Organization Design for AI · Whitepaper
The CLEAR Framework: Modeling Processes When AI Does Part of the Work
Five roles, two rules and one foundation for processes in which people and AI agents work side by side.
9 min read
October 13, 2026
HandsOn Insights
In brief. Flowcharts, swimlanes and RACI tables all rest on one assumption: every actor in the process is a person. That assumption no longer holds. AI systems now draft the email, update the CRM and file the task. The CLEAR Framework replaces RACI’s four letters with five roles and two rules, puts a Context Layer underneath, and puts time on every step. This article summarizes our whitepaper. The full version with all exhibits is available for download at the end.
Where the AI step disappears in a process model
In a classic process model, an AI step usually ends up in one of two places. Either it disappears inside a human’s box (“Sales prepares the proposal”, even though an assistant writes most of it), or it gets a box of its own labeled “System” with no one attached to it.
Both hide the question that now matters most: who is accountable when the machine does the work? Companies bringing in AI one tool at a time feel the result in two failure modes.
Failure mode 01
Paralysis
Nobody dares to let the AI do anything meaningful, because nobody knows who would answer for a mistake.
Failure mode 02
Drift
The AI quietly does more and more, and accountability dissolves along the way.
A process model that includes AI has to answer three questions for every single step. Classic models were never built to answer them.
Question 1
Who leads?
Which person owns the outcome, however much of the work a machine does?
Question 2
Who or what executes?
Is this step done by a person, an AI agent, or both together?
Question 3
Where must a human sign off?
At which points does work stop until someone checks it?
Five roles, two rules
CLEAR keeps what made process models useful: a plain table anyone can fill in, one row per process step. What changes is the columns. Every step is modeled with five roles, and each role states whether a person, an AI or either may hold it.
C · Human or AI
Consults
Gives input to the work: a colleague, an expert or an agent. The data itself sits in the Context Layer.
L · Always a human
Leads
Owns the outcome. Exactly one per step.
E · Human or AI
Executes
Does the work.
A · Human where it matters
Approves
The gate: nothing moves on without this sign-off.
R · As designed
Reviews
Checks finished work afterwards, often as a sample.
Two rules apply to every model, with no exceptions:
Rule 1
01
Leadership is human-only.
An AI system can execute, consult or pre-check. It never leads a step. Someone with a name is always accountable for the outcome.
Rule 2
02
No AI chain runs without a checkpoint.
When one agent hands work to another, there is either a human touchpoint or a defined, logged machine check between them.
Rule 1 is good management, and it is also the core of what the EU AI Act means by human oversight. CLEAR builds it into every process step. Within the HandsOn AI Operating Model, CLEAR is the design instrument of domain D05, Process & Workflow Architecture.
The foundation: the Context Layer
Roles tell you who does what. They don’t tell you whether the work can be done well. A person can ask a colleague, notice that a spreadsheet is outdated or remember what the client said last week. An AI agent can only work with what it can reach through your systems. That is why every CLEAR model sits on a Context Layer, the system layer underneath the process.
Information becomes context when a machine can find it, trust it, and is allowed to use it.
In practice, that means five things:
Component 01
Source of truth
For every piece of information, which one system holds the valid version? Agents read from there and never from copies.
Component 02
Labels
Does every document and record show who owns it, how current it is and how confidential it is?
Component 03
Permissions
Can every agent access exactly what its tasks need and nothing more, traceable back to a responsible person?
Component 04
Connections
Do agents reach your systems through proper, controlled interfaces instead of exports, uploads and copy-paste?
Component 05
Write-back
Do the results flow back into the systems in a form the next step, human or AI, can use? If they don’t, the layer decays.
The Context Layer decides how much you can actually hand over. If an agent is meant to work on its own, the information it depends on has to meet all five points. Until it does, a human checks every output. That gives you one of the most useful sentences in any AI project:
“Your process model says the agent can do this alone. Your systems say a human still has to check every result.”
The CLEAR Framework · Context Layer
The gap between the two is your roadmap, and it usually lies in the data and systems.
The time dimension: where work waits for a person
A process model that only shows who does what still misses the question every managing director asks first: how long does it take? So in CLEAR, every step carries two kinds of time. Working time is when someone, human or AI, is actually doing the work. Waiting time is when the work sits in an inbox, a queue or a to-do list until someone gets to it.
Add both up across all steps and you get the lead time of the process, from trigger to result. In most processes, waiting time is by far the larger part. The work itself takes minutes; the waiting takes days. CLEAR compares lead time in three states.
State 1
Today
The process as it runs now, with its real working and waiting times.
State 2
Target state as planned
AI executes wherever the model allows it. Working time shrinks; waiting at human gates remains.
State 3
Target state as permitted
What the Context Layer allows today. Where it isn’t ready, human checks and their waiting time come back in.
That comparison shows something AI enthusiasm usually hides. AI shortens working time dramatically: a draft that took an hour takes a minute. It barely touches waiting time at the steps that still need a person. Every Approve is a point where work stops until a human gets to it.
If that person is in meetings all day, an agent that finishes in seconds still waits for hours.
So the lead-time view tells you exactly where the human bottlenecks are:
Bottleneck 01
Approval queues
Too many outputs need sign-off from too few people.
Bottleneck 02
Concentration
All approvals land on one desk, typically the boss.
Bottleneck 03
Handoffs and batching
Work waits for the weekly meeting, the next shift or the end of the month.
Bottleneck 04
Office hours
The agent works at night. The approver doesn’t.
The fix is better design. A review deadline keeps gates from turning into queues. Where the risk allows it, a Review afterwards replaces an Approve before. Approvals arrive bundled, in the channel the person already uses. And each phase gets one checkpoint at most.
There is one more honest step. The target state you plan and the target state you are allowed to run are not always the same. So CLEAR always calculates the target state as permitted.
How modeling with CLEAR works
Modeling a process with CLEAR takes five passes, from process steps to business case.
Pass
01
List the steps from the outcome.
Many current steps only exist because people work in batches, hand over by email or keep office hours. An AI-supported process doesn’t have to inherit them.
Pass
02
Assign the five roles to every step.
One Lead, always human. An Executor, human or AI. Consult, Approve and Review where they are needed.
Pass
03
Place the checkpoints on purpose.
Put the Approve gate where a mistake is expensive, irreversible or visible to a client. Everywhere else, a lighter Review is enough.
Pass
04
Check the context underneath.
For every step an AI executes, ask where its information comes from and whether it meets the five Context Layer points. Where it falls short, keep a human gate in place and put the gap on the roadmap.
Pass
05
Put time on every step.
Estimate working and waiting time today and in the target state. Compare the lead times and mark where work waits for a person. Those are your bottlenecks, and your business case.
A live example: modeling an AI chief of staff
The client was an owner-led agency: campaigns, content production and talent management running in parallel, with one managing director keeping it all together. The MD’s day was split across four systems: email and calendar, a CRM for the sales pipeline, a project tool for the team’s production tasks, and a wiki of briefs and playbooks.
The goal was an AI chief of staff: an agent the MD talks to over a messenger, which reads across all four systems, briefs the day, keeps track of deadlines and handles the admin. The tempting approach is to connect the agent to everything and let it help. We mapped the systems first and modeled the process second. This is the resulting CLEAR matrix:
| Step | C · Consults | L | E · Executes | A · Approves | R · Reviews | Reads from | Writes back to |
|---|---|---|---|---|---|---|---|
| Morning brief: what’s about to slip | – | MD | Agent | – | MD, while reading it | Calendar, CRM, project tool, wiki | – |
| Capture a personal reminder | – | MD | Agent | – | MD, via the agent’s confirmation | – | Calendar |
| Create a production task for the team | Project lead (capacity) | MD | Agent | – | MD, via the agent’s confirmation | Brief in the wiki | Project tool |
| Log a meeting summary in the CRM | – | MD | Agent | – | MD, via the agent’s confirmation | Meeting notes | CRM |
| Draft an email to a client | Project lead (status) | MD | Agent drafts | MD sends | – | CRM history, past threads | Email outbox |
| Create a deal or move it in the pipeline | – | MD | Agent prepares | MD confirms | – | CRM | CRM |
| Update briefs and playbooks | Agent (flags outdated content) | MD | Team (humans only) | – | – | Project tool, CRM | Wiki |
MD = managing director. Shaded rows carry an Approve gate. Fictionalized from a small-organization engagement; client identity removed.
Six design decisions the model made visible
Decision 1
The MD leads everything, and that is the point.
In a small company, one person often holds L, A and some of the E. The value of the model then lies in drawing the line between human and machine clearly and putting the checkpoints in the right places.
Decision 2
Reversible work runs without a gate.
Reminders, internal tasks and meeting notes all have an empty A column. A mistake takes ten seconds to fix. So the agent just does it, then reports exactly what it wrote and where. That confirmation is the Review.
Decision 3
Anything that leaves the building or touches money has an Approve.
The agent drafts emails and the MD sends them. It prepares deal changes and the MD confirms with a short “ok”. The line sits where a mistake becomes expensive or visible to a client.
Decision 4
Some systems are read-only.
The agent reads from the wiki and may flag outdated content as a Consult. Only the team writes back to it. That one cell became a technical setting: read-only access.
Decision 5
The Context Layer came first.
Before assigning a single role, we set the source of truth for each kind of information: appointments in the calendar, deals in the CRM, tasks in the project tool, briefs in the wiki. The agent reads live with its own scoped access and writes every result back.
Decision 6
The bottleneck was the MD.
Putting time on the steps showed the pattern clearly. The agent finished in minutes. The waiting happened at the MD’s approvals.
30–45 min
Morning brief today
Clicking through four tools. In the target state the brief is ready before the day starts.
Days
Meeting summary today
Logged when there’s time, often batched at the end of the week. In the target state minutes after the meeting, no approval needed.
1 week
Deal update today
Updated in the weekly review. In the target state the same day, after the MD’s “ok”.
Illustrative orders of magnitude, based on estimates for a setup like this.
Where there is no Approve, lead time collapses from days to minutes. Where there is one, lead time now depends almost entirely on how fast the MD responds. So we designed for that: approvals come to the MD in the messenger they already use, and a short “ok” is enough. The gate stays, and the queue in front of it shrinks.
The process model became the agent’s configuration.
Every E cell defined what the agent is allowed to do. Every A cell became a confirmation step built into the tool. Every C cell named whose input the agent asks for. The Context Layer columns set which systems it reads from and writes back to. The agent’s specification came straight off the model.
What a CLEAR process model gives you
The whitepaper lists eight outcomes. These six carry the business case.
Outcome 01
One standard across every process
Sales, finance, operations and HR are modeled with the same five letters and the same two rules. You see at a glance where AI already works, where humans sign off and where the gaps are.
Outcome 02
Agents built from the model
Once a process is modeled, the agent’s job description is already written: the E cells are its tasks, the A cells its limits, the Context Layer its data access.
Outcome 03
Accountability that never dissolves
Every step has one named person who owns the outcome, however much of the work a machine does.
Outcome 04
Oversight where it matters
A sign-off on every output looks safe, but within weeks people approve without really looking. CLEAR places the Approve gate where a mistake would be costly and uses lighter Reviews everywhere else.
Outcome 05
A model that grows with trust
As an agent proves itself, a step can move from “MD approves” to “MD reviews afterwards”. The table stays the same. Only the checkpoint changes.
Outcome 06
A business case you can calculate
Because every step carries working and waiting time, the effect of a change can be measured: lead time today against lead time in the target state, as permitted by the Context Layer.
Fewer checkpoints, taken seriously, beat many that nobody reads.
Where to start
Pick one process where AI already does some of the work, or where you want it to. Write down the steps and ask five questions for each one:
01 Who leads?
02 Who or what executes?
03 Where must a human sign off?
04 Where does the information for this step come from, and can a machine find it, trust it and use it?
05 How long does this step take today, and how much of that is just waiting?
You’ll see quickly where accountability is clear, where it has quietly gone missing, and where your process is really waiting for a person. That model is the starting point for every agent you build afterwards.
Whitepaper · 18 pages · PDF
The CLEAR Framework
The full version with all exhibits: the five roles, the Context Layer, the lead-time comparison and the complete case example with its CLEAR matrix.
Next step
Model your first process with CLEAR
CLEAR is part of the HandsOn AI Operating Model. We model one process with you where AI already does part of the work, and show where accountability, context and time get lost.