We optimize tasks, not a system
Factories that swapped steam for electric motors gained almost nothing for decades. The gain came when the factory floor was redesigned (Agrawal, Gans and Goldfarb).
Learn
The model, from friction to flow
AI Mapping does not reinvent service design. It adds what is missing to decide where AI goes: a flow layer, five decision criteria and an idea portfolio.
01
Why AI does not shorten the queue
The numbers all tell the same story. In the second quarter of 2026, 19.2% of Canadian businesses used AI, three times more than in 2024 (Statistics Canada). Eighty percent of Canadian nonprofits use it in some form, and only 10% have a formal policy (Imagine Canada, 2026).
Yet nearly 9,000 executives in seven countries estimate that AI added about 0.2% per year to their company’s productivity over three years (NBER, 2026). In Denmark, among 25,000 workers in exposed occupations, time saved averages 2.8% of hours, with no measurable effect on wages or hours worked (Humlum and Vestergaard, 2025).
AI goes where it is easy to see
Communications, the website, emails. Rarely where work gets stuck.
Time saved goes nowhere
Twenty minutes saved by someone who is not the bottleneck become twenty minutes of queue a little further on. Or rushed work that lands on a colleague’s desk.
02
A service’s throughput is set by a single step
Everything is a service, even in a business that sells products. A client waiting for a quote, an invoice to send, a technical support request: every request goes through a series of steps. One of them limits the result of all the others. That is the bottleneck.
The idea has a name: Eliyahu Goldratt’s Theory of Constraints (The Goal, 1984). It comes down to four practical consequences.
12Request
9Qualification
4Writing
6Review
8Sending
Fictional example: service proposals in an SME, capacity per week. The firm sends 4 proposals a week, not one more.
An hour saved elsewhere is a mirage
An hour saved at the bottleneck is an hour saved for the whole system. Speeding up an upstream step swells the queue.
The bottleneck is often a rule
A complete form before starting, a mandatory signature, a habit from another era. A rule can change without a budget.
Waiting can be calculated
Little’s law, valid on average when the system is stable: wait = queue ÷ throughput. Sixteen requests waiting, four per week: four weeks.
The bottleneck can be demand
If your team is waiting for clients, the constraint is marketing. AI helps there too: campaigns, content, analysis, follow-ups. And each flow (selling, delivering, invoicing) has its own bottleneck.
« Technology can bring benefits if, and only if, it diminishes a limitation. »
Goldratt added a condition almost nobody meets: also change the rules that had adapted to that limitation. If AI drafts a note in thirty seconds, but the same line-by-line review is still required, nothing has changed.
03
The usual maps show friction, not flow
Service design gave organizations two valuable tools. The journey map follows a person through the service and shows where it hurts. The service blueprint goes into internal processes and shows what the organization does to deliver each step.
Both share a blind spot: time. As practiced today, neither shows queues, capacity or throughput. Irony of history: Lynn Shostack, who invented the service blueprint in 1984, timed every step. AI Mapping brings the stopwatch back.
The service blueprint, with a flow layer
- 1
The journey
What the person lives
Their efforts, their waits, their low points. Built in workshops with the people who live the service, from realistic personas, not among managers.
Friction - 2
Internal processes
What the organization does
Staff and support, plus two lanes nobody draws: external requirements (clients, insurers, funders, government), and AI as a full actor.
Causes - 3
The flow
How work moves over time
For each step: actual work time, total lead time, queue, capacity, and the share of work that arrives usable at the next step.
The bottleneck
The third layer almost always holds a surprise. In services and offices, the time a file actually moves typically represents 1 to 10% of its total lead time (Martin and Osterling, 2013). Most of the delay is waiting and rework at handoffs. AI that speeds up work anywhere but the bottleneck misses the target.
An example: a quote request
The same case as the online tool: an SME writing its proposals. Draw the steps as columns, then fill in the three layers, top to bottom, with the team.
- 1
The journey: the client drops off at writing. Four weeks without news is the low point of their experience.
- 2
Internal processes: at the same step, prices are copied from one software tool to another, and management must approve everything.
- 3
The flow: 16 requests wait in front of writing, versus 3 at most elsewhere. All three layers point to the same step. That is the bottleneck, and that is where AI belongs, after questioning the approval rule.
04
The biggest gain often comes from a changed rule
Documented results in public services, health care and business, client-facing and internal. The main lever was not technology: it was a rule. That is exactly the point.
Utah
Workforce services
Decision time
15 days11 days
Caseloads more than tripled, budget cut by over US$9M. The bottleneck: eligibility workers constantly interrupted by "where is my file?" calls and working on incomplete files. The fix: remove non-priority tasks and require a complete file before it reaches them. The approach was later extended to 24 state agencies.
Governing, 2012
New Zealand
Chemotherapy clinic
Average patient wait
129 min16.5 min
The bottleneck was a rule: confirm each patient before preparing their drugs. By changing it, the clinic also cut nurses’ overtime by 67%.
Health Systems, 2017
Ford
Accounts payable (internal service)
Accounts payable staff
500≈ 125
About 500 people matched purchase orders, receipts and invoices. The bottleneck was a rule: pay on invoice. Ford changed it to pay on receipt of goods, once the order was confirmed. No more invoices to match, and far fewer errors to fix.
Hammer, Harvard Business Review, 1990
IBM Credit
Client financing approval
Response time to the client
6 days4 hours
Every request went through four specialists, one after the other. The salesperson waited, and the client had time to shop elsewhere. The fix: one person handles the request from start to finish, and specialists step in only for unusual cases.
Hammer and Champy, 1993
Health care
42 implementations
Average wait
−50%
A review of 42 Theory of Constraints implementations in health care reports on average 50% less waiting and 34% more throughput, most often at low or no cost. The authors flag a likely publication bias. The direction of the result is consistent.
Bacelar-Silva et al., 2022
When AI comes afterward, it works on a bottleneck already cleared of useless rules. Then it makes a real difference.
05
Five criteria before AI
Every AI idea goes through the decision matrix. Five criteria, in order. A criterion not met does not mean "never". It means "not now, and not like this".
The questionIf the criterion is not met
- 1FailureDoes this demand exist because we failed upstream?Fix the cause.
- 2RuleIs the bottleneck a rule rather than a resource?Change the rule first.
- 3LimitationWhich limitation, at the bottleneck, does AI reduce? Which rules must change with it?Personal comfort, not a project.
- 4FrontierIs the task inside the zone where AI is reliable? Who checks, and how long does it take?Narrow the scope or add a check.
- 5HumanDoes the step decide for the person? Does it build the relationship?AI stays in support, under real human control.
Criterion 4 deserves special attention
In a study of 758 consultants, the same AI tool improved quality by more than 40% on some tasks, and cut the share of correct answers by 19 points on a neighboring task (Dell’Acqua et al., 2023). And AI that produces faster often moves the bottleneck to the person who checks.
Criterion 5 rests on Quebec’s framework
Quebec’s Law 25 governs decisions based exclusively on automated processing and requires a privacy impact assessment for any new system handling personal information. Token human validation is not enough.
06
The idea portfolio: where to act, where to protect
Retained ideas are placed on two axes: the effect of the step on the bottleneck, and its relational load. The four quadrants are the same as in the AI Canvas matrix. Only the first axis changes: effect on the bottleneck replaces level of irritation.
highRelational loadlow
Touchpoints
Preserve
They do not limit throughput, and they build what brings people back. Nothing to optimize here.
Discovery call, kickoff meeting, informal exchanges.
Moments of truth
Augment, never replace
The moments when the person judges the service. Relationship time is often the bottleneck itself. AI works around it, never in its place, and the time freed goes back into it.
Preparing the meeting, keeping the record, following up.
Routine support
Equip
Useful as self-service, framed by a usage policy. Not a transformation priority, even though most current AI use sits here. Unless the bottleneck is demand: then campaigns and posts move to the top of the list.
Better-written emails, posts, competitor watch, formatting.
Critical internal processes
Automate first
Work the person never sees, but that sits on the bottleneck’s path. AI can go in decisively, provided the rules change and verification is planned.
Invoicing, double entry, searching for information, incomplete files.
lowEffect on the bottleneckhigh
One rule sits above all four quadrants: no AI deciding who deserves help.
07
Three traps the method avoids
Automating failure
A chatbot that answers "where is my file?" faster industrializes failure. Volume goes up, the cause disappears from view, and the dashboard shows success. In a British municipal housing call center, John Seddon measured that 80% of calls were failure demand.
Sorting people
Austria, Poland and Flanders used algorithms to classify job seekers. All three went badly. In Poland, caseworkers could override the ranking: they did so in about one case in a hundred, mostly for lack of time.
Compressing the relationship
In Quebec’s health network, Lean was sometimes used to compress work: a home visit standardized at 30 minutes when it used to take two hours (IRIS, 2024). We measure the work, never the people. And freed time has a destination decided in advance with the team.
08
Measure outcomes, not minutes saved
A map or a prototype is an output. What matters is the outcome: mission throughput, the number of people who reach the result the service exists for.
Mission throughput, by profile
How many people reach their goal, without skimming the easiest cases.
Lead time at each step
Where people wait.
Failure demand rate
How much work we create ourselves.
Destination of freed time
The relationship, more people, or nowhere.
Next expected bottleneck
Where the constraint will go once this one is released.
The bottleneck will move. That is the sign you acted in the right place. The iteration review spots it, and the cycle starts again.
Glossary
- Journey map (Carte de parcours)
- The steps, efforts and emotions of a person through the service.
- Service blueprint (Plan de service)
- The journey, plus everything the organization does internally.
- Flow layer (Couche de flux)
- Lead times, queues, capacity and rework added to the blueprint.
- Bottleneck (Goulot)
- The step that limits the throughput of the whole service.
- Failure demand (Demande de défaillance)
- Demand created by our own failures upstream.
- Key insight (Insight clé)
- The finding that steers what comes next; here, the bottleneck.
- Decision matrix (Matrice de décision)
- The five criteria that filter AI ideas.
- Idea portfolio (Portefeuille d’idées)
- Retained ideas, placed on two axes.
- Wizard of Oz (Magicien d’Oz)
- A prototype where a person plays the role of the AI.
- Output, outcome (Livrable, retombée)
- What we produce, and what it changes for people.
A family of free tools
Where, what, how
Three KODRA tools, at three scales. They work together.
Where to act?
AI Mapping
The whole service. Find the step that limits everything, before choosing anything.
What to build?
AI Canvas
One process. Nine boxes to design a solution starting from the real work.
ai-canvas.orgHow to delegate?
The AI In-Basket
One task. A team game to decide what a person hands to AI, and what they keep.
ai-basket.orgAll three tools are free (CC BY-SA 4.0) and published by KODRA Conseil, which trains and coaches Quebec organizations in AI adoption.
From reading to your own service
The online tool finds your bottleneck in two minutes. The templates are for the workshop.