Who should prepare this work?
The morning order pile
Three orders to group: ready work and waiting work.
Your team checks before use.
BUFFER / FOR ESTONIAN COMPANIES
Small AI teams. You make the call.
Buffer builds small AI teams for Estonian companies. AI prepares work, such as a morning order brief. You decide where it belongs, and your team checks the result before use.
Review a question before email. simo@buffer.ee
RATHER READ? / TWO INDEPENDENT PAPERS
Who should prepare this work?
The little shelf note
Four blue mugs on shelf B. Leave a count and reorder note.
Your team checks before use.
One paper is a complete visit. Unfold the whole source from a result line, then choose who would prepare the work. The other paper can wait.
Use Keep to save each task answer on this device. Your work notes, inquiries and the older lesson last for this visit; reload clears them. Your earlier saved work stays separate.
Private work and saved papers stay separate.
OPTIONAL / REAL OFFICE DECISION
Three stamps. Different questions.
Read why the Office joined one support shift, and what three dated checks actually showed. Reading is enough.
OPTIONAL / FICTIONAL PRACTICE · IN ENGLISH
Three small worktables. Which question is yours?
Go straight to:
You have earlier Website work saved on this device. Continue it in the archive.
Your private rule is saved on this device. Open your rule and next check.
Your notebook kept its pages.
SAVED ON THIS DEVICE / REAL OFFICE
Open the dated Office file ↘
Your choice file and its open question.
INSPECTION DESK / FICTIONAL PRACTICE
Good papers.
Bad rule?
You decide what must match. Send three made returns through your proposed check, then see what slipped past.
Fictional exercise. Nothing is sent. These results cover only the three made returns below.
01 / Build a proposed check
Choose any details. Only selected details must match. You can also try checking nothing.
The whole request
02 / What the check did
All three carry the same checked reply, payment record, queue priority and Office file. A convincing reply can still hide a changed request.
All tested rules
Carry the last tested details as a fictional suggestion. Your own task starts blank.
This made workshop proves no real customer return. An intact made return can pass without promising a bank date.
Built trials use their own save. Earlier practice and Office choices stay separate.
Your practice papers kept your marks.
Your private rule comparison stays separate. Open your rule comparison
SAVED ON THIS DEVICE / THE RETURN DESK
Two slips. Two places to pick up.
Each stamp belongs to its own public record. A newer, unrelated release leaves these slips alone.
Checking this task’s last decision…
DATE / CHECKING
Next: open the task’s dated card.
Read this task’s source ↗ Open this task's dated card Checking its public record…The 27 and 29 September checks cover selected Website states. They do not show that every claim is caught.
LAST CHECK / 29 SEPTEMBER 2026
Next: compare the two dated checks. No new answer is readable yet.
Read the last live check ↗Your private rule and email stay in their own steps. These slips contain public work only.
Earlier lesson · a source check and the workshop summary
Fictional example
The lamp workshop paper
12 desk lamps packed. New dimmer tested.
Marten’s next task is missing from the draft.
Noor’s note is missing. Confirmation and date are unknown.
Restore a task. Can you restore a missing note?
FIELD NOTE / 01 · SOURCE: NIST AI RMF CORE 1.0 (2023)
ONE HUMAN CHECK
A confident answer needs a record and a reviewer.
NIST’s AI Risk Management Framework Core asks teams to define human oversight (Map 3.5) and measure performance in conditions like real use (Measure 2.3). Read the NIST AI RMF Core 1.0 (2023)
Buffer’s practical rule: Before a person releases an AI claim, name the reviewer and the record they will check. If either is missing, HOLD the claim.
In the made refund below: the Support lead checks the payment record. Even a logged request does not prove completion or an arrival date. A person checks each reply, even for €42.
Limit: This is Buffer’s example of applying the framework. NIST does not certify this rule. The made reply does not test an AI system at work or prove customer results.
01 / JUDGE A REAL WEBSITE CHECK
What did the check prove?
On 27 September 2026, Riin checked a Buffer Website change on desktop and phone. Pick the claim the public report can support.
27 September 2026 / live Website reportBOUNDED
Supported result: Four selected live Website requirements passed at 1440×1000 and 390×844. A visitor made a fictional feedback call and saw its consequence. A saved public task also led a return visit after reload.
Proof limit: The report covers those selected Website states. It does not test every AI claim or any customer work.
Public source: Read Riin’s 27 September report and its Office file link ↗
Now try Buffer’s rule on a made refund ↘A FIRST TRIAL / INTERNAL WEEKLY SUMMARYOne summary.
One summary.
Room for the missing note.
Buffer could prepare a draft from agreed notes. Your team lead checks it before anyone acts. First, see what that work includes.
The lamp workshop below is fictional. This sample runs no AI and starts no real trial.
01 / SAME INPUT FOR BOTH WAYSMADE NOTES
The week’s notes
Lamp workshop · week ending 2 October
Lea / 1 October
Packed 12 desk lamps. Order packing sleeves by 5 October.
Marten / 2 October
Tested the new dimmer. Update the fitting guide by 6 October.
Noor / note not supplied
The carrier update has an empty place. No collection confirmation or agreed date is known.
Both ways start here. Nobody gets an extra note.
02 / AI PREPARATION + HUMAN CHECKCHECK THE GAPS
A tidy draft. Two gaps.
Fictional example
12 desk lamps packed. New dimmer tested.
Lea orders packing sleeves by 5 October.
Fix from notes
Marten’s next task is missing from the draft.
Added from Marten’s note · not yet reviewed
Ada’s review is pending. After the correction, she can check Marten’s fitting-guide task against his note.
Source beside the line: Marten’s 2 October note says to update the fitting guide by 6 October.
Ask the note owner
OPENNoor: Did the carrier confirm Friday collection? What collection date is agreed?
No note supplied. The pencil leaves this place empty.
Collection stays held. Noor must supply confirmation and the agreed date; Ada then checks the updated paper.
One gap has a source. The other needs Noor’s answer.
Ada, the made team lead, checks every line before use. Collection stays on hold until Noor answers. This sample resets on reload.
Your draft has a bookmark. Back takes you to the words you were checking.
Earlier draft / unchanged
Fictional example
12 desk lamps packed. New dimmer tested.
Next task: Lea orders packing sleeves by 5 October.
Marten’s next task is missing from the draft.
Collection: Noor’s note is missing. Confirmation and date are unknown.
03 / COUNT THE WHOLE JOB
Two ways. The same notes.
The team lead checks facts, missed tasks, owners and dates in either version.
People prepare and check
Gather the notes. Use Lea’s and Marten’s notes. Mark Noor’s absent update.
AI drafts. People check and fix.
Prepare the same notes. Give AI Lea’s and Marten’s notes. Mark Noor’s absent update.
By hand · write the summary
Include work done, next tasks, owners and dates. Leave missing facts open.
With AI · prepare a first draft
AI arranges the work and next tasks. The made draft above misses Marten’s task.
By hand · Ada’s check
Compare facts, missed tasks, owners and dates with the notes. Correct errors. Ask Noor for confirmation and the collection date.
With AI · Ada’s check
Compare facts, missed tasks, owners and dates with the same notes. Restore Marten’s task. Ask Noor for confirmation and the collection date.
Count gathering, writing, checking, fixing and owner follow-up. Handwritten work can miss things too.
Count preparing the input, drafting, checking, fixing and owner follow-up. Draft speed alone cannot show usefulness.
No times were measured here. This example chooses no winner and proves no savings.
Know what would make it worth keeping.
Agree one weekly-summary task, its sources and reviewer, and what a useful result means. Compare all work against doing that task by hand.
Keep only if…
The checked summary helps the agreed task. Important facts and tasks are present, owners and dates are correct, and the total work is worthwhile.
Stop if…
Checking and fixing take longer than doing the task by hand, or important errors remain. A quick draft cannot cancel that rule.
Missing source? Ask its owner and pause the affected work. Fixing Marten’s line gives no answer about Noor’s collection.
05 / YOU MAKE THE CALL
Would this help your work?
Reload clears this reason.
Fictional example
YOUR PRACTICE DESK / FICTIONAL
THREE WORKTABLES / THREE SEPARATE MADE CASES
What would you
like to work out?
See a little of the work. Choose one question to open its exercise. Nothing is sent or run.
Where would AI help?
Weigh useful preparation against the source checks and human decisions still needed.
OFFER A / REFUND REPLYSettle the €42 refundRequest, receipt, photo. Ada decides.
OFFER B / MORNING ORDERSTidy the morning order pileReady orders and a missing supplier date.
OFFER C / SHELF NOTEAutomate an entire shelf note4 blue mugs. One very small empire.
Your call: keep a trial idea, wait for its reviewer, or keep the work with people.
Open the three task papers.Can this draft say that?
A fact and a promise share one reply. Give each sentence its own record check.
MADE REPLY / YOUR PENCIL NEXT
We issued your €42 refund.
It will reach your bank tomorrow.
Two lines. Two checks. ✎Your call: keep or remove each line. See what the customer could learn.
Open the reply and its records.Who does what next?
Give incoming work a team, or reopen a finished day when a late paper arrives.
THE DISPATCH DESKWork waiting for handsAI prepares. Ada checks one draft.
THE CUP SHOPA day that looks finishedTwo approvals. One closed envelope.
CHOOSE YOUR DESK
Your call: divide the work or repair changed work. Start either story on its own.
Two independent starts. Neither needs the other.RUN A TEAM / TWO FICTIONAL DESKS
One desk waits.
One says it’s done.
Choose where to begin. These are separate stories. You can open either without finishing the other.
Divide the work
Three incoming papers. One AI turn. One human review turn. Decide who takes what.
D-103 / COLLECTION NOTICEA draft is already hereAda can check it against its paper.
D-104 / DELIVERY REPLYWaiting for an AI draftThe delivery record is ready.
D-105 / SUPPLIER QUESTIONA date is still missingA person must ask the supplier.
ONE HUMAN STAMP
AI helpers prepare. Ada, the dispatch lead, decides what can be used.
Open the incoming papers.Repair changed work
A finished day is supplied. Read its approvals, make a guess, then open the late envelope.
P-12 / COLLECTION NOTICECollect Friday from 14:00Approved / Ada Ruut
C-12 / SHELF NOTE4 blue mugs. No reorder.Approved / Jo Kask
Late envelopeStill closed. Guess before opening.
TWO GOOD STAMPS. FOR NOW.
Ada approved collection. Jo approved the shelf note. Which decisions still hold?
Open the supplied finished day.SAVED MARKS / ON THIS DEVICE
Right where your pencil stopped.
Open a marked paper. A finished practice can still hold a reply.
No practice marks yet. Try a case to leave your first one.
UNFOLDED / YOUR MARKS KEPT
NEXT / OPTIONAL · FICTIONAL REFUND A
Give confidence a red pencil.
A matching payment has arrived. Check two sentences. Choose what the customer could learn.
ONE REPLY / TWO SEPARATE CHECKS
Keep the fact.
Question the promise.
The customer asks: “Was my €42 refund issued? When will it reach my bank?” These records are made up. Nothing is sent.
01 / THE PAYMENT SENTENCE
We issued your €42 refund.
Select a sentence, then open its record.
YOUR RED PENCIL
Open the selected sentence’s record before keeping or removing it.
Two lines need a check. A green stamp cannot add a bank arrival date.
THE CUSTOMER’S VIEW / UNSENT PREVIEW
WAITINGWhat would the customer learn?
NO CHOSEN WORDING YET
No reply readyNo reply ready
Check both lines before saving the preview.
What is still unanswered?
Was the refund issued?
The record says the refund was issued. Its sentence still needs your choice.
When will it reach my bank?
The arrival question needs a bank record. P-781 gives no arrival date.
Next: the Support lead checks P-781, the wording and the missing bank record.
Edit the sentences ↑Keep the wording, questions and earlier calls together on this device.
Choices stay on this device. Nothing is sent.
THE CASE TRAY / MADE PRACTICE
Pick what could go wrong.
Fictional cases. Open two tests, compare replies, or run a support shift.YOUR PRIVATE COMPARISON / ON THIS DEVICE
One rule. Two papers.
ORIGINAL RULE / KEPT AS WRITTEN
RULE AFTER THE SECOND PAPER
Your words stay here. Made cases cannot prove a rule works in real company work.
Earlier comparisons on this device
THE LITTLE CUP SHOP / A FICTIONAL FINISHED DESK
Done.
Until this.
Two checked papers. Two good stamps. Then one late envelope asks your desk to think again.
Made people and papers. Nothing is sent, no AI runs, and no real shop work changes.
ORIGINAL ENDING / BEFORE THE ENVELOPE
Both decisions matched their papers.
Work was approved for these versions. No notice was sent in this exercise.
01 / GUESS BEFORE OPENING
One paper has changed.
Which decisions need another look?
Your guess cannot change the papers. The late envelope is still closed.
Read the finished work above, then keep your first guess.
LATE PAPER / P-12 · VERSION 2
Friday collection is cancelled.
Order B-108 is still packed. The shop is closed Friday afternoon. Collection moves to Monday from 10:00.
P-12 version 2 replaces version 1. Count slip C-12 has not changed. No collection end time is given.
Collection noticeNeeds a new decisionAda Ruut approved Friday collection. The time changed, so she must decide again.
Shelf noteEarlier decision still validJo Kask approved the shelf count. It excludes packed orders, so the collection change does not affect it.
Ada Ruut approves collection notices. She is away today. Jo can prepare one, but cannot approve it.
A fresh draft
is not a fresh stamp.
AFTER THE ENVELOPE / NOTHING SENT
One stamp stays. One decision waits.
Original ending kept: Friday notice and shelf note approved. Now: collection notice needs review, shelf approval stays.
Collection notice / B-108
Needs a new decisionChanged fact: Friday is cancelled. P-12 version 2 says Monday from 10:00.
OLD NOTICE / WITHHELD
Your blue mug, order B-108, is packed. Collect it on Friday from 14:00.
The Friday wording no longer fits the current paper.
Ada Ruut approved the Friday notice using P-12 version 1. That decision stays in the original ending. It cannot approve Monday’s notice.
Ada Ruut / required reviewer
Away from the shop today. Jo Kask is available to prepare work, but cannot approve collection notices.
Shelf note / shelf B
Earlier decision still validCount slip C-12 version 1 is unchanged: 4 blue mugs, shelf B, reorder at 2. Packed orders are counted separately.
Shelf B: 4 blue mugs. No reorder needed.
Jo Kask approved this same note and count slip. The new collection time does not change the shelf count. Keep Jo’s approval.
THE LITTLE CUP SHOP / A MADE FRIDAY MORNING
Big offers.
Small print inside.
Which job would you give AI first? Pick an offer. Then open all three papers before deciding.
Fictional practice. The shop, people and papers are made up. No AI runs, no money moves, and nothing is sent.
Pick before seeing the work inside. You can change your mind later.
Three offers. The checking work is still folded inside.
02 / THE PAPERS ARE OPEN · YOUR FIRST PICK STAYS
What is worth a small try?
Keep your pick, choose a different task, or keep all three with people. None of these choices starts a trial.
Final choice still open. Reading papers did not make the choice.
Keep a choice or a stopping point on this device. Nothing is sent.
No choices saved yet.
YOUR CHOICE / MADE PRACTICE
Choice ready · not savedOne morning. Your call.
Trial not runKept with people
Next check
This keeps an idea and its limits. It gives no approval and proves no real work result. Nothing sent.
FIRST PAPER KEPT / NOW TRY ANOTHER
The next paper did not read your rule.
Choose a different mistake below. Your first call and original rule stay in the comparison.
SECOND PAPER / YOUR RULE MEETS THE FACT
Did your rule catch this?
ORIGINAL / STILL HERE
Changing a rule cannot approve a held draft. The case still needs its human check.
Write a rule before saving.
Choose a case above. Your other marks stay on their own papers.
Device storage is unavailable. Your made choices work here but may not return after reload.
THE LITTLE DISPATCH DESK / FICTIONAL
More hands.
Same papers.
Direct the AI helpers. Give Ada one review turn. Keep the ending, then try a different team.
Made people, parcels and papers. Nothing is sent. Each arrangement has one AI preparation turn and one human review turn.
01 / Arrange the desk
1 AI helper
Pip prepares one waiting paper.
One helper takes the first waiting paper. Two helpers take both. Neither supplies the missing date.
Ada gets one turn. A task without a draft or fact uses that turn waiting. Preparation gives no approval.
Optional. Use the same AI turn now, or let the helpers work while Ada reviews.
AI preparation: 1 turn available
Ada’s review: 1 turn available
1Papers to prepare
1Drafts for Ada
0Fictional releases
1Held tasks
02 / Incoming papers. Drafts and sources stay together.
Review the collection notice. Pip prepares the supplier question in the same turn.
Ada’s call belongs only to the exact draft and source shown above.
Three papers arrived. Only the collection notice already has a draft.
Open earlier endings Kept while you arrange this team
Keep an ending to save it on this device. Nothing is sent.
03 / ENDING SHEET · NOTHING SENT
Every paper has a next step.
Ready for review, fictional release and missing facts are different endings. The task order stays fixed.
01 / Delivery reply · D-104
02 / Supplier date · D-105
03 / Collection notice · D-103
Your turn, away from this desk. Name a different work task AI could prepare. Which fact would a person need to check?
TWO MADE TESTS / THE BIKE WORKSHOP
Very reassuring.
Open the covers.
AI writes collection replies. Ada checks the job number and date against the booking before a reply leaves.
Fictional practice. These events are already written. Your calls change the explanation, never the reports. Nothing is sent.
The feared mistake
A draft gives the wrong collection date. The rule: Ada checks both the job number and date, then holds any mismatch.
TEST A / COVER SUMMARY
A tidy finish.
No unsupported reply releasedTwo drafts. One reviewer. What does this ending show?
TEST B / COVER SUMMARY
Another tidy finish.
No unsupported reply releasedTwo drafts. One reviewer. What does this ending show?
Make both first calls. Opening the covers keeps those calls for comparison.
Both covers look clean. The papers are still closed.
COVERS OPEN / FACTS STAY FIXED
Same ending. Different test.
The cover told you what left. It did not tell you what tried to leave.
TEST A / FIXED REPORT 1
The mistake never arrived.
0 wrong dates · 2 matching drafts
What we expected
Ada would hold a wrong date. She checked two made drafts against the bookings below.
What happened
Job J-041
- Source
- Booking J-041: 14 October.
- AI draft
- “Job J-041: collect on 14 October.”
- Reviewer action
- Ada matched the job and date. Released unchanged.
Job J-042
- Source
- Booking J-042: 15 October.
- AI draft
- “Job J-042: collect on 15 October.”
- Reviewer action
- Ada matched the job and date. Released unchanged.
No unsupported reply released. There was no wrong date to catch.
What this leaves unanswered
Would Ada catch a wrong date? This test did not try one. It also tried no wrong job number.
Your call / Test A
- First call / kept
- No first call
- After opening / current call
- No fresh call yet
First call kept. A fresh call is waiting.
The two drafts matched. This test did not face the feared mistake.
TEST B / FIXED REPORT 1
The mistake met Ada.
1 wrong date · caught before release
What we expected
Ada would hold a wrong date. She checked two made drafts against the bookings below.
What happened
Job J-051
- Source
- Booking J-051: 14 October.
- AI draft
- “Job J-051: collect on 14 October.”
- Reviewer action
- Ada matched the job and date. Released unchanged.
Job J-052 / deliberate wrong date
- Source
- Booking J-052: 15 October.
- AI draft
- “Job J-052: collect on 14 October.”
- Reviewer action
- Held before release. Ada changed 14 to 15 October, then checked the revision.
- Released reply
- “Job J-052: collect on 15 October.”
No unsupported reply released. One wrong date was caught and corrected.
What this leaves unanswered
Ada caught one wrong date with these bookings. Other wrong dates and wrong job numbers remain untried.
Your call / Test B
- First call / kept
- No first call
- After opening / current call
- No fresh call yet
First call kept. A fresh call is waiting.
Ada caught one wrong date under these conditions. That is one example.
These fictional tests do not show that the rule always works.
ANOTHER SAMPLE / YOUR PROPOSAL
A question for the next envelope.
Not run · No sample chosen
Choose a made sample, or keep the question open. Nothing runs and no private plan starts.
This is a proposal, not another result.
Save when you want to keep these calls on this device.
No calls yet.
FICTIONAL PRACTICE / PAPER MOON NOTEBOOK SHOP
Two right numbers.
One small question.
Unfold two sales slips. Choose the paper that answers the question. Then make your call.
01 / ASK BEFORE YOU ADD
How much did the shop sell?
Requested dates 21–27 September 2026
Count paid notebook sales, including VAT. Leave shipping out.
The question can move. The ink on the slips cannot.
02 / Unfold the evidence
Read either paper first. Opening is not choosing.
Source A
A sales slip from the shop’s till.
FIXED MADE RECORD / ORIGINAL INK
- Dates
- Counts
- Leaves out
Not chosen
Source B
A sales slip from the shop’s till.
FIXED MADE RECORD / ORIGINAL INK
- Dates
- Counts
- Leaves out
Not chosen
03 / YOUR CURRENT QUESTION
Check this summary
Needs a new reviewRequested 21–27 September 2026
Choose an open source to prepare a summary.
No paper chosen.
Read a paper, choose evidence, then review the words.
Earlier question
A right total still needs the right question.
SUPPORTED MADE ANSWER / YOUR REVIEW
This paper answers this question.
Requested
Nothing sentMade practice only. A person checks real records before sharing a real summary.
Progress stays on this device.
FICTIONAL PRACTICE / YOU RUN EVERY TURN
Two requests.
One free chair.
Choose who goes first. Check both records. Then take the customer’s seat.
1 REVIEWERUrgent stamps may move the chair.
They cannot move money.
They cannot move money.
These people, requests and records are made up. Nothing is sent. Cleared wording still needs a person’s review.
01 / ATTENTION IS YOUR CALL
Who gets the chair?
- CHAIR UNASSIGNED
Lea / €42 refund
- CHAIR UNASSIGNED
Jo / workspace access URGENT
Nobody has the chair yet. Choose who gets attention first.
02 / THE CHAIR MOVES. THE RECORDS STAY.
Check what each reply can say.
Choose an order first. Then open each record before choosing its reply.
Lea / refund
NO ORDER YETFIXED MADE PACKET / ORIGINAL RECORD 1
Payment P-781
“Issued” has support. A bank arrival date does not. Holding a reply cannot undo this payment.
NO WORDS CHOSEN
The payment and the reply are separate.
Jo / access
NO ORDER YETFIXED MADE PACKET / ORIGINAL RECORD 1
Access log L-208
A queued email is a next step. It is not a successful sign-in, however large the urgent stamp.
NO WORDS CHOSEN
Urgency is not a successful sign-in.
Choose a reply or silence for both requests to open the customer view.
03 / THE OTHER SIDE OF YOUR WORDS
Now you are the customer.
NOT SENT
Your saved shift is here. The same records, wording and customer turns are kept.
LEA’S INBOX / ONLY CLEARED WORDING
Was my refund issued?
No reply ready
An arrival answer needs a bank record. P-781 gives no bank arrival date.
Choose Lea’s next move. No action adds a bank date.
JO’S INBOX / ONLY CLEARED WORDING
Can I get into the workspace?
No reply ready
Choose Jo’s next move. Checking email does not unlock the workspace.
Choose each customer’s next move to finish this shift’s turns.
Earlier choices in this shift
The old three-request queue and refund exercises keep their own saves.
Open the earlier pending-refund queue ↗Choices stay on this device. Nothing is sent.
FICTIONAL PRACTICE / TWO VERSIONS OF ONE REFUND
The stamp has a twin.
The payment does not.
Same request. Same amount. Same AI reply. Open each set of papers, then decide whether its words can go to a person for review.
Nothing is sent. These requests and payments are fictional. A person still checks both replies. Neither set of records proves a bank arrival date.
Reply A
REFUND R-042 / ORDER B-104MADE AI DRAFT
“We issued your €42 refund.”
READYAI SAYS SO
Fictional request
- Refund
- R-042
- Order
- B-104
- Amount
- €42
- Request status
- Logged
Fictional payment P-781
- Refund
- R-042
- Order
- B-104
- Amount
- €42
- Payment status
- Completed
Same refund. Same order. The completed payment supports “issued.” Bank arrival is not recorded.
Reply B
REFUND R-042 / ORDER B-104MADE AI DRAFT
“We issued your €42 refund.”
READYAI SAYS SO
Fictional request
- Refund
- R-042
- Order
- B-104
- Amount
- €42
- Request status
- Logged
Fictional payment P-779
- Refund
- R-019
- Order
- B-099
- Amount
- €42
- Payment status
- Completed
Right amount. Wrong refund. This payment belongs to a different order. R-042 is only a logged request.
Make both calls, then compare them with the records.
THE RECORD GETS THE LAST WORD
Two replies. One matching payment.
Your current calls are below. A completed refund and a waiting reply can both be true.
Record supports “issued.”
- Payment for R-042 / B-104
- Refund issued
- Customer reply
- Reply waiting
Next human check: The Support lead matches R-042 and order B-104 to completed payment P-781, then checks the exact reply. No bank arrival date is supported.
Change A’s call ↑Wording stays held.
- Payment for R-042 / B-104
- Request logged
- Customer reply
- Reply waiting
Next human check: The Support lead finds the payment status for R-042 and order B-104. P-779 belongs to a different order. Until then, the reply cannot say “issued.”
Change B’s call ↑CURRENT CALLS / HISTORY KEPT
Your latest call leads.
Your calls and chosen check stay on this device.
FIELD NOTE / 02 · DATA MINIMISATION
The name has left the field. Has it left the draft?
Follow one stray name through the paper trail. Decide what a draft may show to AI before a person reviews it.
FICTIONAL PRACTICE / NOTHING IS SENT
The case of the second Anna
This fixed ticket is made up. The page makes no AI call and asks for no real customer details. The AI input below changes only when you edit this example.
01 / THE MADE TICKET
- Name field
- Anna
A name can travel in more than one place. Find both copies.
02 / EXACT INPUT THE AI WOULD SEE
Name: Anna My parcel is late. Call Anna about the delivery.
MADE DRAFT PREVIEW
“Hi Anna, I’ll look into your late parcel.”
03 / TRACE THE NAME
First hide the separate name field. Then look for the name still written in the message.
04 / BUFFER’S REVIEW DESKHOLD
HOLD · The made ticket still contains Anna. Check the exact AI input before a person reviews any draft.
The greeting uses a name from the input. Hiding one field does not erase the same name in the message.
Source: GDPR Article 5(1)(c) (2016) says personal data should be adequate, relevant, and limited to what is necessary. The ticket and its outcome here are fictional.
Buffer’s review rule: Inspect the whole input, including free text. Hold this made draft while the name remains. A cleaned input can go to a person for review; it is never approved or sent automatically. This is Buffer’s example rule, not a rule stated by the GDPR.
FIELD NOTE / 03 · FICTIONAL PRACTICE
Choose what the AI needs.
A made parcel ticket has three facts. Put only useful facts into the draft, then check what a person still needs to verify.
THE MADE TICKETNO REAL CUSTOMER DETAILS
Question“Where is my parcel?”
The courier record says a label exists, but there is no delivery scan. A label does not prove arrival.
01 / ORDEROrder D-204
Connects the question to the right order.
02 / COURIER STATUSLabel exists. No delivery scan.
Shows what the record can and cannot confirm.
03 / NAMEMarta
This question can be drafted without the name.
Marta stays out of the AI input.
EXACT MADE AI INPUT0 of 3 facts
Task: Prepare a draft answer. A person checks it before any reply. Customer question: Where is my parcel? Chosen facts: none yet.
Stays out: order D-204, courier status, Marta.
MADE DRAFT PREVIEWUNSENT
“I need the order and courier record before I can draft a useful answer.”
This page makes no AI call, upload, or send action. The draft changes only from this made example.
THE REVIEW DESKAWAITING FACTS
Choose facts, then place the draft here. You can change any choice and compare again.
Which paper must a person check?
A person must check the courier record for a delivery scan before any reply. The label alone cannot confirm delivery.
Select the paper you would hand to the reviewer.
FIELD NOTE / 04 · FICTIONAL PRACTICE
The AI said cancel. Then page four arrived.
A made tool has few users and a higher price. Make the call, open the late contract page, and see what your choice costs.
THE MADE RENEWAL / TOOL R-4NOT A REAL COMPANY DECISION
01 / RECORDS ON THE DESK
Usage log · last 30 days4 of 20 seats active
Low use in this made month. It does not say why people stopped using the tool.
Renewal invoice€1,800 / year
Last year: €1,200. The new price is €600 higher.
02 / EXACT FACTS THIS MADE AI SAW
Task: Suggest a renewal call. A person decides. Usage log: 4 of 20 seats active in the last 30 days. Renewal invoice: €1,800 for the next year; €1,200 last year. Contract: not supplied.
AI'S MADE NOTE / NO AI REQUEST RUNS
CANCEL · PROVISIONAL
Low use and a €600 price rise point toward cancellation, but the contract term is missing. Do not act yet.
03 / YOUR HUMAN CALL
Which stamp goes on this desk?
Choose once, then change your mind when the late paper arrives. A choice here runs no action.
MISSING PAPER: The contract term is not on the desk. The AI cannot price the access risk.
ARRIVED LATE / PAGE 4
Made contract term
For this made 1 October renewal, written notice was due by 1 September. The file has no notice by that deadline. Late cancellation may still leave €1,800 due. Workspace and export access end when cancellation takes effect. This page does not confirm that effective date or a complete export.
04 / CONSEQUENCE ON THE DESKAWAITING YOUR CALL
AI'S MADE ADVICECANCEL · PROVISIONAL
YOUR REVERSIBLE CALLNo call chosen yet
THE VISIBLE COST
Choose renew, cancel, or hold to see its money and access consequence.
PERSON AND RECORDS
The Operations lead must check the signed contract, effective date, €1,800 invoice and payment status, and export status before a real decision.
This is a made exercise. No purchase, cancellation, email, or AI request happens here.
FIELD NOTE / 05 · FICTIONAL PRACTICE
One free person.
Three requests.
Choose who gets help first. The queue moves, and the other two tickets tell you what waiting costs.
THE COST OF WAITING
All three have a reason to ask.
Read each delay before you choose. “Help first” reserves the only free chair in this made queue. You can change your call.
-
ACCESS / BLOCKEDCannot sign in for the next shift
A team member sees a sign-in error. No one has checked the account or the error log.
If it waits: the next shift may start without its work queue.
NOT CHOSEN -
REFUND / STATUS“Where is my €42?”
The made record logs a refund request. It does not show a completed payment or arrival date.
If it waits: the customer still has no checked answer about their money.
NOT CHOSEN -
EXPORT / TIME-SENSITIVEFigures needed before a meeting
A finance export is queued. The file is not confirmed ready for the made 16:00 meeting.
If it waits: the meeting may start without checked figures.
NOT CHOSEN
YOUR MADE PRIORITYONE CHAIR FREE
HELP NOW
No request chosen yet.
STILL WAITING
Choose a ticket to see why the other two wait.
PERSON'S CHECK
These people and requests are made up. This page makes no support request, AI call, email, refund, or service change. A person would confirm the records and decide what to do.
FIELD NOTE / 06 · FICTIONAL PRACTICE
A neat handoff.
One loose end.
The AI has marked a made launch READY for Jo. Open the original note before deciding what Jo should receive.
AI'S SHORT HANDOFFREADY
“The launch email and help page are prepared. Jo can announce at 09:00.”
This stamp is only the AI's draft. Check the original work note before Jo acts.
THE ORIGINAL WORK NOTE
The short handoff is missing a line from the source. Find it before choosing.
SOURCE OPEN / STAMP RECHECKEDHOLD
- Draft launch email prepared
- Help page updated
- Confirm who covers Friday support before Jo announces
The cover line is unchecked. If Jo announces now, people may ask for help when no one is assigned to answer.
YOUR NEXT CALL
Who gets the loose end?
These choices change only this made case on your device.
CONSEQUENCE FOR THE NEXT PERSONAWAITING SOURCE
Open the original note. The AI's READY stamp is not a finished decision.
Jo has no checked support cover in this draft.
HUMAN CHECK BEFORE ANY ANNOUNCEMENTThe Support lead checks the Friday rota and confirms who accepted cover. Jo checks that record and the updated handoff before any announcement.
Jo, this launch, and every note here are made up. No choice sends a message, invites anyone, or runs work. Real Office records sit outside this case.
MADE CASE 06 / A TRIAL WEEK
Friday says done.
Monday asks who checked?
A fictional assistant drafted three routine replies. Read each paper below: the correction and review time are part of the work.
FRIDAY PILE03drafts made
THE PAPERS / FICTIONAL
What the tidy pile hides
- 01 / REFUND REPLYReview time · 11 min
AI draft: “Your refund is complete and will arrive tomorrow.”
Person corrected: The record shows only a request. Remove the completion and arrival claim. Check the payment status before replying.
- 02 / ACCESS NOTEReview time · 8 min
AI draft: “Your account is active. Please reset your password.”
Person corrected: Account status was not checked. Ask the Support lead to inspect the sign-in error and identity before suggesting a reset.
- 03 / EXPORT UPDATEReview time · 13 min
AI draft: “The file is ready to share before the meeting.”
Person corrected: The export is queued. Check permission, job status, and contents before promising a file.
32 minutes of made review for three made drafts. These are story times, not measured customer results or proven savings.
YOUR MONDAY CALL
Give the assistant its next job.
Choose what happens to this one made task. You can change your mind; each call moves work on Monday's desk.
MONDAY / THE HUMAN DESK
A PERSON STILL HAS TO
BEFORE TUESDAY / YOUR LIMIT
What would make you change this job?
Pick one stop rule before opening the next made record. The rule will be tested against what Tuesday shows.
MADE TUESDAY / JOB LOG OPENEDQUEUED
THE ASSISTANT'S NEW TUESDAY NOTE
“Export finished. File ready for the meeting.” DONE
THE NEW JOB LOG
Tuesday's export job 418 is still QUEUED. Mira owns the check. The made review took 6 minutes. There is no finished file to approve.
This device could not save this case. Your choices work now but may not return after reload.
This whole week and Tuesday's log are fictional practice. The review times are made up, not measured savings or customer results. Nothing runs, uploads, or sends. Your job choice, stop rule, and opened log stay on this device.
Earlier work and the Playground
MORE TO EXPLORE
Earlier work lives here.
Old made demos and saved Website choices remain on this device in the earlier experience.
About older proof
Older proof can expire. In the archive, that state is marked This proof is no longer current.
Check its own date and source.
WATCH REAL WORK / PUBLIC RECORDS
READING DESK / 1 OCTOBER 2026 RECORDS
One shift. Three stamps.
The Office chose to join queue priority, reply checks and the customer’s next move in one fictional support shift.
RECORDED REASON / PRESERVED HISTORY
Keep the same facts through the shift.
The preserved plan passage joins the work so visitors can see what a supported reply tells the customer, and what the customer still needs to ask.
Recorded alternative: build the customer turn separately. No tested result for that alternative is recorded.
Attributed to the former Website plan section “Active product outcome: Run one made support shift”. That section is absent from today’s plan. The original decision time is not recorded here.
Read the published outcome · original English ↗The outcome’s exact Result passage · original English
Visitors run a made support shift, change who goes first, check each reply and choose the customer's next move. Reload brings back the same records, replies and unanswered question.This records the intended return on release 161941e4. Publication and browser findings are separate. This is the quote kept with a private view.
Original decision identities
- Task
- milestone-20261001085443681-6bfa6243
- Outcome
- outcome-job-20261001085446228-2f773549
- Original release
- release-20261001091505028-161941e4
Three open papers. Fold any you’ve read.
Each stamp belongs to its own release. The two passes tested different returns.
01 / CUSTOMER RETURNrelease 161941e4
No saved customer choice.
FAILEDDid the customer’s bank-date question return after reload?
Expected ask-arrival; found none. The full phone journey was not established. The cause is unknown; the report does not separate a missed choice from a lost save.
Preserved dated excerpt. Its former report path now contains finding 03, with another report ID and release. That newer record cannot replace this failure.
visual-qa-20261001091612125-3c4ff53902 / OFFICE FILE RETURNrelease 51693617
The Office file returned.
PASSEDDid the saved Office file return with its failed-check limit?
The saved reason and failed customer finding returned at 1440×1000 and 390×844. This checks the Office file, not the customer’s move in the fictional shift.
Read report 02 · original English ↗visual-qa-20261001124623608-7d2c2ad803 / LATER CUSTOMER RETURNrelease 8a138cfb
The customer’s question returned.
PASSEDDid the same customer question survive before and after reload?
The supported reply, records, changed priority and ask-arrival choice were present before and after reload at 1440×1000 and 390×844. This is a matching customer check for release 8a138cfb only. It does not rewrite finding 01.
Read report 03 · original English ↗visual-qa-20261001231346042-d53cb843YOUR PENCIL / OPTIONAL / PRIVATE
What would you keep in mind?
A personal view, saved only on this device. It approves no Office work and sends nothing.
The exact Result passage you kept · English
The outcome’s return promise, not a new check. Original publication: release 161941e4.
Earlier private views · unchanged
Read the same public sources again.
A fresh read changes neither your note nor the dated excerpts. An unavailable source does not prove it was deleted.
Earlier Office papers and saved calls · English
WATCH REAL WORK / THE DESK
Open the choice. Check what followed.
Public work has its own dates and records. Your private review rule stays on this device.
FILE 01 / REAL WEBSITE WORK1 October 2026 records
THE CHOSEN APPROACH
Run one made
support shift.
Join queue priority, reply checks and the customer’s next move. Follow this real choice about a fictional exercise.
Preserved history from the former Website plan section, “Active product outcome: Run one made support shift”. That section is absent from the current plan. The original decision time is not recorded here.
Published
1 October 2026 · release 161941e4
The outcome records publication, verified at 12:16 Tallinn time. Read the published outcome ↗
01 / THE PLAN’S REASON · 1 OCTOBER RECORD
One shift, the same facts.
The plan brings queue priority, record checks and the customer turn into one shift. Visitors can see what a true reply tells the customer and what they still need to ask.
THE RECORDED ALTERNATIVE
Build the customer turn separately. The plan records replacing that approach with the joined shift. It gives no tested result for the alternative.
Preserved, attributed history from the former support-shift plan section. That section is absent from the current Website plan ↗. Its current file date is not the decision time.
02 / THE LIVE FINDING · 1 OCTOBER 2026
A return question stays open.
FAILEDThe report shows a held-answer view and several passed desktop checks. It then stops at the customer choice after reload.
- Expected
- Ask the bank-date question
- Found after reload
- No saved customer choice
The customer choice after reload was not proved. The complete phone journey was not established. The cause is unknown.
Preserved report visual-qa-20261001091612125-3c4ff539 · captured at 12:16 Tallinn time on release 161941e4. Its former path now opens the later customer report ↗, not this original failure.
The plan’s benefit remains unanswered.
Intended return: Keep priority, checked replies, original records and the customer’s unanswered question together. The outcome promises this return. The failed check does not prove it.
The report does not show whether the question appeared before reload. It cannot separate a missed choice from a lost save.
TWO DATED LIVE CHECKS / ONE CALL FROM YOU
A new stamp.
An old question.
The customer’s bank-date question did not return in the earlier check. Does the newer passed check answer that question?
Make a call, then open both papers. You can change your call.
01 / Earlier check
FAILED
Live Website · release 161941e4
Tested question inside ↓
WHAT THIS CHECK TESTED
Does the customer’s bank-date question return with the made support shift?
No saved customer choiceExpected “Ask the bank-date question” after reload. Found no choice.
The full phone journey was not established. The report cannot tell whether the choice was missed before reload or lost on return. The cause is unknown.
Preserved dated excerpt from report 3c4ff539. Its former path now holds the separate later customer report ↗.
Original record identities
- Task
- milestone-20261001085443681-6bfa6243
- Outcome
- outcome-job-20261001085446228-2f773549
- Release
- release-20261001091505028-161941e4
- Report
- visual-qa-20261001091612125-3c4ff539
02 / Later check
PASSED
Live Website · release 51693617
Tested question inside ↓
WHAT THIS CHECK TESTED
Does the saved Office choice file return with its failed-check limit?
The Office file returned.Its reason and failed customer finding stayed together after reload, on desktop and phone.
Tested at 1440×1000 and 390×844 on this release. This supports the Office file’s return. It does not prove the customer’s return on either release.
Dated excerpt from the later live report ↗.
Later record identities
- Task
- milestone-20261001121827700-00ed70e6
- Outcome
- outcome-job-20261001121830556-b5e521ae
- Release
- release-20261001124512990-51693617
- Report
- visual-qa-20261001124623608-7d2c2ad8
BOTH PAPERS OPEN
Two different questions. The passed stamp stays on the Office file.
Within these two reports, the customer’s return was not proved.
The second check returned the paper about the failed check. It did not test whether the customer’s choice returned. A third, matching customer check belongs to another release; read it at the new reading desk.
NEXT / FICTIONAL WORKSHOP
The right words can reach the wrong desk. Try to fool a check, then teach it to notice.
FICTIONAL WORKSHOP / PROPOSED CHECKS
Right words.
Wrong desk?
Pick a check. Guess what it will notice, open a made return, then improve the check on the same facts.
These are made returns. Nothing is sent. This runs no real customer test.
Want to choose each detail yourself? Test all three papers together.
The whole request
- Fictional ticket / original desk
- T-042
- Customer’s question
- Customer’s bank-date question: “When will my €42 refund reach my bank?”
- Checked reply
- “We issued your €42 refund.”
- Queue priority
- Refund first. Locked access next.
- Original record
- P-781 · €42 issued. No bank-arrival date.
- Office file
- Reason and failed customer finding.
What will your check see?
Choose a proposed check above. Only the details it covers must match.
- ○ Ticket number
- ○ Customer’s bank-date question
- ○ Checked reply and payment record
- ○ Queue priority
- ○ Office reason and failed finding
The question check reads the words. Only the ticket check asks whose words they are.
Still in its envelope
Make your guess before opening a return. The ticket stays sealed too.
- Fictional ticket / returned desk
- Customer’s question
- Checked reply
- “We issued your €42 refund.”
- Queue priority
- Refund first. Locked access next.
- Original record
- P-781 · €42 issued. No bank-arrival date.
- Office file
- Reason and failed customer finding.
YOUR GUESS / BEFORE REVEAL
RESULT / FICTIONAL RETURN
EARLIER TRIAL / KEPT AS IT WAS
CURRENT TRIAL
This made workshop proves no real customer return. The reading desk keeps the separate later customer report and its release.
Holding every reply is no shortcut. A return that keeps everything must pass. Accepting the return still promises no bank date.
All trials in this comparison
Choose a proposed check to start.
Your choices stay on this device, separate from the real Office file.
Revise your call or keep it to return after reload. Both papers keep their ink.
Your call stays separate from what these checks prove.
DATED EXCERPTS / FRESH SOURCE ACCESS
Fresh public records are unavailable here. The attributed 1 October excerpts remain readable as historical records.
- Fresh plan: not checked.
- Original outcome: not checked.
- Original report path: not checked.
- Office-file outcome: not checked.
- Office-file report: not checked.
This earlier comparison keeps its original two reports. The new reading desk separately presents the 23:14 UTC customer pass on release 8a138cfb. Read current Office work for updates ↗
Keep the question with the file.
This saves your selected choice on this device. It does not direct Office work or save a fictional customer turn.
Earlier task slips and private notes stay separate.
CHECKING THE OFFICE
Finding a public task…
The public record will name it here.
LAST DATED DECISION
Checking the public record…
Checking for a newer decision…
One task stays here, even when newer work starts beside it.
See other saved tasks in the archive ↗
OPEN / AN INVITATION
Read the current public record ↗
CHECKING FINISHED WORK
Checking the latest release…
Reading its matching public result.
Read the matching live check ↗ Read the public outcome ↗Other tasks saved on this device
Open an earlier slip without losing the task on your desk.
Inspect another real Website claim
REAL WEBSITE WORK / THE RED PENCIL
Follow a claim to its check.
Two real Website checks are on the desk. Mark the words that reach further than their records.
QUESTION / PHONE CLAIM
The claim on the draft:
“Our Website always catches every confident claim before it goes out.”
What needs the red pencil?
The big words are crossed out. See what the checks actually covered.
THE SMALLER STATEMENT
These checks show selected Website states on a desktop and phone.
On 27 September, a made feedback choice showed its consequence. On 29 September, one made refund reply was held when its record supported only a logged request. Neither check proves that every claim is caught or that any customer received a result.
CHECK 01 / CHOICE VISIBLE
Release c9bbc8e5
Riin’s live report passed four selected Website checks at 1440×1000 and 390×844. A visitor made a fictional feedback call and saw its consequence.
Limit: This shows the selected made call and its displayed consequence at those sizes. It does not test every claim or customer work.
Report visual-qa-20260927175016744-d86930da
Read the 27 September report ↗CHECK 02 / CLAIM HELD
Release 7c285e8a
Riin’s live report passed three selected Website checks at 1440×1000 and 390×844. In the made refund exercise, approval of a confident draft ended at “HELD / NOT SENT” because the record showed only a request.
Limit: This checks one fictional refund path on the Website. It does not prove a general detection rate, a real refund, or an email being sent.
Report visual-qa-20260929175213299-45f740a5
Read the 29 September report ↗Source: Riin’s 27 September live Website report and 29 September live Website report. Each report has a stable ID and a link to its Office file.
Proof stops here: Selected page states and interactions passed at the tested sizes. These reports do not establish how an AI team performs for customers.
THE CLAIM HEARING / MADE REFUND
Call the records as witnesses.
The Website was published. The made reply was held. Which record can speak to each claim?
Choose both public records. Each witness gets only the words its source can support.
Release record waiting to be heard.
It records publication of this Website lesson. It has no payment transaction.
Read the 29 September release outcome ↗Live page check waiting to be heard.
It covers selected Website states at desktop and phone sizes, including one made refund path.
Read the 29 September live report ↗ONE LAST CLAIM
Did the made refund actually happen?
VERDICT / PAYMENT STILL UNPROVED
Neither record proves a refund happened.
The made payment record logs a request, not completion. A real customer refund would need its own payment proof and human review.
KEEP THIS QUESTION
Return to this same public question. A new answer needs a matching published result and a passed live check.
The 29 September check is the last dated answer here.
ASK SIMO / UNSENT
Put your question on paper.
Review the exact words. This page sends nothing.
Reload clears this draft and review. Copy the reviewed words before leaving if you want to keep them.
Add a subject and message before reviewing.
UNSENT / REVIEWED HERE
simo@buffer.ee
Open this draft in your email app ↗
Reload clears this draft and review. Keep a copy before leaving.
You decide whether to send it.
Leave a bookmark here while you check the fictional paper.
Private notebook · in English
YOUR NEXT STEP · PRIVATE
A page for your next test.
Keep a plan and a guess. Come back with what surprised you.
PRIVATE NOTEBOOK / ONE SMALL TEST
Guesses meet surprises.
This page runs no task and sends nothing. Your words stay in this browser on this device.
YOUR EARLIER RULE · KEPT AS WRITTENHOLD
AI PREPARES
RECORD TO CHECK
HUMAN REVIEWER
SMALL SAMPLE MUST SHOW
Original words and record state kept. No prediction or test report was saved with this rule.
Why I stopped
Stopping is a decision. It does not mean a test ran or passed.
KEPT PAGES / ON THIS DEVICE
The old guess stays on the page.
Earlier pages
AFTER THE TEST / YOUR JUDGMENT
What met your guess?
This is your report, not a Buffer check or an approval. Selecting a report does not save it.
CLOSE THIS PAGE / UNSAVED ENDING
You can stop here.
You do not need to run another test. Choose your reason, then keep this ending.
Your kept plan and guess stay together. Nothing is saved by choosing a reason.
Could not save this rule on this device. It may not return after reload. Try again before leaving.
BEFORE THE TEST / UNSAVED PAGE
What will you put to the test?
Draft edits stay here for this visit. Your kept pages stay as written.
Both choices require human review before release. No draft goes out during this plan.
YOUR PRIVATE REVIEW RULEHOLD
BEFORE THE TEST / KEEP THIS GUESS
A kept page starts Not run yet. A revised page cannot borrow an earlier result.
An unsaved page is folded here.
A found record is your statement, not a verified fact. A person still checks and decides.
SEPARATE / UNSENT QUESTION
Ask Simo in your own words.
Your notebook stays here. Add only what you choose to this separate draft.
UNSENT / TO SIMOHOLD
No plan, prediction or report is copied into this draft. Review the exact words before opening your email app.
Add a subject and message before reviewing.
Reviewed copy · still unsent
Open unsent email draft ↗
Your email app opens with these words. You choose whether to send them.