The AI Operating Model: Workflow Redesign, Change Management, and the 10-20-70 Rule
Three months after launch, the usage chart for your best AI tool looks like a comb with missing teeth: a few enthusiasts, long gaps, and one department that never logged in. The tool is fine. What is missing is an AI operating model, the written answer to four questions: who decides, who builds, who pays, and who is accountable for the result. This lesson shows you how to write one: split the work between a central hub and the business units, cut a first redesign of one workflow, draft the role changes and take them through workforce consultation where it applies, and test your budget against the 10-20-70 rule. You leave with a signed AI Operating Model Charter and a consultation plan for the role changes.
The Problem: A Good Tool Dropped Into a Process That Never Changed
Look closer before you blame the tool. It passed its evals, runs on the governed platform, and demoed well. The gaps are not random: each missing tooth is a team where nothing around the tool changed.
Walk the floor and you find out why. The process is exactly what it was last year: the same four approvals, the same handoff by email, the same job descriptions. Managers are not sure they may change any of it, so they have not. And the central AI team has been told to "drive adoption" in departments it does not run, with people it does not manage, against targets it does not own.
Then look at the budget. There are lines for models, licences and compute, and none for training hours, process work or the manager time it takes to change how a team operates. The money went to the machine, not to the people standing around it.
What did Moderna reportedly do about this?
It reportedly put people and technology under one executive. A caution: the story reaches us through a secondary source quoting a paywalled Wall Street Journal article, so treat every detail as reported, not confirmed.
As FlexOS reported in May 2025, citing The Wall Street Journal, Moderna combined human resources and technology into a single remit held by Tracey Franklin, with the title Chief People and Digital Technology Officer. The company is reported to have deployed more than 3,000 custom GPTs since 2023. The number matters less than one sentence attributed to Franklin:
"I've shifted from workforce planning to work planning."
That is the whole lesson in eight words. Workforce planning asks "how many people do we need in each box?" Work planning asks "what is the work, and who or what should do each part of it?" The reported practice at Moderna is that roles are created, removed or reshaped depending on whether a task suits people or machines.
The unit of design is the work, not the headcount and not the tool. And the question "who or what should do this task?" has to belong to somebody. In most companies it falls between HR, who own roles, and IT, who own systems, so nobody asks it. You do not need to copy Moderna's merger. You do need one named owner for work design in each business domain.
What does leaving the process alone cost?
It costs you most of the financial return, if the survey evidence is any guide. One caveat applies to every number in this section: these are self-reported survey correlations from firms with a commercial interest in the answer, meaning consultancies that sell transformation services and, in Cisco's case, a vendor of AI infrastructure. They show which practices travel with results; they do not prove cause.
- Redesign is the biggest lever and the least used. McKinsey's State of AI report of March 2025 (survey fielded July 2024, 1,491 respondents) tested 25 attributes and found workflow redesign had the largest effect on EBIT impact from generative AI. That is a different cut from the KPI-tracking finding in Lesson 19, which ranks 12 adoption practices, so the two do not contradict each other. Only 21% of organizations using gen AI said they had fundamentally redesigned at least some workflows.
- The gap is widening. McKinsey's State of AI, published 25 August 2026 (1,719 respondents), reported that nearly three-quarters of high performers have redesigned workflows around AI, up from 55% in 2025, against about one-quarter of everyone else. High performers remain about 6% of respondents.
- The hard part is people. BCG's study of 24 October 2024 (1,000 executives) found roughly 70% of AI implementation challenges are people and process issues.
- Few have a plan for the change. The Cisco AI Readiness Index of 14 October 2025 reported that 91% of its top group, the "Pacesetters", have a change-management plan, against 35% overall. The Register, on 13 November 2025, cited Cisco as finding that only 32% of companies had identified which human tasks they want agents to supplement or replace; the figure is not in Cisco's press release, so treat it as second-hand.
Put that last figure next to the Moderna quote: if it holds, about two-thirds of companies in that survey had not done the work planning. They bought the tool before deciding what work it was for.
The opposite mistake is redesign for the wrong reason, the Klarna pattern told in Lesson 19's honest ROI ledger: a customer service cut steered by cost and partly reversed when quality suffered. A redesign driven by a headcount target alone is still a redesign, and it can still be wrong.
Where this sits on the ladder: this is Level 4, AI Scaled, and the last lesson of the level. Lesson 19 on measuring AI ROI honestly gave you a value ledger, an honest record of what each initiative is worth and how you know. This lesson asks who owns each line in that ledger, and what must change in the organization so the value shows up. Lesson 21 on AI-native process redesign opens Level 5 by redrawing the process from a blank page and deciding, step by step, what is agent work, plain automation or a person's decision. Here you make the first cut the operating model needs: enough to know whose roles change, who owns the result and who must be consulted.
How an AI Operating Model Works: The Boat, the Motor and the Crew
Start with a boat. An eight-person rowing boat, eight rowers facing backwards, one person at the stern calling the stroke.
Now someone bolts an outboard motor to the back and changes nothing else. All eight rowers stay at their oars, facing the wrong way. The boat is a little faster and far more awkward. Nobody knows whether the caller or the motor sets the pace. Half the crew row harder to prove they are still needed; the other half quietly stop.
To get the benefit of the motor you have to do three things that have nothing to do with the motor. You change the seating. You decide who steers. And you work out what the rowers do now, because some of them are about to become navigators, mechanics and lookouts. In a real organization there is a fourth: the crew has a say in the new seating plan, and in many countries that say is written into law.
The motor was the cheap part. Rearranging the boat and crew is the expensive part, and nobody budgeted for it.
What is an AI operating model, in plain terms?
An AI operating model is a short written document that says who decides, who builds, who pays and who is accountable for AI results. In the boat, it is the answer to "who steers, who maintains the motor, who bought the fuel, and who answers if we miss the harbour?"
Unwritten, those answers get settled by default, and badly: the central AI team decides, builds and pays, then is held accountable for outcomes in departments it does not run. That is the mechanic being blamed for the route.
What is workflow redesign?
Workflow redesign means changing the end-to-end sequence of steps, handoffs and checks, instead of making single steps faster. Speeding up one step is fitting the motor. Redesign is changing the seating.
The test is simple. Put the process map from before the tool beside the map from after. If the boxes and arrows are the same and one box now says "with AI", you have not redesigned anything. A real redesign removes steps, merges handoffs, and moves human attention to the places where judgment is needed.
Satya Nadella made the same point, as SiliconANGLE reported on 29 April 2025: productivity change "requires software and also management change", and he pointed to the roughly 50 years factories took to exploit electricity.
What is a hub and spoke model?
Hub and spoke is a structure in which a central team, the hub, provides the platform, standards, risk controls and training, while the business units, the spokes, own the use cases and the outcomes. The hub is the boatyard: it supplies and services the motors, sets the safety rules and trains the crews. The spokes are the boats, and each crew decides where to go and answers for getting there.
If you are asking how to structure an AI team, this is the comparison you will meet first. The alternative is an AI center of excellence, a single central team that builds and approves most AI work itself. Be honest about the evidence: no rigorous, non-survey study shows one structure beats the other. What McKinsey's March 2025 survey does show is a pattern in what firms centralize. Risk, compliance and data governance are often fully centralized. Tech talent and adoption are more often hybrid. Firms under $500 million in revenue centralize more.
That is a sensible default: centralize what one standard protects, distribute what local knowledge decides. Then write the split down, because the argument you are trying to prevent starts "I thought that was your team's job".
What is the 10-20-70 rule?
The 10-20-70 rule is BCG's rule of thumb for where AI effort should go: 10% on algorithms, 20% on technology and data, and 70% on people and processes. BCG has used it since at least October 2019, when its report How to Win with Artificial Intelligence told companies to consider AI "10% about algorithms, 20% about technology, and 70% about business process transformation". Its October 2024 study restated the split as how AI leaders allocate their resources, with the 70 described as people and processes. In the boat, the motor is the 10, the fuel line and mounting bracket are the 20, and the seating, steering and retraining of the crew are the 70.
Treat it as a budget check, not a law of nature. It comes from a consultancy that sells transformation work, and nobody has shown that 70 is the right number for your firm. Use it as a mirror: if your spending on people and process is close to zero, you do not need a precise target to know something is missing.
Why can't HR simply sign the new roles on the day?
Because in much of Europe, a change to how work is organized is not the employer's alone to finalize. Where a works council, union or collective agreement applies, employee representatives have a legal right to be informed and consulted before the change takes effect, and in some areas to co-decide. AI change management that skips this step does not save time; it produces a go-live date that moves, or a dispute.
Three layers matter to an EU transformation lead:
- The EU floor. Directive 2002/14/EC of 11 March 2002 sets minimum rights to information and consultation in undertakings with at least 50 employees or establishments with at least 20 (each member state chooses which). Article 4 covers "decisions likely to lead to substantial changes in work organisation", requires information at a time that lets representatives prepare, and frames consultation as "with a view to reaching an agreement".
- National law, which is often stronger. Germany's Works Constitution Act (official English translation, including the amendments of 19 July 2024) requires the employer to inform the works council of plans for "working procedures and operations including the use of artificial intelligence" and to consult it "in good time" (section 90). The works council co-determines "the introduction and use of technical devices designed to monitor the behaviour or performance of the employees" (section 87(1) no. 6), and where it must assess the introduction of AI, calling in an expert is deemed necessary (section 80(3)). Other member states have their own regimes; check each country with employment counsel.
- The AI Act for high-risk workplace systems. Under Article 26(7) of the EU AI Act, deployers who are employers must inform workers' representatives and the affected workers before putting a high-risk AI system into use at the workplace. Annex III point 4 classes as high-risk systems intended to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate workers' performance and behaviour. After the Digital Omnibus on AI (Regulation (EU) 2026/1744), Annex III duties apply from 2 December 2027. Confirm the details against the Official Journal text.
The consequence is about timing. A charter session can draft every role change but cannot finalize one, because consultation must happen while the decision can still change, and its length is not yours to set. So the session ends with a draft signed by HR, a written proposal for the representatives, and a go-live date after consultation is expected to close. Outside the EU, check collective agreements for change-of-work clauses before you set a date.
What does a working model look like?
Seven moves, in this order. Treat them as your AI adoption framework; the workshop makes them for one workflow.
- A sponsor at CEO level, and a named value owner for each business domain.
- The hub owns platform, risk, standards and training. The spokes own use cases and P&L outcomes.
- Redesign the end-to-end workflow before you automate individual steps.
- Rewrite the roles and the review points, and consult on them where employee representatives have rights.
- Build an AI champions network and role-specific training.
- Keep a KPI ledger and review it quarterly.
- Make fewer, larger bets.
On the first move: in McKinsey's March 2025 report (July 2024 survey), 28% of respondents said their CEO oversees AI governance, among the attributes most correlated with bottom-line impact. Treat that 2024 figure as dated. The point stands: someone senior enough to move budget between departments has to chair this.
The Workshop: Run an Operating Model Charter Session
The method has three parts: one working day to draft the charter, a consultation period wherever employee representatives have rights over the role changes, and a 60-minute review every quarter after go-live. It produces one document, the AI Operating Model Charter. Do it for a single workflow, not the whole company.
Who needs to be in the room?
Eight people, each holding a decision nobody else can make, plus one conversation that happens before the day:
- The executive sponsor, who chairs. Often the COO, CIO or a chief transformation officer acting for the CEO.
- The head of AI, speaking for the hub.
- The HR business partner for the affected teams, because roles will change and HR owns job descriptions, grading and the consultation calendar. With works councils or recognized unions, bring the employment relations specialist too.
- A finance partner, to move money and to sign the outcome measures.
- The business unit head who owns the workflow, who will leave the room as its value owner.
- One frontline manager and one practitioner from that workflow. They know what actually happens on a busy day, which the process map does not.
- Risk or compliance, to say which human checks are fixed and whether any system in scope is high-risk under the AI Act.
The conversation before the day: HR tells the works council chair or union representative in writing that a change to work organization is being prepared, and offers an observer seat. An observer is good faith, not consultation, which starts later on a written proposal.
Bring six inputs: the paved-road charter from Lesson 16, the permission matrices from Lesson 17, the runbooks from Lesson 18, the value ledger from Lesson 19, the current process map and job descriptions for one workflow, and the collective and works council agreements covering the affected roles.
What are the six steps?
- Name the owners. Write down the executive sponsor. For each business domain, name one value owner, accountable for the outcomes in the ledger: a person, not a committee. Add the succession rule now: if a value owner leaves, the sponsor names an interim individual within a week, because a "leadership team" holding the line is an orphan with a nicer title.
- Draw the hub and spoke split, and the funding rule. List ten to fifteen decisions and assign each to hub, spoke or shared. Head the table with one sentence: the hub enables and sets standards; it is not a queue. Then decide who pays for the platform and who pays for use cases.
- Map one workflow, then cut a first redesign. Choose an end-to-end workflow the ledger shows is worth scaling. Map every step, elapsed time and handoff, and let the practitioner correct the map. Then work backwards from the outcome at the level of who does what: which handoffs disappear, which work moves to AI, and where the Lesson 17 approval gates stay. Stop there; the full rebuild belongs to Lesson 21.
- Draft the role changes and plan the consultation. For each affected role, write what the person stops doing, starts doing and now reviews. Add four lines most sessions forget: any effect on grade or pay, the redeployment or reskilling path for freed capacity, the performance objective that rewards the new way of working, and whether any new tool monitors individual performance. HR signs the draft, not the final version. Where a works council, union or collective agreement applies, the draft goes to consultation before go-live, and the charter records that consultation is required, the date it starts and the employee representative's name.
- Test the budget against 10-20-70, and fund the 70. Sort every line of this workflow's budget into algorithms, technology and data, or people and process. Look for what the 70 is missing: training hours, process work, manager time and consultation hours. Pick champions from practitioners, not only managers, and set role-specific training on this workflow's real cases, built on the role-based plan from Lesson 4 on AI literacy that sticks.
- Set the measures, the review and the stop list. Choose adoption and outcome measures that describe the workflow, not the licence count, and add them to the value ledger. Fix the first quarterly review date and its chair. Then list the initiatives you will stop to fund the larger bets.
What does it look like filled in?
Here is an illustration; everything in it is invented. Picture a 3,000-person wholesale distributor with offices and warehouses in Germany and the Netherlands, whose ledger shows that supplier onboarding is slow, costly and a good candidate for scaling. Procurement is the spoke. The German entity has a works council.
Step 2, the decision table. The group assigns the decisions that keep causing arguments.
| Decision or activity | Hub | Spoke (procurement) | Shared |
|---|---|---|---|
| Gateway, logging, model approval | Owns | Uses | |
| Risk classification and intake review | Owns | Supplies facts | |
| Choice and funding of use cases | Advises | Owns | |
| Workflow redesign and role changes | Facilitates | Owns | HR signs drafts; consultation where required |
| Golden dataset and expert review | Provides harness | Owns content | |
| Outcome metric in the value ledger | Reports | Accountable | Finance signs |
| Training | Core curriculum | Role-specific practice |
The hub does not own the choice of use cases, the redesign, or the outcome. The head of procurement is accountable for the number, and finance signs it so nobody marks their own homework.
Step 3, the first cut. An onboarding clerk walks the room through what happens today. The group then starts from the outcome ("an approved supplier, correctly set up, quickly") and decides who does what.
| Workflow step (before) | After the first cut | Human role |
|---|---|---|
| Buyer emails supplier for documents | Agent requests and chases documents | None, logged |
| Clerk keys data from documents | Agent extracts data, flags low confidence fields | Clerk reviews flagged fields only |
| Compliance checks each file | Agent pre-checks against rules | Compliance approves every new supplier (gate kept) |
| Manager approves by email | Approval in one queue with the trace attached | Manager approves exceptions |
The chasing step loses its human but stays logged. The clerk moves from typing to reviewing only the fields the agent is unsure about. One gate does not move: compliance still approves every new supplier, as the Lesson 17 permission matrix required. Redesign is not a contest to remove people; it decides where a person's attention is worth most.
Step 4, the role change sheet. For the onboarding clerk, the room writes: stops keying data from documents; starts handling supplier queries that the agent escalates; now reviews flagged low confidence fields. Grade unchanged, pending HR's job evaluation. Freed capacity moves to supplier queries, with two days of practice on real escalations. The performance objective becomes "supplier records right first time", not "records processed per hour".
Then the HR partner raises the line that changes the plan. The new queue shows how many flagged fields each clerk clears per hour. In Germany, a technical device designed to monitor performance is a co-determination matter for the works council, and a per-person count is exactly the kind of feature that raises the question. The room can report the count by team, or put the individual view into consultation. It chooses team-level reporting and writes that into the proposal. Risk notes that an individual view could also bring the queue under Annex III point 4, which would add the Article 26(7) duty to inform workers; team-level reporting keeps that question out of version one. HR signs the sheet as a draft, the proposal goes to the German works council the following week, and the Dutch HR lead checks the local position with employment counsel. Go-live moves from the vendor's suggested date to after consultation is expected to close.
Step 5, the budget check. The finance partner sorts the workflow's budget into three columns. In this illustration the people and process column is empty: licences and integration, but no line for training, process change or consultation time. The session moves money from licences into a three-week process and training effort and books HR and manager hours for consultation, often the most valuable thing the day produces.
What goes in the AI Operating Model Charter?
Twelve fields. Copy this table, fill the right-hand column in the session, and have the sponsor sign it before people leave the room. The role change sheet stays marked as a draft until consultation closes.
| Charter field | What to write |
|---|---|
| Executive sponsor | One name, with authority to move budget across departments |
| Value owner per domain | One named person per business domain, plus the interim rule if they leave |
| Hub and spoke decision table | Ten to fifteen decisions, each marked hub, spoke or shared |
| Funding rule | Who pays for the platform, and who pays for use cases |
| Redesigned workflow map | Before and after steps, with every human review point marked |
| Role change sheet | Stops, starts and reviews, grade effect, redeployment path and performance objective for each role; signed by HR as a draft |
| Consultation required | Yes or no per country, start date, employee representative's name, expected close |
| Budget split against 10-20-70 | Each budget line sorted into the three groups, with gaps named |
| Champions list and training plan | Named practitioners, and role-specific practice for each role |
| Adoption and outcome measures | Workflow outcomes added to the value ledger, not licence counts |
| Quarterly review date | A date in the calendar and the name of the chair |
| Stop list | Initiatives that end so the larger bets are funded |
The funding rule is the field people skip. A workable pattern funds the hub's platform centrally and lets each spoke fund its own use cases, because a unit that pays for a use case behaves like its owner.
The stop list is the other field that gets left blank. BCG's October 2024 study noted that leaders pursue about half as many opportunities as their peers. You cannot fund the 70 on fifteen initiatives at once. Something has to end, and it is kinder to end it in a meeting than to starve it slowly.
Here is what it looks like when it works, continuing the illustration. Two quarters later the onboarding queue has one owner, a number finance trusts and a role description the clerks recognize, and the works council agreement on team-level reporting is filed with the charter. For the next workflow, the decision table, funding rule and consultation template already exist, so the second session takes half the time. That reuse is what "scaled" means.
Failure Modes, How to Spot Them, and What to Measure
Seven things go wrong, and each has a tell you can check in under an hour.
| Failure mode | How to spot it | What to do |
|---|---|---|
| Tool on an old process | The process map looks the same before and after | Rerun step 3 from the outcome backwards |
| Gatekeeper hub | Business units describe the central team as "approval", and requests queue | Revisit the decision table; move use case choice to the spokes |
| Orphan outcomes | No business owner's name sits against the value in the ledger | Name a value owner or pause the initiative |
| Consultation as an afterthought | A go-live date exists before employee representatives have seen a written proposal | Move the date; send the draft role sheet as the proposal |
| The 70 unfunded | The budget has lines for licences and compute and none for training, process change or consultation time | Reallocate before launch, as in step 5 |
| Executive optimism | Leadership surveys are positive while frontline strain is high | Ask the practitioners separately and compare |
| Cost-only redesign | Headcount targets are set before quality measures | Set the quality measure first, then the cost target |
Consultation as an afterthought usually starts with a vendor timeline drawn before HR saw the role changes. Ask to see the written proposal that went to the employee representatives, and its date. If there is none, the go-live date is a hope.
Executive optimism hides well. McKinsey's August 2026 survey put negative AI-related strain at 31% among executives and 47% among mid-level managers and individual contributors. Survey only the steering committee and you measure the calmer group, so survey practitioners and executives separately and compare.
How do you get people on board without a big campaign?
Involve the people who do the work in building and judging the system. An OpenAI customer story about Morgan Stanley (undated, and a vendor source) describes advisors grading the evals before release, and reports that over 98% of advisor teams are active. The modest inference: asking domain experts to grade the tests doubles as change management, because people trust what they helped to grade.
Which metrics should you track?
Six, reviewed at the quarterly session:
- Workflows fundamentally redesigned. A count of end-to-end workflows, not tools deployed. McKinsey's August 2026 benchmark: nearly three-quarters of high performers.
- Share of initiatives with a named business value owner. Anything below 100% is a list of orphans.
- Share of AI budget spent on people and process, compared with the 70 guide.
- Workflow outcome change after redesign, taken from the value ledger.
- Consultation lead time, from written proposal to agreement, per country, which tells you how early the next session must start.
- Strain and confidence, reported by practitioners and by executives, surveyed separately.
Notice what is missing: licences, logins, prompts sent. Those tell you the motor is running, not whether the boat is going anywhere.
Frequently Asked Questions
Who should own AI in a company: IT, data or the business?
The business owns the outcomes, and a central hub, usually under the CIO, CDO or head of AI, owns the platform, risk controls, standards and training. Write that split into a decision table and name one value owner per domain, or the central AI team inherits accountability for results in departments it does not run.
Should AI be run from a center of excellence or a hub and spoke model?
Choose by size and maturity and write the split down, because no rigorous study proves one structure beats the other. McKinsey's March 2025 survey found risk, compliance and data governance are often fully centralized while tech talent and adoption are more often hybrid, and firms under $500 million in revenue centralize more.
How much of an AI budget should go to people and change management?
Roughly 70% of the effort, if you use BCG's 10-20-70 rule of thumb, in its publications since at least 2019: 70% on people and processes, 20% on technology and data, 10% on algorithms. Treat it as a mirror rather than a target: if training hours, process work, manager time and consultation time have no budget line at all, move money before launch.
Do we have to consult the works council before AI changes people's jobs?
Usually yes, where a works council, union or collective agreement applies. Directive 2002/14/EC gives EU employee representatives information and consultation rights on decisions likely to lead to substantial changes in work organisation, and Germany's Works Constitution Act requires consultation in good time on working procedures, explicitly including the use of artificial intelligence. Treat role changes from the charter session as drafts and confirm each country's rules with employment counsel.
Do AI champions networks actually work?
They are common practice, but no measured study of their effect has been found. Use them anyway, choose champions from practitioners rather than only managers, and judge the network by workflow outcomes rather than launch attendance.
Key Takeaways
- An AI operating model is four written answers: who decides, who builds, who pays, and who is accountable. Unwritten, all four default to the central AI team.
- Design the work, not the headcount and not the tool. Each domain needs one named value owner, and an interim rule for when that person leaves.
- In a hub and spoke model the hub provides platform, standards, risk controls and training; the spokes own use cases and outcomes. The hub is not a queue.
- Cut a first redesign of one workflow at the level of who does what, keep the human gates that risk requires, and leave the blank-page rebuild to Lesson 21.
- Role changes signed in a one-day session are drafts. Where works councils, unions or collective agreements apply, consult before go-live and record it in the charter.
- Use the 10-20-70 rule, in BCG's material since at least 2019, as a budget mirror: if training, process work and consultation time have no line, move money before launch.
- Measure redesigned workflows, named owners, consultation lead time and practitioner strain, not licences.
Skill.re