later.

Store the data consulted, AI proposal, human edits, and final approval separately. This makes errors traceable and turns corrections into reusable operational learning.

You do not need a large system to start. As in the template below, write one row per item processed, with the AI output and the human judgment in separate columns. The requirement is simple: one row should show what AI looked at, what it produced, and what a person changed. The second row fills in the fictional example above.

Do not allow the edit reason to be left blank. With reasons on record, you can count how often the same correction recurs, and that count tells you whether to fix the prompt, the business rule, or the reference data.

05Operate a measurable improvement loop

Track completion, exceptions, human edits, cycle time, and downstream rework, not accuracy alone. Accuracy can hold steady while exceptions climb and work quietly flows back to people, and in that case the workflow has not improved.

When quality drops, isolate whether the cause lies in reference data, business rules, the prompt, or the model. Fixing the order of investigation in advance means you do not have to suspect everything every time. The steps below check inputs and assumptions before the AI output itself.

Give operators a simple way to record exceptions and correction reasons, then feed validated changes back into the workflow definition. Small, continuous lessons are what turn AI adoption from a one-off project into an operating capability.

06Prove value in a small, controlled scope

Start with one workflow in one team. Take a baseline before rollout and run a four- to eight-week pilot comparing time saved, quality, and adoption. Also check that the saved time has not simply moved into some other manual task.

Decide and write down the metrics before the pilot starts. Metrics chosen mid-pilot tend to be the ones that already look good. If the workflow includes a monthly close, include at least one close in the pilot so you see peak-period exceptions too.

Once the value is confirmed, expand to adjacent work that can reuse the same data definitions and approval patterns. The goal is reusable operating capability, not a collection of isolated automations.