←
AI-Ready Foundation
Visionary · M1 · lesson 1 of 5 · in progress
Preview — browse every lesson free. Enroll to mark lessons complete, open partner links and save your progress. Login & enroll →
AI-Native Process Redesign: Rebuilding End-to-End Workflows Around AI
📖
now learning

AI-Native Process Redesign: Rebuilding End-to-End Workflows Around AI

15 min

A storm has passed, a household's fridge has failed, and the family lodges a claim for a few hundred dollars of spoiled food. At Allianz in Australia, seven specialized agents check that claim and put it in front of a claims professional for the payout decision in under five minutes, according to the insurer's own account of Project Nemo, published 3 November 2025. That is AI-native process redesign: drawing a process from the outcome backwards as if AI had always existed, then deciding which steps are agent work, which are plain rule-based automation, and which decisions stay with a named person. Many firms give the handler a copilot instead: it drafts the note in thirty seconds, and the customer is still paid on the same day as last year. This lesson shows why speeding up one step rarely moves the result, what Allianz actually changed, and how to run a three-hour workshop that ends with a one-page Process Redesign Charter.

The Problem: Faster Steps, Same Process, No Profit

Now picture the common version in your own claims department (an illustration, not a real firm). A handler opens a new claim, clicks a button, and the copilot drafts the assessment note before the coffee is cool. Then the note waits two days in the same review queue it always waited in. It passes the same four handoffs, finance pays it in the same weekly batch, and the customer is paid on the same day as a year ago.

Meanwhile the AI dashboard shows plenty of "hours saved" and the P&L shows nothing. Nobody lied: the tool really did save the handler fifteen minutes, which fell into a queue and disappeared. The pattern is expensive because it looks like success: the tool works, the licence count grows, the process is untouched.

What did Allianz do with its claim queue?

Allianz rebuilt one narrow slice of the claims process around agents instead of handing its handlers a chatbot. It is a published case of AI agents in claims processing with a human kept at the money; the details below come from Allianz's article of 3 November 2025.

After natural catastrophes in Australia, Allianz faced surges of small food-spoilage claims. Each is simple, but they arrive all at once.

The company built what it calls Project Nemo. Seven specialized agents each do one job: a planner, a cyber agent, a coverage agent, a weather agent, a fraud agent, a payout agent and an audit agent. They work only on low-complexity claims under AUD 500. A claims professional still makes the payout decision, by design, and the audit agent summarizes every other agent's decision so that person can see what was checked and why.

Allianz reports that processing went from four or more days to about a day or a matter of hours, and cites an 80% reduction in processing and settlement time. It says the system went live in under 100 days from a July 2025 launch, with travel delay, simple motor and property claims planned next.

A caveat before you put this on a slide: the 80% figure is Allianz's own and has not been independently audited. Use the design; treat the number as company-reported.

Copy the design, not the number:

  • A narrow slice. One claim type, one peril, one value ceiling.
  • Single-responsibility agents. Each can be tested and blamed on its own.
  • Audit built in. Part of the design, not a report written afterwards.
  • A human at the money-moving decision. Agents prepare the payout; a person decides it.
  • An end-to-end measure. Processing and settlement time for the whole claim, not minutes saved per task.

What does bolting AI onto a process cost?

It costs the gap between rising AI spend and flat profit. The survey numbers below are self-reported, so read them as signals, not audited accounts.

  • Profit is mostly not moving. Only 37% of respondents report any EBIT impact from AI, roughly flat year over year, according to McKinsey's State of AI survey, published 25 August 2026 (1,719 participants).
  • Redesign is the biggest visible difference. In that same McKinsey survey, nearly three-quarters of AI high performers have redesigned workflows around AI, up from 55% in 2025, against about one-quarter of everyone else.
  • Most firms are behind, and the value is in core operations. BCG's study of 1,250 executives, released 30 September 2025, put 60% of companies in a laggard group with minimal gains and located 70% of AI's potential value in core functions such as sales, manufacturing, supply chain and pricing, with IT at 13%.
  • Spending is doubling. BCG's press release of 30 September 2026 says AI spending roughly doubled, from about 1.7% of revenue in late 2025 to 3.3% in 2026.

Spend is doubling while fewer than four in ten firms see profit move, and the firms that see the most are far more likely to have changed the work. Survey correlations do not prove that redesign causes the outperformance, but it is the strongest behavioural difference with numbers behind it.

Who owns this problem?

The business process owner owns it, with the COO: the head of claims, customer operations or finance operations, not the CIO. The CIO or head of AI supplies the platform, guardrails and engineers. The outcome (days to payment, cost per case, complaints) belongs to the person who runs the process.

The money already reflects this. The same BCG release of 30 September 2026 reports that more than 80% of AI spending now sits outside the enterprise IT budget. If the budget lives in the business line, so must the redesign. Chaired by IT, a redesign tends to produce a better tool; chaired by the process owner, a better process.

Where this lesson sits on the ladder

This is Level 5, AI Native, lesson 1 of 5: the top of a ladder that runs from AI Aware through Capable, Operational and Scaled, and the rung where AI stops being a project and becomes the way the work is designed.

In Lesson 20 on the AI operating model and the 10-20-70 rule you wrote an AI Operating Model Charter: who decides, who builds, who pays and who is accountable, plus the budget mirror. That charter is an input today. What is new at Level 5 is the unit of change: Lesson 20 set the rules for the whole organization, and this lesson rebuilds one end-to-end process inside them. Lesson 22 on governed autonomy, agent identity and kill switches then picks up the question your redesign will raise: how much each agent may do on its own, under whose identity, and how you stop it in seconds.

Picture the destination. An AI-native enterprise can absorb a surge of claims after a storm without a hiring round, change a policy rule once and see every case follow it the same day, and show an auditor, for any decision, what was checked, by which agent and who signed. Its rivals can rent the same models. What they cannot rent is a process built around them, and that takes one slice at a time.

How AI-Native Process Redesign Works: The Dishwasher in the Old Kitchen

A busy restaurant buys a dishwasher that cleans plates five times faster than the old one. But the machine sits at the back, so staff still carry plates across the dining room one stack at a time, and someone still dries every plate by hand. On Friday night, dinner is not served one minute sooner.

Now picture a kitchen designed around the machine. The pass, the racks, the machine and the storage sit in one line. Plates go in dirty at one end and come out ready at the other, with nobody walking. The same machine now changes how many covers the restaurant can serve in an evening.

The machine was never the constraint. The walk was. In your next steering committee, when someone shows you a tool that makes a task faster, ask, where is the walk?

What is AI-native process redesign?

AI-native process redesign is rebuilding a process from its outcome backwards so that agents, fixed automation and human decisions each sit where they fit best. Bolt-on AI, by contrast, adds an assistant to one step and leaves everything around it as it was.

Four ideas make it work.

Why does speeding up one step not speed up the process?

Because a process is a chain, and a chain moves at the speed of its slowest link. Speed up any other link and the work simply arrives sooner at the queue in front of the slow one. The queue gets longer; the result arrives on the same day.

Touch time vs wait time is the most useful split in this lesson. Touch time is the time someone (or something) is actively working on a case. Wait time is the time the case sits untouched, in an inbox, a queue or a batch. In many office processes wait time dwarfs touch time; step 3 of the workshop shows whether yours is one. A copilot attacks touch time. The days are usually in the waits.

People who run AI at scale say the same thing. JPMorgan Chase's Chief Analytics Officer is reported by AI News on 16 December 2025 as saying that in end-to-end processes, task-level time savings "often just shift bottlenecks". Google's DORA (DevOps Research and Assessment) report (23 September 2025, about 5,000 respondents) found the same in software delivery: AI adoption tracks higher throughput but still lower delivery stability, an "amplifier" of existing strengths and weaknesses.

So the first rule of redesign: whenever you make one step faster, ask what queue grows next, and redesign that step in the same release.

Does every step need an AI model?

No. Many steps should be deterministic automation: fixed rules that do the same thing every time, such as routing a form by type, posting a credit note on approval or sending a survey when a case closes. They are cheap, predictable and easy to audit. Use a model only where a step needs judgment over messy inputs, such as reading a contract clause or weighing evidence.

This is normal engineering. In February 2026 Salesforce told IT Brief it uses "deterministic triggers" for steps such as sending surveys, an intentional and standard design choice, while disputing reports that it had cut about 4,000 support roles because of AI. The staffing story is contested; the design point is not.

In the workshop you will sort every step into one of four bins:

BinWhat it meansTypical example
Agent recommendsA model gathers evidence and proposes an answer, with reasonsAssemble the contract clause and propose a credit
Deterministic automationA fixed rule acts the same way every timeRoute the dispute by type; post the approved credit note
Human decidesA named role makes the call, with an audit summaryApprove the payout or the credit
DeleteThe step exists only because of the old designA monthly report nobody reads once a live view exists

The fourth bin is the one teams forget. Many old steps exist only to move paper between people; when the paper stops moving, the step can go.

Where should the human sit in an AI-native process?

Put the human at the decision that is costly or hard to reverse, not sprinkled across every step. In Project Nemo that decision is the payout. The agents prepare it, the audit agent explains it, and a claims professional makes it.

Scattering checks everywhere feels safe but trains people to click "approve" without reading. The AI News writer warns of reviewer complacency when AI is right 85 to 95% of the time. In that range errors are rare enough that a reviewer stops expecting them, and then stops looking. One well-placed human with an honest summary and a clear threshold is worth more than five tired ones.

Why measure instead of asking people?

Because people's sense of their own speed can point the wrong way. In METR's randomized trial of July 2025, experienced developers believed AI had made them 20% faster while the stopwatch showed them 19% slower; Lesson 19 on measuring AI ROI honestly tells that story in full, with its limits. A questionnaire asking "did this save you time?" is not evidence. A timestamp at the start, a timestamp at the end and a comparison group are evidence.

The Workshop: Run a Process Redesign Session and Write the Charter

This is the part you can run this week: one three-hour session plus about two hours of data pulling beforehand, ending with a one-page Process Redesign Charter. The steps are recommended practice synthesized from the cases in this lesson, not a published standard.

Who needs to be in the room, and what do they bring?

Eight people, and the chair is the process owner, not IT. You need the people who know where cases wait, the people who will carry the new queue, and the people who sign off on money, risk and jobs.

  • The process owner (chair). Owns the outcome and signs the charter.
  • Two front-line practitioners. They know where cases really wait, which is rarely where the process map says.
  • One team leader from the middle layer. They carry the new approval queue.
  • The AI platform lead. Says what the platform can do safely, and at what cost.
  • A risk or compliance partner. Places the human decision point and the thresholds.
  • A finance analyst. Owns cost per case and keeps the baseline honest.
  • An HR business partner. Owns step 9: the role sketch, the redeployment plan and the order in which employee representatives are told.

Before the session, pull the inputs: 90 days of case data (volumes, a timestamp per step, rework, complaints), the current process map, the cost per case, and your Operating Model Charter from Lesson 20. Many case systems hold only open and close timestamps. If per-step timestamps do not exist, reconstruct the twenty cases for step 3 from the audit trail or email headers, and note the gap as a data requirement on the charter.

What are the nine steps?

  1. Choose the slice. Pick one end-to-end process in a core function, then narrow it to a high-volume, low-severity segment: one case type and one value ceiling, the way Allianz chose food-spoilage claims under AUD 500. Write the ceiling down: a slice without one grows until it includes the hard cases that sink the project.
  2. Baseline before touching anything. Record end-to-end cycle time, cost per case, first-time-right rate, complaint rate, rework rate and demand volume, each with its date range. No baseline, no redesign: without one you argue about feelings, as Lesson 10 on pilot baselines and the go or no-go decision showed.
  3. Walk the timeline. Take twenty recent cases. For each step, mark touch time against wait time. Circle the three longest waits. These are your targets, not the slowest tasks.
  4. Redraw from the outcome backwards. Write the outcome in one sentence ("the customer receives a correct decision and, if due, the money"). List only the decisions needed to reach it. Then sort every existing step into the four bins.
  5. Redesign the next bottleneck in the same release. If the first steps became instant tomorrow, which queue grows? Usually review, a compliance check, testing or a finance batch. Redesign it now. "Phase two" is where shifted bottlenecks go to live.
  6. Set the human decision point and the audit summary. Name the role that decides, what they see on one screen, and the value threshold above which they must approve each case individually.
  7. Plan the rollout ladder. Rung one is shadow mode: agents run, humans decide as before, and you compare. Rung two is human approval on every case. Rung three relaxes approval by value threshold, such as batch approval below a ceiling. Write each rung's exit criterion before you start, so nobody moves the goalposts.
  8. Measure, do not survey. Define a holdout team, region or week that keeps the old process, and track the step 2 metrics on both sides, plus cost per case including tokens so model costs sit next to labour.
  9. Hold workforce decisions and sketch the new roles. Agree in writing that no role changes are announced until quality and demand data have held for a sustained period you define now (for example, two full quarters). Then sketch what each affected role does in the new process, write the redeployment and retraining plan, and set the order in which people are told.

How should you spend the three hours?

The timeline walk always runs long, so protect the last 70 minutes for steps 5 to 9, the ones that make the redesign real.

TimeStepsOutput on the wall
0:00 to 0:201 and 2Slice definition with value ceiling; baseline numbers with dates
0:20 to 1:103Touch and wait table for twenty cases; three longest waits circled
1:10 to 1:504Outcome sentence; every step in one of four bins
1:50 to 2:205 and 6Next-bottleneck redesign; decision role and threshold
2:20 to 2:407 and 8Rollout ladder with exit criteria; holdout design
2:40 to 3:009Workforce commitment; role sketch; who is told, in what order

What does it look like filled in?

An illustrative example, not a real company: picture a 900-person regional equipment-leasing firm handling about 1,400 invoice disputes a month, most under 300 euros. A copilot already drafts the replies, and the team likes it. Yet a dispute still takes nine working days from the customer's email to the corrected invoice.

The head of customer finance chairs. Step 1 sets the slice (billing disputes under 300 euros), step 2 records the baseline over the previous 90 days, and the group walks twenty cases:

Step in today's processTouch timeWait timeRedesign binRedesigned step
Customer emails dispute5 min1 day in shared inboxDeterministic automationIntake form routes by dispute type on arrival
Handler pulls contract and invoice20 min0.5 dayAgent recommendsRetrieval agent assembles contract clause, invoice lines and usage log
Handler drafts assessment (copilot)15 min0Agent recommendsAssessment agent proposes credit or rejection with cited evidence
Team leader review10 min4 days in queueHuman decides (threshold)One-screen audit summary; batch approval under 300 euros twice daily
Finance posts credit note5 min3 days (weekly batch)Deterministic automationCredit note posted on approval, same day
Monthly dispute report2 hoursn/aDeleteReplaced by a live dashboard

Add up the touch time on a single dispute (the monthly report is shared across all of them) and you get 55 minutes of work inside nine days of elapsed time. The copilot attacked the 15-minute drafting step, the one step with no wait in front of it. The real targets were the four-day review queue and the three-day finance batch: the dishwasher at the back of the kitchen.

Step 5 asks what queue grows next. If intake, retrieval and drafting become near-instant, the team leader's queue fills faster than ever, so review is redesigned in the same release: one audit summary per dispute, twice-daily batch approval under the ceiling, individual approval above it. One step further, approved credits would pile up in front of the weekly finance batch, so posting becomes a deterministic rule on approval.

Step 6 names the decision: the team leader still decides every credit. Step 7 sets the ladder: four weeks in shadow mode, exiting when agent proposals match the human decision in an agreed share of cases; then approval of every case; then batch approval under 300 euros. Step 8 picks a holdout region that keeps the old process for a quarter. Step 9 records three commitments: no role changes before two quarters of quality and demand data, freed team-leader time goes first to complex disputes over 300 euros, and the employee forum hears before shadow-mode results are shared.

Their target, measured against the holdout: two working days end to end, with the team leader still deciding every credit. Note what they did not write: "hours saved".

What goes in the Process Redesign Charter?

The Process Redesign Charter is a one-page record of what you will change in one process, how you will prove it worked and what you promise people while you find out. Fill in every row; an empty row is a decision nobody has made yet.

FieldWhat to writeExample entry (illustrative)
Process and sliceProcess name, segment, value ceilingInvoice disputes; billing disputes under 300 euros
Business ownerOne named person who owns the outcomeHead of customer finance
Outcome statementOne sentence, from the customer's sideCustomer receives a correct decision and, if due, a credit note
Baseline metricsThe step 2 metrics, with datesNine working days median; 90 days of data to quarter end
Timeline and binsTouch and wait per step from twenty cases; each step's bin55 minutes of touch in nine days; two agent, two rule, one human, one delete
Next-bottleneck redesignThe downstream step redesigned in the same releaseTeam-leader review and finance posting
Human decision pointRole, what they see, approval thresholdTeam leader; one-screen audit summary; batch below 300 euros
Rollout ladderThree rungs, each with an exit criterionFour weeks shadow; match rate agreed in advance
Measurement designComparison group, metrics, review dateHoldout region for one quarter; review at quarter end
Workforce commitmentsHold period, role sketch, redeployment plan, who is told in what orderNo role changes for two quarters; freed time to complex disputes; employee forum first
Data requirementsTimestamps or fields the measurement needs but the system lacksPer-step timestamps for review and posting, logged from shadow mode
Budget lineWhere the money sits and who holds itCustomer finance operating budget, held by the business owner

A point for the risk partner. If your slice touches a use the EU AI Act lists as high-risk in Annex III, those duties apply from 2 December 2027 under the Digital Omnibus on AI, Regulation (EU) 2026/1744, and two land directly on this charter. Under Article 26 of the AI Act, Regulation (EU) 2024/1689, the person at your human decision point must have "the necessary competence, training and authority" (Article 26(2)), and an employer must inform workers' representatives and the affected workers before putting a high-risk system into use at the workplace (Article 26(7)). Note both, and confirm against the Official Journal text.

How do you redesign the roles around the new process?

From the four bins, not from the org chart. Every step that left a person's hands frees time, and every new decision point or exception queue needs an owner. The leasing example again, still illustrative:

Role todayWhat moves to agents or rulesIn the new processDecide before rollout
Dispute handlerPulling documents, drafting the assessmentWorks disputes above the ceiling and the exceptions the agents route outTraining on exception types; how exceptions count in targets
Team leaderReading every case from scratchApproves in batch below the ceiling, individually above it; reviews override reasons weeklySpan of control: how many summaries one person can read properly in a day

Test the span of control in shadow mode: count approvals per approver per day and ask whether each summary was really read. If not, you have built reviewer complacency into the org chart, and the fix is a lower batch ceiling or a second approver, not a reminder email.

Then sequence the conversations: employee representatives (a works council or union, where you have one) before any announcement, affected staff before shadow-mode results circulate, everyone else last. Local labour law may require more, and earlier, so bring in employment counsel before shadow mode starts.

Failure Modes, How to Spot Them, and What to Measure

Redesigns fail in a small number of predictable ways, and two widely reported reversals of 2025 were not technical failures at all.

What happened when Klarna and Commonwealth Bank cut staff for AI?

Both reversed course in 2025, because capacity was removed before quality and demand had been checked.

Klarna said on 27 February 2024, in its own press release, that its AI assistant handled 2.3 million conversations in its first month, two-thirds of its customer-service chats, doing the equivalent work of 700 full-time agents. These are company claims, not audited figures. Entrepreneur reported on 9 May 2025 that headcount had fallen 22% to 3,500 by December 2024, mostly through attrition and a hiring freeze, and that the CEO said it was hiring human agents again, that customers should know "there will always be a human if you want", and that the AI route had produced "lower quality". The lesson: cost was tracked, quality was not.

Commonwealth Bank of Australia announced in July 2025 that 45 customer-service roles would be replaced by an AI voice bot. On 21 August 2025, as ACS Information Age reported, it reversed the decision, said its assessment "did not adequately consider all relevant business considerations", said the roles "were not redundant", and apologized. The union said call volumes were rising. The lesson: check demand before removing capacity.

The surveys show the same overreach at scale: in McKinsey's 2025 State of AI survey, 32% of respondents expected AI to cut their headcount over the following year, yet in the August 2026 edition only 14% reported an AI-driven workforce decline.

Which failure modes should you watch for?

Six account for most stalled redesigns, each visible in data you hold.

Failure modeWhat it looks likeHow to spot it
Bolt-on copilotAI added to a step; the process around it unchangedTask time falls, end-to-end cycle time does not move
Shifted bottleneckThe work arrives faster at a queue nobody redesignedA downstream queue (review, compliance, finance) grows after launch
Savings by surveyBenefits claimed from what people feel"Hours saved" from a questionnaire, with no comparison group
Headcount firstThe business case is written as job cutsRole cuts announced before a quarter of quality and demand data
Reviewer complacencyThe human approves without readingOverride rate drifts toward zero while complaints or rework rise
Middle-layer overloadTeam leaders absorb the new approval loadTeam leaders' approval queues and overtime rise

Why do the people in the middle need special attention?

Because they carry the redesigned process and already feel the strain most. McKinsey's survey published 25 August 2026 reports that 47% of mid-level managers and individual contributors feel AI-related strain, against 31% of executives. That is why a team leader sits in the workshop and why the span of control is tested in shadow mode. Measure strain among the people who run the queue, separately from executives.

Which metrics should you track after the redesign?

Track five, on both the redesigned process and the holdout.

MetricHow to measure itWhat it warns you about
End-to-end cycle timeTimestamp at request and at outcome delivered; median and 90th percentileBolt-on copilot; shifted bottleneck
Straight-through rateShare of cases with no human edit before the approval pointAgents that need constant correction
Human override rateShare of agent recommendations the approver changes, with reason codesComplacency if near zero; poor agents if high
QualityComplaint, repeat-contact and rework rates against baselineThe Klarna pattern: cost down, quality down
Cost per case including tokensLabour plus model and platform cost per completed caseSavings that exist only on the labour line

Watch the 90th percentile as closely as the median: a redesign that speeds typical cases while the awkward tenth waits longer shows up there first. There is no public benchmark for any of these, such as an acceptable override rate, so judge every change against your step 2 baseline and the holdout. An override rate falling toward zero is good news only if complaints and rework fall too.

Frequently Asked Questions

What is the difference between AI workflow automation and AI process redesign?

Automating a step makes the existing process faster at that point; redesign rebuilds the whole process from the outcome backwards, deleting some steps, giving others to fixed rules and moving the human to the decision that matters. Automating one step usually moves the queue, not the result.

Why did our AI tool save time but not money?

Speeding one task usually moves the queue to the next step, such as review or compliance. Unless you redesign that downstream step in the same release, the total cycle time and cost per case stay where they were.

Which process should we redesign first?

Start with a high-volume, low-severity slice of a core function, with a clear value ceiling and good case data. Allianz began with food-spoilage claims under AUD 500.

Should we reduce headcount once the new process is live?

Not until quality and demand data have held for a sustained period you set in advance. Klarna and Commonwealth Bank of Australia both publicly reversed AI-driven staffing decisions in 2025.

How do we prove the redesign worked?

Compare the redesigned process against a holdout team, region or period on cycle time, quality and cost per case. Do not rely on staff surveys: in METR's July 2025 trial, developers felt 20% faster while measuring 19% slower.

Key Takeaways

  • AI-native process redesign rebuilds a process from the outcome backwards; a copilot bolted onto one step usually leaves the end-to-end result unchanged.
  • Find the walk, not the slow task: mark touch time against wait time on twenty real cases and target the three longest waits.
  • Sort every step into four bins (agent recommends, deterministic automation, human decides, delete), and redesign the next bottleneck in the same release.
  • Put one named human at the costly or hard-to-reverse decision, with an audit summary and a value threshold, as Allianz did for payouts.
  • Measure with a holdout and timestamps, not a survey, and read override rate together with quality.
  • Hold workforce decisions until quality and demand data have held for a period you set in advance; sketch the new roles and tell employee representatives first.
  • Leave the session with a signed Process Redesign Charter for one slice, owned by the process owner, not IT.