A guide to where AI helps your organization—and what to improve next.
Here, AI-native means designing work around useful collaboration between people and AI. Keep one team, department or company in mind throughout; “organization” means that unit. The levels describe useful gains from that collaboration. A company that uses little AI may already work very well.
In this guide, productivity means useful results for the full effort, time, attention and cost they take, including review and rework. Count everyone affected, including recipients and colleagues whose work is displaced.
| Level | What is true | Not yet established / keep going |
|---|---|---|
| 0 — No useful AI-enabled gain yet | AI is not used in the work being considered, or examination of its use shows no useful gain after checking, correction and other costs. | Find a use that improves the work for the full effort it takes. This says nothing about how good the team is. |
| 1 — AI produces useful gains in parts of the work | Individual tasks get faster or better, and the gains hold up in review. | A better complete result is not yet established; handoffs, waiting or rework may absorb local gains. |
| 2 — AI improves complete, worthwhile results | Complete results arrive sooner or better, or become practical at all, and people use them at justified total cost. | AI-enabled improvement in how people accomplish shared work is not yet established. |
| 3 — AI-enabled collaboration increases what the organization can accomplish together | People and AI combine knowledge, judgment and contributions to accomplish more valuable or demanding shared work, after counting costs to everyone affected. | Continuing improvement across repeated work is not yet established; existing gains can still last. |
| 4 — AI helps the organization keep improving how it works | AI helps improve methods and people’s capability; recorded comparisons across repeated work show worthwhile gains after the cost of learning and change. | A dependable two-way benefit between strategy and delivery is not yet established; strategic thinking can already occur. |
| 5 — Strategy and delivery improve each other | Changed direction reaches actual work, delivery evidence improves strategic choices, and those choices produce worthwhile results or supported stops. | Keep going: check whether choices remain worthwhile and where follow-through still consumes attention. |
Higher levels retain the useful gains below them. Earlier limitations are overcome, not inherited. Learning, strategy and judgment can exist at every level; the distinction is the AI-enabled benefit you can establish and sustain.
Recognition is a starting hypothesis, not a score. “Not yet” means a productive condition is unestablished, not forbidden. Use the exercise below to support a placement and choose an improvement. Measuring an existing benefit changes what you know; it does not by itself mean the work has improved.
Use this exercise with people who do, review and receive the work. Bring recent examples, including one that went poorly. The result is a provisional placement and one useful next action; this assessment method is still in development.
Keep a short record:
For [scope and period], our supported starting point is [level or unresolved conditions], because [evidence of benefit and full effort]. We still need to understand [gap]. [Owner] will test [change] against [comparison] and review it on [date], using [result, quality and cost checks] to decide whether to keep, change or stop it.
Hypothetical illustration, not an observed company result. A service team considers its standard client setups over the last quarter, from agreed requirements through configuration, training and early support. Bespoke development and sales are outside this assessment; their handoffs into the work are included.
| Question | The team’s answer in this illustration |
|---|---|
| What supports a starting point? | Setup records and client feedback show successful use with less total delivery effort through AI-assisted configuration and testing. Review, correction, training and early support are included. Comparison with similar earlier setups includes troublesome cases and checks for changes in difficulty and staffing. This supports provisional level 2 for the stated work. |
| What remains unresolved? | Trainers discover mistaken assumptions late and ask specialists to rebuild the explanation. Better shared investigation and decisions remain unestablished; separate AI sessions alone do not determine placement. |
| What will change? | The specialist and trainer will examine requirements together with shared AI context before configuration. The agent will trace agreed corrections into the setup and training materials. |
| How will they check it? | The delivery lead will compare client usability, late corrections, total effort and interruptions, including context upkeep. Review date: 30 November 2026, using setups with at least a month of client use. If too few comparable cases have finished, leave the conclusion open. |
| What decision follows? | Keep it if client results improve or total effort falls at the required quality. Change or stop it if upkeep and interruptions erase the gain. A useful trial supports further testing, not an automatic team-wide level 3. |
The examples are hypothetical. They follow the same service unit, which configures customer-support systems, trains client staff and supports them afterward. Success means clients can handle enquiries well at worthwhile cost.
“Enabled by” describes typical setups. Tools and autonomy do not earn a level; useful gains do. Personal assistants and shared agents can coexist. Human-initiated work qualifies when the same benefit holds at the same total burden.
Typical setup: personal assistants with supplied context and manual copying between applications; no AI use is also possible.
Producing a training guide with AI is quick. Making it trustworthy requires someone to reconstruct the context, check claims against the system, find omissions and correct errors. In this example, that effort consumes the drafting gain. Careful manual AI use can also produce worthwhile gains; copying alone does not determine placement.
Do next: choose a recurring task and define a usable result. Give AI the relevant sources and ask it to expose unsupported claims and missing requirements. Check against sources, requirements or tests, and count generation, validation, correction and recipient effort. Another AI’s agreement alone is insufficient verification. Keep the approach when a useful gain survives the whole process.
Enabled by: assistants that search project documents, read a repository or analyze data, guided by project instructions and sources people can check.
A specialist asks AI to compare client requirements with a proposed configuration. The resulting proposal is sound and takes less total effort, including validation and rebuilding context between sessions. The gain holds for this part of the work.
The wider engagement still depends on people carrying context between tasks. A better proposal may leave delivery with the same clarification burden or fail to resolve a conflicting sales promise.
Do next: follow one engagement from promise to actual use. Find where people repeat explanations, wait or redo work. Give AI the context and checks to address one break, then compare the complete result and full effort.
Enabled by: agents working in the tools where results belong: preparing a pull request with tests, writing a report in the shared document system, or routing work for approval.
An agent checks configuration against requirements and detects a mismatch before training. A specialist decides the correction; it is implemented and tested. The client successfully uses the service, with better quality, less total effort, shorter delivery time or newly practical capability.
Agree what “done” means. A report ready for review can complete a drafting assignment; a service fix needs the agreed checks and successful use. Human review counts in the effort. An agent may also build a temporary tool to test an otherwise impractical idea; evidence for stopping the idea can be a useful result.
Colleagues may still contribute through their usual handoffs. Better shared investigation and decisions remain unestablished.
Do next: choose work needing several people’s knowledge or judgment. Try shared AI context to examine requirements, challenge assumptions and develop the result together. Compare quality and full effort with the usual handoffs.
Enabled by: multiplayer AI—people working with shared agents and context, building on one another’s contributions and continuing the work through shared sessions, decision records and appropriate access.
Sales, engineering and delivery develop a proposal together. A delivery specialist challenges an assumption; the agent traces its consequences, prepares alternatives and updates the agreed proposal and implementation plan. Colleagues can understand and continue the work without one person repeatedly relaying context between private sessions.
Their combined understanding produces better commitments with less reconstruction and rework. They can investigate harder problems or design solutions that isolated contributions would miss. Count review, context upkeep and interruptions when checking the gain.
This can extend across projects: AI notices two engagements competing for a specialist, prepares options and carries the agreed resolution into affected plans. It pursues missing follow-up within its remit. Shared work within one project can also demonstrate the distinction.
Do next: use a recurring training or handoff problem to test continuing improvement. Have AI assemble cases, propose an explanation and prepare a change. People decide; agents maintain the agreed experiment and follow-through. Compare subsequent results, including learning costs.
Enabled by: records of repeated work, including quality, cycle time, rework, recurring problems and people’s demonstrated skills.
AI notices repeated training difficulties and brings the cases to a debrief. People test its explanation and try better discovery questions, guidance and practice for newer specialists. Agents carry the change into later work, check whether it is used and follow its effects without repeated managerial prompting.
Later engagements show fewer repeat problems and more people handling demanding cases without senior rescue. The benefit justifies learning and change costs. Failed experiments remain visible; some cycles can stall or worsen. Level 3 gains may already last, and colleagues may already learn from each other. Level 4 adds demonstrated continuing improvement in methods, capability and results.
An engineering example follows the same logic: identify missing tests in returned pull requests, try a change to guidance and compare later review effort and production defects. Fewer returns alone could mean weaker review; better agent evaluation scores alone cannot establish better work.
Do next: connect this learning to a live strategic choice, such as serving more demanding clients. Have AI assemble customer evidence, constraints and capability gaps. Agree a bounded test and evidence for continuing, changing or stopping it.
Enabled by: linked strategy and delivery evidence, clear decision authority and agents able to carry agreed changes into affected work.
The service unit is pursuing larger clients. AI connects repeated delays and support demands to assumptions behind that direction, assembles the evidence and brings options to leadership. People decide to narrow the target segment. Agents identify affected opportunities, plans and training needs, carry through agreed changes and track the results.
The connection also works when leadership initiates a change: AI explains its implications, exposes conflicts and carries decisions into action. Subsequent delivery evidence helps people reconsider their choices. Better commitments and worthwhile results follow while existing obligations remain served and people develop the needed skills and relationships.
This two-way benefit must operate dependably across the assessed work. A CEO’s insightful conversation with a personal assistant about an investment may be a local gain; its strategic subject does not establish level 5. Neither does one successful cycle or a result outside the stated scope.
Keep improving: examine both directions—strategy into work, and experience into decisions. Find stale assumptions, repeated human chasing or proactive activity that creates noise. Fix the break and check the effect on results and attention.
AI skill, sometimes called “taste”, belongs throughout: define good work, delegate, check against sources, and decide when to trust, redirect or stop AI. People retain judgment and the ability to handle consequential choices and failures. At level 4, check their own development as well as assisted output.
Useful initiative follows notice → understand the implications → act or bring a decision → follow through → check the result. A notification leaves most of that work undone. Give agents relevant context, responsibility and a clear remit; keep people able to challenge and redirect them.
Count context switching, reconstructing sessions, interruptions and supervision. Agents can absorb these burdens or add to them. Who starts a routine does not determine its level; repeated chasing matters through its cost and fragility.
Permissions and human sign-off should match the consequences of an action. Routine actions can proceed within an agreed remit. Sandboxing and sovereign processing depend on the work’s sensitivity and obligations at any level.
This matrix helps turn recognition into a check.
| Level | Look for | Insufficient on its own |
|---|---|---|
| 0 | No AI use in the stated scope, or examined work showing no useful gain after full costs. | Missing evidence, manual AI use, or low AI use as proof of poor general productivity. |
| 1 | Useful local work that improves after review and correction costs. | More prompts, drafts or apparent speed. |
| 2 | A result accepted and used for a real need, with a justified gain in quality, full effort, time or feasible capability. | A completed ticket or polished deliverable nobody uses. |
| 3 | Better shared investigation, creation, decisions or delivery through people and AI combining their contributions, after counting review, context upkeep, interruptions and costs to others. | Shared channels, agents, more visible activity or a reported absence of conflict. |
| 4 | AI-assisted improvement carried through into changed practice and demonstrated skill, with recorded gains across repeated work after learning and supervision costs; failed experiments remain visible. | Debriefs, reminders, training attendance or an improving dashboard alone. |
| 5 | AI contributes in both strategy/delivery directions: evidence changes choices, choices reach work, and worthwhile outcomes or supported stops follow at justified total burden while existing obligations remain served. | Current strategy documents, management visibility or rapid reprioritization without better outcomes. |
Use the exercise’s comparison and time period throughout. Combine records of effort, elapsed time, correction and recurrence with evidence of quality, actual use and demonstrated skill. Note changes in difficulty, staffing or demand. A stronger result can justify greater time or cost; make that trade-off explicit.
For placement, check the productive condition across the stated scope. A strong example can reveal a possibility without establishing a team or company level. Record gaps and mixed conditions; do not average team levels. If complete delivery already includes improved collaboration, acknowledge both conditions rather than inventing an extra adoption step.
The six levels and placement method draw on research and practitioner experience. The sources below shaped the design. The operating examples illustrate conditions to look for.
Use the guide to place a team, department or company and choose a testable next action, keeping uncertainty visible.