Redesign Work Around AI, Not Just Individual Tasks

In one sentence

The paper asks what changes when a business designs its work around artificial intelligence, or AI, rather than adding AI tools to existing tasks. Its answer is an operating model—a way to organize decisions, people and processes—in which work produces information that can help improve later work. Consider a hypothetical supplier-approval process. An AI system might gather documents and flag missing information, while a person decides whether to approve the supplier. The team would record corrections and outcomes for later review. That example illustrates the paper’s distinction between changing a whole workflow and merely making one document task faster; it is not a reported deployment. The World Economic Forum and Kearney developed their blueprint through engagement with more than 50 enterprises and thinkers, drawing on interviews, workshops and consultations with more than 150 executives and experts. They describe five connected building blocks: an ‘intelligence engine’ that uses information from work to inform later decisions; an adaptable collection of data and AI tools; redesigned operations; human–AI teams; and choices about what AI-enabled value to offer customers. The blocks are presented as a lens for design, not a sequence every company has been shown to complete.

Key concepts

• AI-enabled versus AI-first: An AI-enabled company uses AI within existing tasks. In the paper’s definition, an AI-first company redesigns workflows, roles and decision rights so AI is integral to delivering its work.
• Intelligence engine: The authors’ name for a system that connects business goals, operational information and outcomes. A feedback loop means checking what happened and using that information to guide a later cycle; the paper organizes these loops around speed, scale and scope.
• Modular technology stack: Data, AI models and the connections between tools are arranged as replaceable parts. Orchestration is the set of rules that directs a task to a tool and governs how it proceeds; the authors argue that companies need control over these rules and data access as models change.
• Human–AI teaming: The paper asks teams to decide task by task whether AI acts, assists or stays out. People retain responsibility for judgement, accountability and decisions with the highest stakes.
• Business model canvas: The authors extend this planning framework—which asks how a business creates, delivers and earns value—with questions about AI, including how an AI-enabled offering is presented to customers.

A concrete example

Hypothetical illustration, not a finding: A procurement manager chooses supplier approval as a workflow to examine. The team maps its steps, lets AI draft a financial review for a person to check, keeps final approval with an authorized employee and records errors for the next review. The question is whether the complete process works better, not how many AI drafts it produces.

What the researchers measured

The white paper reports company examples rather than a single measured test of its five-block blueprint. In its Gamma case study, the authors say the AI-native presentation company’s inference-related gross margin moved from approximately 31% to approximately 77% six months after launch. Gross margin here describes the share of revenue left after the relevant costs; these figures concern Gamma’s inference-related measure, not every AI-first company. In a separate Claryo warehouse-platform example, the authors report an annualized profit-and-loss gain of $2.56 per square foot. That is presented as Claryo’s reported figure, not as a measured result for the blueprint or for other warehouse systems. The paper also describes practices and targets at other named organizations; those should not be read as outcomes achieved by Gamma or Claryo.

Why it matters

The distinction matters because making a task faster does not, by itself, change who makes a decision, how a handoff works or whether the organization learns from an outcome. The authors propose examining those connections together. For a working team, the immediate use is to ask where a process starts and ends, what information it needs, who can intervene and what evidence would show improvement.

Where it might help

As possible uses of the blueprint, a leader could choose a high-volume workflow, assign human and AI responsibilities at each step, set checks before any change goes live and decide how to compare the redesigned workflow with the existing one. These are planning uses, not deployments tested by the white paper.

Impact across sectors

• Possible application in financial services: A team could examine where document checks might be assisted by AI while people retain responsibility for consequential decisions. This is not a demonstrated outcome of the blueprint.
• Possible application in warehousing: Managers could examine whether live operational information helps them spot delays and decide when to intervene. Results from a named company in the paper should not be assumed to apply to other warehouses.
• Possible application in customer support: A service team could map handoffs and decide which requests need a person, which might be assisted and what feedback to retain. The paper does not establish the outcome of that hypothetical redesign.

Where the evidence stops

The authors say AI-first enterprises are still at an early stage. They state that it is unclear which approaches will work best across industries, how broadly they will scale and what the long-term effects will be on productivity, employment and competition. They also say it is not yet clear which operating models will prove most resilient over time. The company accounts in this white paper do not amount to one common evaluation of the blueprint. PtoP note: this is a collaborative white paper, not a reported controlled test or a verified implementation guide.

In one sentence

The study asks whether a video generator can show a hand completing an everyday task from a starting image and a goal, without being told every step. The authors built Ego2Act to test this in first-person videos, where hands can hide objects and the view moves with the person. In the paper’s travel-case example, the toothpaste needs a cap and the toothbrush needs folding before both can be packed and the case closed; more than one order of preparation is possible. The authors also built Ego2ActJudge to check generated videos against the goal without requiring a single ‘correct’ reference video.

Key concepts

• A goal-directed prompt states the desired outcome, leaving the generator to work out intermediate actions. The travel-case goal, for example, does not prescribe which item to prepare first.
• A subgoal is one observable part of the task, such as capping the toothpaste. Checking subgoals lets the evaluator notice a missing preparation step even if the case appears closed.
• Task completion asks whether each action starts, takes place and leaves the required result. The checks stop at the first failure.
• Physical plausibility separately asks whether objects remain consistent and whether their movements, contact and resulting positions make physical sense. An omitted action is not itself scored as a physical interaction.
• Reference-free evaluation allows different valid routes to the same goal. Ego2ActJudge derives subgoals from the starting scene and goal, then examines the video rather than comparing it with one prescribed recording.

A concrete example

Hypothetical illustration, not a study result: A generated clip shows a hand put an unfolded toothbrush into a travel case and close it. A reviewer would check the folding subgoal and the final state separately, rather than crediting the clip simply because the case closes.

What the researchers measured

The authors report a benchmark of 110 real-world cases and 2,640 videos, including human recordings and outputs from six generators. On a human-rated panel covering 25 cases, Seedance-2.0 had the highest overall generator score in the main results table: 64.0 on a 0–100 score combining task completion and physical plausibility. That is a rubric score, not a percentage of tasks completed. On the 600-video human-rating panel, Ego2ActJudge’s final scores had a correlation of 0.69 with the combined human ratings; correlation describes how scores vary together, not exact agreement. The authors report recurring omitted or incomplete steps and errors in detailed object interactions.

Why it matters

A video may show an appealing end scene while skipping an action needed to get there. By separating goal completion from physical plausibility, the authors make those different kinds of failure visible. Their automated judge is intended to support comparisons across many recordings, while the paper cautions against treating one clip’s score as a definitive verdict.

Where it might help

As a possible research use, the benchmark could help teams compare video generators intended to depict multi-step object handling. Its scores are evidence about generated recordings in this benchmark, not evidence that a model can carry out a task in the physical world.

Impact across sectors

• Possible use in robotics research: inspect proposed training videos for omitted actions and implausible contact before considering them as demonstrations; this was not tested as a robot-training deployment.
• Possible use in instructional-video development: make a checklist of intermediate outcomes that a generated demonstration ought to show; the paper did not test learners or teaching outcomes.
• Possible use in video-model development: compare how generators depict everyday organization or personal-care tasks; these are benchmark settings, not proven workplace applications.

Where the evidence stops

Human ratings covered 25 of the 110 cases, rather than the whole benchmark; automated evaluators scored the benchmark videos they could score. The authors say Ego2ActJudge tends to score executions above human raters, misses many brief errors that unfold between sampled frames, and varies across runs on an individual video. Some action-category groups are small or overlap, and the paper reports that its tested feature associations did not survive correction for multiple comparisons. Generated clips were limited by the models’ available durations. The authors also say baseline evaluators needed adaptations for this setting. This source is an arXiv preprint, not presented here as peer-reviewed work.

In one sentence

A model can get credit for naming a tool without issuing the structured request needed to use it. Santillana compared two Spanish-language security models that received nearly identical scores from a keyword-based test. He then checked their actual output, starting with examples from their training data and moving to prompts containing unfamiliar details. He also examined the models’ likelihood of starting a tool call and, after fine-tuning the larger model, whether the part of the model representing the call-start token had changed. The comparison suggests that training-data composition matters for how readily this format appears, but the models differ in more than their data mix, so it does not isolate a cause.

Key concepts

• A tool call is a structured request naming an available tool and supplying its arguments. Mentioning the tool in ordinary prose is not the same action.
• A keyword harness is an automatic grader that looks for words in an answer. The paper’s tool-use harness could award credit when a tool name appeared without a valid call.
• A verbatim-reproduction check gives a model a real training prompt and asks whether it produces a valid call. The authors distinguish a sensible new argument from an exact copy of the training answer.
• A first-token probe checks how likely the model is to begin its answer with the special token that opens a tool call. It helps locate a format failure without depending on a generated answer.
• A generalization battery uses new prompts to test tool choice, argument details and when not to call a tool. Success on familiar training prompts alone could reflect copying.

A concrete example

Hypothetical illustration, not a paper result: A security worker asks whether a particular vulnerability appears in a catalog. An answer saying that a catalog-checking tool exists would satisfy a loose word match; a tool call would instead name that tool in the required structure and put the requested vulnerability identifier in its argument.

What the researchers measured

On the authors’ single-seed keyword harness, VectraYX-600M at training step 154K scored 0.660 on tool use, while the recorded VectraYX-1B score was 0.650. The authors identify the latter as unreliable because the grader could credit a tool name in prose. In a stricter check using real training examples, the unfinished VectraYX-600M checkpoint produced valid calls with new, sensible arguments on 6/6 examples; the tested VectraYX-1B checkpoints before repair produced the structure on 0/4–6 examples per run. These counts are examples passing the specified check, not a deployment success rate. After a targeted fine-tuning run, the reproduced repaired VectraYX-1B produced well-formed calls on 0.959 of 269 corpus rows, compared with 0.100 for its pre-repair parent and 0.926 for the VectraYX-600M checkpoint. The authors say the repaired model mostly copied answers on its own corpus. On the call-requiring portion of their new-prompt battery—166 prompts about entities absent from the training corpus—the repaired VectraYX-1B passed the combined tool, argument-shape and entity check on 0.536, versus 0.428 for VectraYX-600M. The wider battery contained 238 prompts, including prompts on which a call was not wanted. On no-call prompts mentioning a vulnerability or shell command, the VectraYX-600M answered without a call on only 0.09 of items and the repaired VectraYX-1B on only 0.17. A weight check found that the repaired model’s call-start token representation was essentially unmoved; the authors locate the change in the surrounding network, within the limits of that check.

Why it matters

For work that depends on a model invoking a tool, the grader must recognize the required request, not merely related words. The paper also shows why testing when a model should refrain matters: both the smaller model and the repaired larger model often called a tool on prompts that only mentioned a vulnerability or shell command.

Where it might help

A possible use is to add a strict output check to an internal model assessment before considering a model for a tool-connected workflow. The paper tests model outputs; it does not establish a deployed security assistant.

Impact across sectors

• Possible security-team use, not a tested deployment: distinguish a model that talks about vulnerability lookups from one that formats a request for the declared lookup tool.
• Possible model-development use, not a tested deployment: check familiar and unfamiliar prompts separately after fine-tuning, so copied training answers are not mistaken for generalization.
• Possible software-procurement use, not a tested deployment: ask for valid-call and no-call checks alongside a vendor’s keyword-based tool-use score.

Where the evidence stops

This is a preprint, not presented as peer-reviewed work. The VectraYX-600M run stopped permanently at 64% of its planned schedule. The two models share an architecture family and tokenizer but differ in size, training volume and curriculum; their comparison is a natural experiment, not a controlled test of data composition. The authors’ trajectory probes also overlap with or closely resemble the larger model’s earlier training examples: an early apparent tool-call tendency was exact copying, not demonstrated generalization. The original successful repair checkpoint and its parent were lost, so the primary strict results use a reproduced run from a surviving sibling checkpoint. The new-prompt battery was built by the authors; it does not cover unseen tools or multi-call sequences. The authors report no evidence that the repaired larger model answers ordinary conversational questions usefully. Harness figures are single-seed, and the repair is one recipe. In a separate set of five fine-tuning runs using a different numerical setup, the apparent benefit of a diverse corpus for avoiding unwanted calls was not reproduced by a partial second seed; the authors retain it as a hypothesis, not an established effect.

In one sentence

A drawing program can run correctly and still make the wrong picture. Imagine asking for a sun to move behind a hill: the code might produce a video, but put the sun in front throughout. The authors call this mismatch the Program-to-Visual gap. Their MaLiang-Harness has a multimodal language model—one that works with both words and images—write drawing or animation code, then inspect the rendered result and revise it. A renderer is the tool that turns the code into an image or video. The system keeps earlier versions and links each inspection to the version that produced it, so an old approval does not stand in for a check of the latest edit.

Key concepts

• Program-to-Visual gap: code that executes successfully can still miss a requested position, appearance or action.
• Persistent Executable Generation: the harness keeps the artwork’s code, assets, requirements and plan across revisions, or saved versions.
• Traceable Generation Process: recorded operations connect code changes to the visual results inspected after those changes; this is not a record of the model’s private reasoning.
• Revision-aware Editing and Verification: earlier artwork can be restored, while the current revision needs its own checks before delivery.

A concrete example

Hypothetical illustration, not a study result: a designer requests a poster with a small orange sun above a navy mountain. If the first rendering puts the sun behind the mountain, the designer could inspect that version, change the drawing code and check the new rendering rather than treating error-free code as a finished poster.

What the researchers measured

In the authors’ MaLiang-IBench evaluation, GPT-6-Astra produced a successful output for every image task; 48 of 50 tasks met all the image-quality thresholds. That count means the outputs passed the paper’s separate checks for following the prompt, appearance and composition, as judged on successful images by GPT-6-Sol. In the authors’ MaLiang-VBench evaluation, GPT-6-Astra produced a successful output for every video task; 10 of 13 tasks met all the video-quality thresholds, which also include motion coherence. A successful output means a decodable file passed the harness’s completion checks, not that it met those separate quality thresholds.

Why it matters

The authors separate producing a usable file from satisfying the visual brief. They also report that similar scores on a general model-capability measure did not necessarily correspond to similar results on their image tasks. For someone choosing a model for this kind of work, their findings point to checking the rendered work against the actual brief rather than relying on code execution or a general score alone.

Where it might help

Possible uses include artwork with precise layout instructions and silent animations with ordered actions. These are potential uses of the approach, not deployments demonstrated by the paper.

Impact across sectors

• Possible use in graphic design: a team could inspect saved poster revisions when a change fixes one layout requirement but affects another. This is hypothetical, not a measured workplace outcome.
• Possible use in animation production: a team could compare sampled moments of a short coded animation with its action brief. This is hypothetical and would not, by itself, establish smooth motion between samples.

Where the evidence stops

The authors say the harness’s visual reviews are model self-assessments, not independent measures of perceptual quality. Their separate video-quality review used one reviewer, was not formally blinded and examined sampled frames; it cannot establish smoothness between frames or rule out brief faults. Image results combine batches, retries and historical runs rather than uniform first attempts, and DeepSeek-V4-Pro was evaluated without visual feedback. Timing and budget conventions also differ across some evaluated configurations, limiting direct cost comparisons. In their examples, finer brushwork did not make a stylized image photorealistic, and refinement could stall or exhaust its token budget, the limit on text the model could use. General PtoP note: this source is an arXiv preprint, not a claim of peer review, and its benchmark results are not a workplace deployment.

In one sentence

Can a language model learn new question-and-answer pairs without quickly losing old ones? The authors found that combining ways to protect earlier learning helped the model retain more answers than using any one of those ways alone. Imagine teaching a model a fictional fact today, then many more facts over the following weeks: the question is whether it can still give the first answer. The authors trained Qwen3-4B-Base on streams of questions and answers, one task after another, without keeping earlier raw examples for later training or telling the model which task a question came from. They compared safeguards that regenerate practice material, encourage the updated model to behave like its previous version, and limit changes to parts judged important. They also compared reusing one small set of trainable parts with folding each task’s changes into the model before starting a fresh set.

Key concepts

• Continual learning means updating one model on successive tasks while trying to retain what it learned earlier.
• Generative replay supplies practice material made by the previous model instead of saving raw examples from earlier tasks.
• Self-distillation asks the updated model to stay close to the previous model’s predictions on the current task’s inputs.
• A weight anchor limits changes to trainable values estimated to matter for earlier learning.
• Merged LoRA is a rule for small trainable model updates: fold one task’s update into the model, then start a fresh update for the next task.

A concrete example

Hypothetical illustration, not a study result: a model learns the answer to a question about an invented town, then learns many unrelated question-and-answer sets. A memory test asks the original question again. The safeguards are intended to reduce the chance that later training overwrites its answer; this example does not show how it would respond to a reworded question.

What the researchers measured

After 100 tasks, the authors report 1.2% average final retention for naive sequential fine-tuning and 34.9% for their method combining all three anchors with merged LoRA. These figures average across Symbol-QA, LLM-QA and Real-QA using Qwen3-4B-Base; final retention is the share of training questions answered correctly across all tasks after the last update, averaged over tasks. The method ranked among the top 3 compositions on each dataset, though a different composition had the highest mean retention on Symbol-QA and Real-QA. In their factorial comparison, the authors identify replay and merged LoRA as the largest average contributors and report a positive interaction between them on all three datasets. The final experiments used three training seeds per method and dataset.

Why it matters

The authors separate learning a new task from remembering it after later updates. Their results suggest that, in this memorization setting, the combination of replay and the rule for retaining updates matters more than choosing a single safeguard in isolation.

Where it might help

A possible use is designing experiments for models that receive successive batches of information. The paper does not demonstrate a working deployment or show that the model can reliably answer newly phrased questions.

Impact across sectors

• Possible, not a tested deployment: a publisher updating a model with successive batches of reference facts could use the study’s retention measures to plan an experiment.
• Possible, not a tested deployment: a support team training on successive sets of standard questions could compare whether old, identically phrased questions remain answerable.
• Possible, not a tested deployment: an education team updating a question-answer model over time could test recall of earlier training questions, while separately checking unfamiliar wording.

Where the evidence stops

The authors tested recall on the same questions used for training, so the results do not establish that answers survive rewording or transfer to new questions. They report that even the stronger compositions continue to forget older material. Stronger memorization also did not preserve general ability: the evaluated methods lost substantial accuracy on separate reasoning and knowledge tests. Their search held task order fixed across training seeds, so those seeds did not measure sensitivity to task order; the later evaluation used a different order, and the search winner was not the final leader on Symbol-QA. General PtoP note: this source is an arXiv preprint, not evidence of peer review or a deployed system.

In one sentence

Instead of asking a student model merely to copy an improved teacher, the authors train it toward the internal change that separated the teacher from its earlier version. Imagine a student working through a math problem. On the same partial answer, the method compares signals inside the teacher with signals inside the model before reinforcement learning—a training process that rewards desired responses. It then sets a target farther along that difference and trains the student’s internal signals toward it. The authors call the method RIDE. This is a training target, not a guarantee that going farther will produce a better answer.

Key concepts

• A student is the model being trained; its teacher is a model already improved through reinforcement learning. In this study, the student begins from the teacher’s earlier checkpoint, or saved version.
• On-policy training means the student generates its own partial answers. The teacher and earlier checkpoint are then compared on those same partial answers, rather than on different responses.
• A hidden state is an internal set of values a model builds while processing text. RIDE compares hidden states at each layer, or stage of processing, before they are turned into probabilities for the next piece of text.
• The residual is the difference between the teacher’s hidden state and the earlier checkpoint’s hidden state on the same text. RIDE uses that difference as a direction for its target, rather than treating the teacher’s state as the endpoint.
• The extrapolation coefficient controls where the target sits. At a value of 1, the target is the teacher’s state; above 1, it lies farther along the measured direction.

A concrete example

Hypothetical illustration, not a paper result: a student begins a competition-math solution by considering one approach. The earlier model and the improved teacher process that same unfinished solution. RIDE measures the difference between their internal signals and sets the student a target beyond the teacher’s signals in that direction. Whether the student’s finished solution is correct must still be checked.

What the researchers measured

The authors evaluated four base-model and reinforcement-learning-teacher pairs on three competition-math problem sets. Their score, Avg@16, averages whether sampled answers are correct across 16 responses per problem, then averages the three problem-set scores; it is reported as a percentage. For the DeepSeek-R1-Distill-Qwen-1.5B and JustRL-DeepSeek-1.5B pair, RIDE’s mean score across three training runs was 56.38, while the fixed teacher’s score was 55.30. The authors report that RIDE’s mean was above its teacher’s score on every pair and above the methods they compared it with. On three pairs, however, the margin over the teacher was within one across-run standard deviation. These are measured math-evaluation findings, not deployment results.

Why it matters

The comparison distinguishes two ways of learning from an improved model: copying what it produces and using how its internal processing changed. The authors report that their internal-state approach performed differently from methods that pushed beyond the teacher using output probabilities.

Where it might help

A possible research use is training a student when a team has both a model improved through reinforcement learning and the checkpoint it started from. The paper does not establish a ready-to-use service or performance outside its competition-math evaluations.

Impact across sectors

• Possible, not tested as a deployment: language-model training teams could investigate whether saved before-and-after checkpoints provide a useful training direction when building math-focused models.
• Possible, not tested as a deployment: competition-math evaluation teams could use the paper’s comparison as a starting point for studying how different training targets affect answers.
• Possible, not tested as a deployment: AI safety teams could examine whether training beyond a teacher’s internal signals also carries forward its errors or biases.

Where the evidence stops

The authors say RIDE requires the checkpoint from before reinforcement learning and models whose internal states can be compared. It uses one global extrapolation setting and was evaluated only on mathematical reasoning with one reinforcement-learning recipe. They note that the two AIME evaluations were near the scoring floor for the Llama-3.2-3B pair, so evidence for that pair rests mainly on AIMO. They also say a student could carry forward or amplify the teacher’s errors or biases; its calibration and safety need independent evaluation before deployment. The paper does not report an analysis of overlap between training prompts and evaluation problems. PtoP note: the supplied source is an arXiv preprint, not evidence of peer review.