Designing for Humans Who Arrive Unfinished
Every speaking human arrives unfinished.
This is the central design principle.
It is not true of some users and false of others. It is not a matter of intelligence, age, emotional state, technical ability, personality, or preparation. It is the nature of live human speech.
Unless a person is reading from a script, teleprompter, prepared statement, memorized line, or written artifact, the speaker does not know the exact next sound wave that is about to come out of their mouth.
The human may know the general topic. They may know the feeling. They may know the urgency. They may know the complaint. They may even know the desired outcome. But the actual expression is still becoming.
That is the condition AI workflow designers must begin with.
Most workflow design begins too late. It begins after the human has already been forced into a category, field, form, menu, or request type. That was necessary in the old computer-interface world. It is not the right starting point for AI voice.
A properly designed AI workflow begins with emergence.
Not extraction.
1. The Beginning of the Workflow Is Not Extraction. The Beginning of the Workflow Is Emergence.
The first mistake in AI workflow design is treating the human’s first utterance as if it were a completed input.
It is not.
The first utterance is the opening movement of emergence.
The person begins speaking, and the meaning becomes clearer as the speech unfolds. The person may start with one concern and reveal another. They may begin with billing and reveal service. They may begin with a complaint and reveal a lost item. They may begin with a vague frustration and discover the actual issue only because the AI reflects something back.
This is not noise.
This is the signal.
Traditional workflow design tends to interrupt this signal too early.
It asks:
- What category is this?
- What department should receive it?
- What fields are required?
- Which branch should the caller enter?
- How do we route this as quickly as possible?
Those questions belong to the old interface mindset.
That mindset made sense when software could not hold ambiguity. Forms need fields. State machines need branches. Databases need structured entries. Traditional software needed the human to become structured as quickly as possible.
AI can do something different.
AI can hold unfinished human expression for a little while. It can converse. It can reflect. It can say, “Tell me what happened.” It can allow the person to continue. It can let the real artifact begin to reveal itself.
That is the magic of AI voice.
The workflow designer does not need to add this capability.
The workflow designer needs to stop removing it.
Do not prompt-engineer the AI into a phone tree.
Do not force category selection too early.
Do not strip out brief reflection.
Do not over-optimize for shortness at the beginning.
Do not turn the AI into a spoken form.
Do not interrupt emergence in the name of efficiency.
The practical instruction is simple:
Let the human talk.
Not forever. Not aimlessly. Not as therapy. But long enough for the actual artifact to begin revealing itself.
Extraction asks, “What fields do I need?”
Emergence asks, “What is this human expression becoming?”
At the beginning of an AI workflow, emergence must come before extraction.
A question that is helpful after emergence may be harmful before emergence.
For example, after the artifact is clear, “What kind of phone was it?” is a useful question.
But before the caller has finished arriving, that same question may collapse the conversation too early.
Fields are not bad.
Categories are not bad.
Routing is not bad.
They are bad when they arrive too early.
Before the artifact is visible, fields feel like interruption.
After the artifact is visible, fields feel like help.
2. Design Around Artifacts, Not Departments
Departments are where unresolved things are sent.
Artifacts are what unresolved things want to become.
Most workflows are designed around routing:
- Send this to billing.
- Send this to the manager.
- Send this to support.
- Send this to sales.
- Send this to legal.
That is not wrong, but it is not the best starting point for AI workflow design.
The better question is:
What artifact is this human expression trying to become?
A lost-item call may need to become:
- A lost-item report.
- A searchable item record.
- A manager alert.
- A verified found-item status.
- A pickup instruction.
- A closed claim.
A billing concern may need to become:
- A billing review.
- A receipt lookup.
- A pending-versus-posted explanation.
- A refund request.
- A manager approval.
- An issued refund.
A complaint may need to become:
- A complaint record.
- An issue classification.
- A customer desired resolution.
- A manager alert.
- A draft response.
- An approved remedy.
A business idea may need to become:
- A memo.
- A landing page.
- An offer description.
- Product copy.
- A proposal.
- An article.
- An image.
- A plan.
A software request may need to become:
- A technical spec.
- A first-pass script.
- A report.
- An automation.
- A test case.
- Working code.
This is the more useful design object.
Do not design the workflow primarily as:
Human → category → department → action item.
Design it as:
Human expression → artifact recognition → artifact formation → status → authority.
That is the AI-native workflow structure.
The workflow is not primarily about where to send the human.
It is about what the human’s expression is trying to become.
3. Do Not Constrain Arrival. Constrain Completion.
A well-designed AI workflow is wide at the beginning and narrow at the end.
Wide at the beginning because every speaking human arrives unfinished.
Narrow at the end because completion must be true.
This is the central asymmetry.
Most bad AI workflows reverse it. They over-engineer the beginning and under-engineer the end.
They force the human into rigid categories early, then allow the AI to speak loosely at the point of completion.
That is exactly backwards.
At the beginning, allow:
- Natural conversation.
- Ambiguity.
- Brief reflection.
- Open-ended explanation.
- The human to finish arriving.
At the end, require:
- Authority.
- Status.
- Verification.
- Precise language.
- Forbidden claims.
The beginning should feel like conversation.
The ending should behave like protocol.
The practical rule is this:
Do not over-engineer the human’s arrival.
Over-engineer the artifact’s authority boundary.
That is the proper use of design effort.
4. Let the AI’s Native Conversational Ability Work
The AI already knows how to receive a human.
It knows how to say:
- “Tell me what happened.”
- “Start wherever it makes sense.”
- “Let me make sure I understand.”
- “What I’m hearing is…”
- “That sounds like it may involve two separate issues.”
- “Keep going. I want to make sure I route this correctly.”
These behaviors are not decorative.
They are part of the translation process.
They allow unfinished human expression to become stable enough to translate.
Designers often remove the very behaviors that make AI valuable.
They say:
- Be concise.
- Do not repeat the user.
- Classify immediately.
- Ask only required questions.
- Move quickly through the workflow.
- Avoid open-ended conversation.
Those instructions sound efficient.
But they can destroy the bridge.
They can turn AI back into old software.
A poor system instruction says:
“Classify the caller’s intent immediately and gather the required fields.”
A better instruction says:
“Allow the caller to explain naturally before selecting a workflow path. Do not force category selection too early. Use brief reflection when it helps the caller’s meaning emerge. Once the likely artifact is clear, gather the missing details efficiently. Never claim authority-bound completion without verification.”
A poor instruction says:
“Be concise and avoid repeating the caller.”
A better instruction says:
“Be concise after the artifact is clear. Before that, allow brief reflection or restatement when it helps stabilize the caller’s meaning.”
A poor instruction says:
“Move the caller through the workflow quickly.”
A better instruction says:
“Do not rush the caller into structure. First identify what artifact their expression is reaching toward. Then proceed efficiently.”
The goal is not fewer words.
The goal is fewer wrong artifacts.
Early patience prevents late repair.
A few extra seconds of emergence can prevent the wrong ticket, wrong department, wrong fields, wrong promise, or wrong artifact.
5. Complete From Pattern. Verify From Authority.
Once the artifact is visible, the workflow must determine whether the artifact is pattern-bound or authority-bound.
Pattern is learned.
Authority is granted.
AI can complete from pattern.
AI must verify from authority.
Pattern-bound artifacts are artifacts AI can often form directly because they are governed by stable patterns.
Examples include:
- Lost-item reports.
- Complaint summaries.
- Meeting recaps.
- Follow-up emails.
- Contract drafts.
- Product descriptions.
- Website sections.
- Proposal outlines.
- Python scripts.
- Training documents.
- Manager alerts.
- Customer responses.
- Technical specs.
- First drafts.
Authority-bound artifacts cannot be completed from pattern alone.
They require a person, system, permission, signature, payment action, verification, or official state change.
Examples include:
- Found phones.
- Issued refunds.
- Confirmed reservations.
- Signed contracts.
- Seller acceptances.
- Bank transactions.
- Price approvals.
- Verified inventory states.
- Manager-approved remedies.
- Legal acceptances.
- Delivery commitments.
The AI can create a lost-item report.
It cannot say the item was found unless verified.
The AI can prepare a refund request.
It cannot say the refund was issued unless issued.
The AI can draft a contract.
It cannot say the contract is binding unless signed.
The AI can suggest an available reservation time.
It cannot say “you are confirmed” unless the system confirms.
The AI can generate a product image.
It cannot imply actual inventory unless inventory has been verified.
This is where workflow design must become strict.
Do not constrain the AI’s conversational emergence too aggressively.
Do constrain the AI’s authority claims aggressively.
6. Build Authority Maps, Not Just Flowcharts
A flowchart tells the AI where to go.
An authority map tells the AI what it is allowed to make true.
AI workflows need both, but authority maps are more important than most designers realize.
For each workflow, define:
- What can the AI create from pattern?
- What can the AI read from a system?
- What can the AI write to a system?
- What can the AI submit for approval?
- What can the AI approve on its own, if anything?
- What requires a manager?
- What requires a system of record?
- What requires a signature?
- What requires payment authority?
- What requires physical verification?
- What must the AI never claim unless verified?
Every serious workflow needs forbidden claims.
For lost items:
- Never say the item was found unless verified.
- Never say someone checked unless someone checked.
- Never give pickup instructions unless the item is confirmed.
For billing:
- Never say the refund was issued unless issued.
- Never say the second charge is only pending unless verified.
- Never promise manager approval.
For reservations:
- Never say confirmed unless confirmed in the reservation system.
- Never promise a seating time unless the system supports that promise.
For contracts:
- Never say accepted unless accepted by authorized parties.
- Never say signed unless signed.
- Never blur a draft with an agreement.
For inventory or delivery:
- Never say available unless inventory is verified.
- Never promise delivery unless an authorized party or system confirms.
These rules are more important than personality instructions.
They protect trust.
The AI does not need much help sounding conversational.
It does need help not falsely completing authority-bound artifacts.
7. Every Artifact Needs Status
A good AI workflow does not merely produce artifacts.
It produces artifacts with honest status.
Status is not bureaucracy.
Status is truthfulness.
Useful status labels include:
- Drafted.
- Prepared.
- Submitted.
- Routed.
- Pending verification.
- Verified.
- Approved.
- Issued.
- Confirmed.
- Completed.
These words tell the human where the crossing actually stands.
A draft is not a final.
A request is not an approval.
A report is not a verified fact.
A prepared refund is not an issued refund.
A generated image is not proof of a real object.
A contract draft is not a signed agreement.
A lost-item report is not a found phone.
The workflow must never allow the artifact’s appearance to outrun its truth.
Instead of:
“You’re all set.”
Say:
“I created the lost-item report and sent it to the manager. I do not yet have confirmation that the phone was found.”
Instead of:
“We’ll take care of the refund.”
Say:
“I prepared the billing review with the details you gave me. A manager must approve any refund before it can be issued.”
Instead of:
“The agreement is ready.”
Say:
“I drafted the agreement. It is not binding until both parties review and sign it.”
The better version may sound less magical.
It is more trustworthy.
8. False Completion Is the Main Failure Mode
False completion happens when the AI claims, implies, or emotionally suggests that an artifact exists before authority has made it real.
False completion is worse than awkwardness.
It is worse than a clumsy phrase.
It is worse than a slightly mechanical voice.
False completion destroys trust.
Examples of false completion include:
- “Your refund has been issued,” when only a refund request exists.
- “We found your phone,” when only a lost-item report exists.
- “You are confirmed,” when the reservation system was not updated.
- “The agreement is accepted,” when only draft language exists.
- “The item is available,” when inventory was not verified.
False completion tells the human the crossing is finished when the human is still standing in the middle.
It changes what the human does next.
They stop looking for the phone.
They wait for a refund that is not coming.
They show up for a reservation that was never made.
They rely on a contract that was never accepted.
This is why the completion boundary must be heavily engineered.
9. Return Humans at the Point of Judgment, Not First Formation
“Human in the loop” is too vague.
The better question is:
Why is the human in the loop?
If the human is only there to convert messy expression into first form, AI may be able to do much of that.
If the human is there for authority, judgment, taste, relationship, ethics, responsibility, or final decision, the human belongs there.
Bad human escalation:
- “Customer complained.”
- “Need agreement.”
- “Need spreadsheet comparison.”
- “Caller had billing issue.”
These are unresolved handoffs.
They force the human to do first formation.
Good human escalation:
“Complaint record created. Customer visited Friday around 8 p.m. Main issues: food delay, cold entrée, and possible billing concern. Customer complimented server and does not want punitive action. Billing concern requires verification. Draft callback prepared. Manager review required.”
“Draft agreement prepared with assumptions marked and open questions listed. Attorney review required before use.”
“First-pass spreadsheet comparison script created. Assumptions, expected behavior, and test cases included. Developer review required.”
This lets the human judge, approve, correct, or decide.
That is where human attention belongs.
10. The Five-Layer AI Workflow Model
10.1 Layer One: Arrival
The human speaks naturally.
The AI does not interrupt too early.
The workflow protects emergence.
Design goal:
Let the human arrive.
10.2 Layer Two: Emergence
The AI allows the meaning to unfold.
The model may reflect, clarify, restate, or ask open-ended questions.
Design goal:
Let the real issue appear.
10.3 Layer Three: Recognition
The AI identifies the artifact the expression is becoming.
Design goal:
Name the artifact before collecting fields.
Example:
“What I’m hearing is that this is partly a service complaint and partly a billing concern. I’m going to separate those so the manager can handle each one correctly.”
10.4 Layer Four: Formation
The AI completes what can be completed from pattern.
Design goal:
Create the artifact.
Examples:
- Report.
- Draft.
- Record.
- Summary.
- Script.
- Memo.
- Alert.
- Request.
10.5 Layer Five: Authority and Status
The workflow verifies what requires authority and reports the artifact’s true state.
Design goal:
Prevent false completion.
Example:
“I created the report and routed it to the manager. The item has not yet been verified as found.”
11. The Designer’s Checklist
Arrival
Are we allowing the human to speak naturally?
Are we forcing categories too early?
Are we stripping out useful reflection?
Are we treating the first utterance as a completed input?
Emergence
Does the AI have room to let the meaning unfold?
Can the AI detect when the issue changes or deepens?
Can the AI handle a caller who starts in the wrong place?
Recognition
What artifact is the expression trying to become?
Are we designing around artifacts or departments?
Does the AI confirm the emerging artifact before extracting fields?
Formation
Which artifacts are pattern-bound?
Can the AI form them directly?
Are we avoiding unnecessary action items?
Authority
Which artifacts require authority?
What system or person makes the state true?
What must be verified?
What must be escalated?
Status
What status labels are available?
Can the AI clearly distinguish drafted, submitted, verified, approved, issued, confirmed, and completed?
Forbidden Claims
What must the AI never say unless verified?
Where is false completion most likely?
Human Return
Where does human judgment belong?
Are humans being used for authority or merely first formation?
Can the AI send humans formed artifacts instead of vague tasks?
12. The Short Version for Designers
All speaking humans arrive unfinished.
Therefore, do not begin with extraction.
Begin with emergence.
Let the AI’s native conversational ability work.
Do not collapse the human into categories too early.
Design around artifacts, not departments.
Once the artifact is visible, complete what pattern allows.
Verify what authority requires.
Every artifact needs honest status.
Prevent false completion.
Return humans at the point of judgment, not first formation.
The simplest version is this:
Let them talk.
Let the artifact emerge.
Complete from pattern.
Verify from authority.
Tell the truth about status.
That is proper AI workflow design.