Process optimisation
Map current processes, identify automation opportunities, and design the future state — the upstream PM work that determines whether AI creates value or noise.
Process mapping for AI PMs
Before you can automate a process with AI, you must understand exactly how it works today. Skipping the as-is mapping leads to automating the wrong steps, missing edge cases, and building AI that creates problems faster than it solves them.
SIPOC — the fast-frame technique
Use in early discovery to align on scope before going deep. Takes 60 minutes with the right stakeholders.
Workday new hire onboarding: Suppliers = HR + IT, Inputs = offer letter + role data, Process = profile creation → system provisioning → training assignment, Outputs = activated employee, Customers = new hire + manager.
Swim lane diagram
Use after SIPOC to show who does what. Each horizontal lane = a role or system. Handoffs between lanes are your AI opportunity zones.
Invoice processing at a 500-person company: 4 lanes (AP Clerk, Manager, ERP System, Vendor). Longest wait = manager approval queue (avg 2.1 days). AI opportunity: route low-risk invoices (<$5k, verified vendor) directly to ERP, bypassing manager queue.
Value Stream Mapping — where time actually goes
VSM extends process mapping by adding time data. Most enterprise processes are 5–15% efficient — 85–95% of elapsed time is waiting. This changes what AI should target.
Time actually spent doing the work (value-add)
Usually 5–15% of lead time
Time the work sits idle between steps (non-value-add)
Usually 85–95% of lead time — the AI opportunity
Total elapsed time: process time + wait time
What the customer / user actually experiences
The highest-value AI target is not the step that takes the most process time — it is the step that creates the longest downstream queue. An AI that automates a 2-minute approval step and eliminates a 2-day wait delivers more business value than one that saves an hour at the end of a process with no downstream bottleneck.
Automation fit evaluator
Not every step should be automated — and not every automation should be full automation. Evaluate each process step against three criteria to determine the right automation level.
How often does this step occur?
How well-defined are the inputs and outputs?
What is the business cost of a wrong AI decision?
Select one option per criterion to see the recommended automation level.
As-Is → To-Be process design
The gap between the as-is map and the to-be design is your product scope. Get the as-is wrong and you will scope the wrong thing. Most PMs rush the as-is — senior PMs know it is the most valuable output.
As-Is mapping
Map exactly what happens today — the actual version, not the ideal one. Include informal workarounds, manual steps people do off-system, and the error rates nobody wants to talk about.
- Walk me through exactly what you do when [trigger event] happens.
- What information do you use to make that decision?
- How often does this step fail or require rework?
- What happens when [edge case]?
- Where do you spend the most time that feels like it should not be manual?
Never ask stakeholders to describe the ideal process — ask them to describe what they actually do. The gap between the ideal and actual is where the real pain lives.
Pain point analysis
For each process step, quantify: volume (how often?), time (how long?), error rate (what % goes wrong?), and downstream wait (how long before the next step picks it up?). Numbers turn anecdotes into requirements.
- What is the volume of this step per day/week?
- What is the average process time and wait time?
- What percentage of outputs require rework?
- What is the downstream cost when this step is delayed?
- Which errors are caught immediately vs. discovered hours or days later?
Pain points without numbers are opinions. Pain points with numbers are requirements. "This is slow" is an opinion. "This step takes 8 minutes and creates a 2-day downstream queue for 400 tickets per day" is a requirement.
To-Be design
Design the optimised process — which steps are automated, which are eliminated, which are transformed. Show the roles and systems in the new state. The gap between as-is and to-be is your product backlog.
- Which steps can be eliminated entirely (not automated — eliminated)?
- Which steps should be automated at Assist level vs. higher?
- What new steps does the AI introduce (monitoring, exception review)?
- How do errors flow in the to-be state — where does the human intervene?
- What does success look like in 6 months? How will we measure it?
Do not skip to to-be design without a complete as-is. The to-be is only as good as the as-is that precedes it. Premature to-be design is the root cause of most AI features that solve the wrong problem.
High-fit vs. low-fit for AI automation
Not every process benefits from AI. Knowing what AI should not automate is as important an interview signal as knowing what it can.
High-volume classification
Routing customer emails, triaging medical claims, labelling support tickets. Repetitive, well-defined outputs, high volume justifies model investment.
Zendesk AI: classifies and routes 60M+ support tickets per month — 85% routed correctly without human triage.
Document extraction
Extracting structured data from unstructured documents — contracts, invoices, medical forms. Well-bounded inputs and outputs.
Workday AI: extracts skills from unstructured CVs into structured competency profiles. Reduced HR admin time by 3 hours per hire.
Draft generation
Generating first drafts from structured inputs — policy docs from templates, email replies from ticket context, reports from data.
Salesforce Service Cloud: drafts email responses for agents using ticket context + knowledge base. Agents edit and send — reduces average handle time by 35%.
Anomaly detection
Flagging unusual patterns in data streams — fraud signals, system failures, quality defects. Pattern recognition in high-volume streams.
JPMorgan COiN: detects anomalous loan covenant clauses in legal documents. Replaced 360,000 hours of manual annual review.
Novel/crisis decisions
Situations with no historical precedent — new market entry, regulatory response, M&A. AI cannot learn from what has never happened.
No AI was trained on COVID response protocols before March 2020. Human judgment is irreplaceable for genuine novelty.
Relationship-critical tasks
Sensitive customer negotiations, HR counselling, executive stakeholder management. Emotional intelligence and context that AI cannot replicate.
Enterprise sales deal closure: AI can surface insights and draft emails, but a human must close. Trust is built through the relationship, not the model.
Change management — the often-skipped PM requirement
AI products fail not because the model is bad but because the humans who use it do not trust it or change their behaviour. Change management belongs in the PRD, not in post-launch retrospectives.
Process optimisation interview Q&A
Five scenario questions that surface whether you can apply frameworks to real-world constraints — not just define them.
How do you identify where to apply AI in an existing business process at a company like Workday?
I start by mapping the current process end-to-end — not the ideal version, the actual one, including informal workarounds. I use Value Stream Mapping to capture both process time (time spent doing the work) and wait time (idle time between steps). Most enterprise processes are 5–15% efficient — 85–95% of elapsed time is waiting.
Once I have the map, I apply three filters to identify AI candidates: volume (how often does this step occur?), structure (are the inputs and outputs well-defined?), and error tolerance (what is the business cost of a wrong decision?). The intersection of high volume, structured inputs, and tolerable error rate is where AI delivers reliable value.
At Workday, I would look specifically at skill extraction from CVs, benefits eligibility verification, and payroll anomaly detection — all high-volume, structured, and with bounded error tolerance. I prioritize the step that creates the biggest downstream wait, not the step that takes the most individual process time.
Common trap: Starting with the AI solution and working backward to find a process it fits. This is the most common AI PM mistake — it leads to building features nobody needs because the process problem was never validated.
You are introducing AI automation to a customer service team. The team lead says the agents will not trust it. How do you handle this?
I treat adoption as a product requirement, not an afterthought. Three concrete actions.
First, I involve the team lead and two experienced agents in the design sprint. They surface edge cases the data does not capture, and their participation creates a coalition of early believers rather than resistors.
Second, I start the rollout at the Assist level — the AI suggests, agents decide. No automation, no pressure. I track suggestion acceptance rate. When agents are accepting more than 60% of suggestions without modification, they have revealed their trust threshold.
Third, I make the AI's reasoning visible. When agents see why the model recommended a response — "based on similar tickets, customers who reported this issue resolved it with X" — they can evaluate the suggestion rather than just accept or reject it blindly. Explainability is not a nice-to-have — it is the primary trust mechanism for non-technical users.
Only after acceptance rate stabilizes above 60% for four weeks do I propose moving to Automate with Review. The rollout plan is in the PRD before sprint 1.
Common trap: Launching at a higher automation level and trying to debug trust after agents revolt. Build trust the way you build any metric — define the target, measure it, and do not advance until you hit it.
Walk me through how you would run an as-is / to-be process workshop for an AI opportunity identification project.
Two workshop sessions.
Session one — as-is mapping. I prepare a SIPOC template to align on scope before the session: who supplies inputs, what the inputs are, the five to seven high-level process steps, the outputs, and who receives them. In the session, I walk through each step with the people who do the work, not just their managers. I capture: what information drives each decision, how long each step takes, the error rate at each step, and — most importantly — where the step sits before a long wait queue. I do not optimize for an hour; I look for a two-minute step blocking a two-day queue.
Session two — to-be design. I present back the as-is with the top three bottlenecks quantified. I use that as the anchor for to-be design: which steps are automated, which are eliminated, which are transformed. The gap between as-is and to-be becomes the product backlog with a clear linkage between feature and business pain.
Common trap: Running only one session and rushing to the to-be. The as-is is the most valuable output — it takes two sessions because the first session surfaces the surface view and the second surfaces the reality under it.
What is Value Stream Mapping and how does it change where you prioritize AI?
VSM extends standard process mapping by adding time data to every step. For each step you capture two numbers: process time (how long the actual work takes) and wait time (how long the work sits idle before someone picks it up). Lead time equals the sum of both.
Most enterprise processes have an efficiency of 5–15% — meaning 85–95% of elapsed time is waiting, not working. This changes the AI prioritization logic fundamentally. The highest-value AI target is usually not the step that takes the most process time — it is the step that creates the longest downstream queue.
A concrete example: if invoice approval takes a manager two minutes but the invoice sits in their queue for two days, automating that two-minute step eliminates the two-day wait for the AP team. An AI model that saves two minutes of manager time also saves two days of lead time. That is the VSM insight — and it is the reason process mapping before AI scoping is not optional, it is the first deliverable.
Common trap: Optimizing the step with the highest process time rather than the step with the highest downstream wait. Most PMs see the work; VSM makes the waiting visible.
Your stakeholder wants to automate a high-stakes hiring decision using AI. How do you respond?
I do not say no immediately — I say let us define what "automate" means and then evaluate whether this use case can support that level of automation.
Hiring decisions in most jurisdictions are classified as high-risk under the EU AI Act and face EEOC scrutiny in the US. That means any automated decision requires an auditable rationale and a human review path — full automation is off the table legally, not just ethically.
But there are real automation opportunities within the process: screening for minimum qualifications (structured, high-volume, low-stakes), extracting skills from resumes into a structured profile (well-defined extraction task), and ranking by competency match (a recommendation, not a decision). These are Assist-level automation — the AI surfaces information, the human makes the decision.
I reframe the question: "Rather than automating the hiring decision, let us automate the triage so your recruiters spend less time reading unqualified applications and more time evaluating strong ones." That delivers 80% of the value at 10% of the regulatory risk. I document this framing in the PRD so the constraint is explicit and defensible.
Common trap: Saying "we cannot do this" without offering an alternative. The PM's job is to find the highest-value thing that can be done within the constraint — not to close the conversation.
I'm an AI PM candidate preparing for a process discovery interview. Give me a scenario: an enterprise company wants to use AI to automate a specific business process. Ask me to walk you through how I would approach this — from as-is mapping to automation level recommendation — then give me feedback on whether I applied VSM thinking, correctly evaluated automation fit, and addressed change management.