AI Changes Where Thinking Happens at Work
AI adoption redistributes processing, memory and contextual judgement. Map all three before optimising a prompt or measuring speed.

The most visible effect of AI at work is speed. A model can draft, sort, compare and summarise in seconds. That makes “time saved” easy to discuss—and makes a deeper change easy to miss. When an organisation introduces a large language model, thinking does not simply become faster. Parts of the thinking move. Processing may shift into the model. Memory may sit in a project file, a decision log or nowhere reliable at all. Contextual judgement may remain with a person, become split across several roles or quietly disappear between steps.
That redistribution matters because a fluent answer can hide a weak arrangement. If the model processes quickly but lacks the organisation’s current decisions, it may produce polished work from stale premises. If a worker supplies rich context but no record is updated afterwards, the next person starts again. If nobody owns evaluation, speed creates more material without creating a better decision. The practical question is therefore not only “Should we use AI?” It is “Where will processing, memory and contextualisation live in this workflow?”
Three functions shape the work
A 2026 Joint Research Centre report proposes processing, memory and contextualisation as a way to analyse human–LLM work in organisations. The value of the framework is not a claim that every task has one perfect design. It gives teams a vocabulary for seeing what their adoption plan may have left implicit.
The three-function map
Processing
Transformation at speed: drafting, classifying, extracting, comparing or reorganising material. A model may do much of this work, but the task still needs a defined outcome and inputs.
Memory
The records that give work continuity: source files, state-of-play notes, decisions, terminology, versions, exceptions and previous failures. Chat history alone is rarely a sufficient organisational record.
Contextualisation
Judgement about why the task matters here: audience, consequence, policy, professional norms, local history and what remains uncertain. People and accountable roles usually carry most of this burden.
The functions are connected. Good processing depends on useful memory and context. Memory improves when people record what changed after the output is reviewed. Context becomes easier to apply when the workflow makes the decision, evidence and owner visible. Weak designs often over-invest in the middle—generation—while leaving the beginning and end informal.
| Speed-first workflow | Distributed-thinking workflow | |
|---|---|---|
| Before the model | Paste what is available and ask for an answer | Frame the decision, audience, boundary and source set |
| During processing | Optimise for a polished output | Use the model for bounded transformations that can be inspected |
| After the model | Edit the prose and move on | Evaluate claims, decide, record changes and update the working memory |
| Failure signal | Try another prompt | Ask which function is missing, stale or owned by the wrong place |
A worked example: a quarterly supplier review
Reading is a start. Practice makes it stick.
Start learningImagine a procurement team preparing a quarterly supplier review. The model can process contract notes, service reports and meeting records into a draft comparison. That is useful processing. It does not know which delayed delivery changed a production plan, which exception was accepted temporarily or which relationship constraint matters in the next negotiation unless those facts are present and interpretable.
Redesign the workflow around the three functions
- 1
Frame slowly
The category lead writes the decision to support, the suppliers in scope, the period, the evidence pack and the claims that require traceability.
- 2
Process quickly
The approved model extracts service events, groups them by agreed criteria and marks every missing or conflicting input instead of resolving it silently.
- 3
Judge deliberately
The lead and service owner compare the draft with source records, add commercial context and decide which issues are material.
- 4
Return memory
The final decision, unresolved questions, accepted exceptions and next review date are written back to the supplier record—not left inside the chat.
The model did not replace thinking. It changed the location and timing of it. Human effort moved upstream into framing and downstream into evaluation, ownership and record keeping. The surrounding memory made the fast middle useful. This is close to the JRC’s “slow–fast–slow” practical pattern: deliberate preparation, rapid model processing, then deliberate review and memory update.
Diagnose the arrangement before improving the prompt
When an AI-assisted task disappoints, teams often treat the prompt as the only adjustable part. The three-function map gives a wider diagnosis. A factual error may come from weak processing, but it may also come from a missing source. An inconsistent answer may reflect absent terminology or an outdated decision. A recommendation that ignores the organisation’s constraints may reveal that contextual judgement was never supplied or assigned.
- Choose one repeated AI-assisted task and write the real decision or work result it supports.
- List what the model processes and mark any transformation that is too broad to inspect.
- Name the records that should give the task continuity. Circle anything stored only in one person’s head or one chat.
- Write the context a competent colleague would need: audience, consequence, policy, local history and unresolved uncertainty.
- Assign an owner to framing, evaluation and memory update.
- Change one weak hand-off before changing the prompt.
Related Bokili guides can strengthen each part of the map. Use the AI workflow hand-off card to make ownership visible: https://bokili.com/en/learn/ai-workflow-handoff-card. Use the source-freshness check when memory may be stale: https://bokili.com/en/learn/ai-source-freshness-check. Use the decision-register method when summaries must preserve what changed and why: https://bokili.com/en/learn/teams-transcript-decision-register.
Design for continuity, not only acceleration
A good AI workflow is not the one that moves the most text through a model. It is the one that places each kind of thinking where it can be performed, checked and carried forward. Let the model supply bounded processing speed. Give people explicit responsibility for framing and judgement. Give the organisation a memory that survives the session. Once those three functions are visible, speed becomes a design choice rather than the only story of adoption.
Practise this in Bokili
Turn what you just read into a ten-minute mission.
- Start learning
Map where the thinking happens
Choose one AI-assisted workflow and assign processing, memory, contextual judgement and final ownership.
- Start learning
Repair one weak hand-off
Change the point where context, evidence or decisions are lost before changing the prompt.
Sources
- Human–LLM thinking in organisations — European Commission Joint Research Centre
- Thinking, faster and slower — European Commission Joint Research Centre
- AI Use Taxonomy: A Human-Centered Approach — NIST
Reading is a start. Practice makes it stick.
Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.
Start learningKeep reading

AI Training for HR: Keep Health Inference Out of Absence Summaries
Train HR teams to produce a minimal operational absence summary without exposing health details, inventing causes or letting AI make a people decision.

Allocate AI Licences to Task-Ready Work, Not Waiting Lists
Use four readiness gates to allocate scarce AI access to work with an approved input, clear outcome, review owner and support route.

AI Skills for Employees: Calibrate Confidence Before Checking
Train employees to record confidence before AI advice, verify the decisive claim and learn from the gap between prediction and evidence.