Assist, Augment, Adapt: A practical guide to AI-native work

AI can now do real work. What should it do, and what stays yours?

If you have just joined The Helix Loop, welcome. This is the first thing I'd want you to read. Everything else on Agent Watch, from tool pages and comparisons to the weekly update, links back here. I wrote it from my own practice: running proposals, briefings and recurring reports with AI over the past 18 months, getting some of it wrong, and slowly working out where the line sits.

The short answer

AI-native work means designing the job around what AI can do and what people must still decide. Assist helps you think. Augment does several steps, you check. Adapt works on its own within your rules. Choose the level for each task, rather than giving a tool permission to do everything.

I start with the work: a client proposal, a policy brief, a council report or a leadership meeting. What needs to arrive at the other end? Which sources can I trust? What am I prepared to put my name to? Answer those first and the tool choice gets much easier.

Mode	What you do	What AI does	Example	What stays yours
Assist: helps you think	Ask, challenge and decide	Offers ideas and challenges	Suggest questions for a stakeholder interview	Choosing the question and making the final decision
Augment: does several steps, you check	Set the brief and review	Carries several steps	Assemble a cited briefing from approved documents	Approving the output and any changes to files
Adapt: works on its own within your rules	Set rules and stay on call	Works on files within those rules	Prepare a recurring report and flag exceptions	Setting the rules, handling exceptions and deciding when to stop

These follow Craft from Helix Lab (checked 17 October 2026): Assist is "AI helps me think. I remain responsible for the decision." Augment is "AI helps carry several steps. I stay in the loop and review." Adapt is "AI works on files in a workspace I control. I stay on call." The modes describe how I hand over a particular job. They are editorial guidance, not vendor certifications.

If this is your first visit, start here. If you already have a task in mind, use the tools directory alongside this guide. Neither a familiar brand nor an impressive demo tells you how much responsibility to hand over.

From asking a chatbot to having AI do the work

A chatbot conversation usually leaves me moving information between places. I ask for a draft, copy it into a document, find the missing source and ask again. AI helps a lot, but I'm still the one coordinating the work.

An agent is AI equipped to take steps towards a goal using tools, with some freedom to choose the next step. That might mean reading authorised files, comparing their contents and saving a draft report. The practical difference is whether I get advice about the work or a finished output produced through several actions.

The distinction is visible in official product descriptions checked on 17 October 2026. Anthropic's Cowork page describes work in selected folders and tools, delivered for review. Microsoft's Copilot Cowork page describes work across apps and systems. These descriptions establish capabilities, not suitability for your particular records or decisions.

Take a monthly operating report. Asking for a clearer introduction is Assist. Having AI gather approved figures, calculate changes and draft the report for checking is Augment. Letting it prepare that report on a schedule, within fixed access and escalation rules, is Adapt. Publishing the report remains a separate permission.

I judge progress by the usefulness of the finished work and the effort needed to check it. A polished document with missing evidence is unfinished. A short draft with traceable sources and clearly marked gaps may be much more valuable.

Assist, Augment, Adapt in one picture

Individual: Assist → Team: Augment → Organisation: Adapt

This picture describes a useful learning journey, not a restriction on who can use each mode. An individual can delegate a bounded recurring job. An organisation can keep a sensitive decision at Assist. Moving right means changing who does what and who answers for it, not simply buying a more expensive subscription.

Assist: helps you think

I keep the work close. AI proposes an outline, challenges an assumption or helps me see alternatives. I choose what to pursue and write the final argument. AI never acts outside the conversation. This is a good place to begin when the task is new or the evidence is hard to judge.

What stays yours: choosing the question, weighing the options and making the call.

Augment: does several steps, you check

A team agrees on inputs, a procedure and who reviews the result. AI can then prepare several parts of the deliverable in one go. For a consultation summary, that means grouping responses, keeping source references and drafting themes. Someone checks whether the themes fairly represent the submissions. Nothing is saved over the originals or sent without your say.

What stays yours: approving the summary and any changes to the original files.

Adapt: works on its own within your rules

An organisation makes the boundaries repeatable: approved sources, permitted actions, how exceptions are handled and who is accountable. AI can carry on without a new instruction for every step. I still need to know how to stop the work, and I keep a history of what it did so I can recover from a mistake.

What stays yours: setting the rules, handling exceptions and deciding when to stop it.

A useful next step is to map out one AI-native workflow: the inputs, the steps, the output and the points where a person must step in. The seven layers below are the one map you need to explain how it works.

The 7 layers of any AI tool

The AI behind it: the model supports every mode

A model is the underlying AI system that generates responses from the information it receives. It helps interpret a question, draft text or reason through a problem. Model quality matters, but it does not tell me which files the app can reach or whether it can send a message.

Start with the tool profiles in the tools directory, such as Claude Cowork, Copilot Cowork and ChatGPT.

For a choice between products, follow Claude Cowork vs Copilot Cowork or Claude Chat vs Cowork vs Claude Code. Compare a representative task with sources you are allowed to use. Check the result before treating fluent writing as sound judgement.

How it works with you: the software that lets it take steps

Around every model sits software that manages instructions, tools and progress. It decides how AI works with you, and it is what makes anything beyond Assist possible. Does it wait for each message? Can it plan several steps, work in files and come back with questions?

Two products built on related models can behave quite differently because the software around them differs. For a briefing, I want to see what it read, where it saved the draft and what remains unresolved. A 'done' message alone doesn't prove the job is done.

Choosing how a tool works with you explains the wider pattern. Keep the deliverable explicit: a source list and a draft memo are easier to inspect than an instruction to “research everything”.

What it can connect to: connections make Augment practical

A connector gives an AI app access to another system, such as a document store. MCP, short for Model Context Protocol, is an open standard for connecting AI applications to external data and tools. The official MCP introduction, checked on 17 October 2026, explains that connection role.

A connection can reduce copying and help AI find working context. It also changes what the tool can touch. Reading a folder, editing a document and sending an email are different permissions. I check them separately, using the narrowest access that can complete the job.

Read what MCP means for work. Not every product supports every connection, so confirm support and access in the product you will actually use.

What it knows how to do: skills support Augment and bounded Adapt

A skill is a reusable set of instructions and supporting material for a task. It might explain how to structure your policy brief, apply a report template or check spreadsheet assumptions. A connection provides access; a skill explains a procedure. Neither establishes that the source material is correct.

I'd begin with a job I can already explain to a colleague. Specify the required sections, acceptable sources, checks and reasons to stop. Include one good example and explain what makes it good. Only reuse the instructions once they reliably produce work someone can review.

Use skills vs MCP vs plugins to distinguish procedures from access. A plugin is a package that can bundle such capabilities. Then follow how to build a skill. Avoid turning an unsettled judgement into a reusable instruction simply because it is easy to save.

The app you open: the product makes each mode usable

The product is the app and service you actually use, including its account settings and controls. I choose it based on where the work happens and the review process it supports. A pleasant writing experience can't make up for the wrong permissions.

Check each product's profile in the tools directory when checking fit, and start with the job rather than the brand.

Names also change. On 17 October 2026, Anthropic's Cowork page calls the product Claude Cowork and places it next to Chat. I keep “Cowork” here so the working pattern stays recognisable. Verify the controls in your actual account rather than relying on a product name.

How much it does on its own: agency defines Adapt

Agency means how much freedom the system has to act without another instruction. Digital labour means AI performing work through actions, within a defined scope. It is useful only when the permissions, expected output and person responsible are clear.

OpenAI's dot guide, checked on 17 October 2026, documents custom rules covering autonomous action, explicit pre-approval, confirmation and hand-off. It also warns that a dot can make mistakes when following rules. Setting a rule does not remove the need for checking.

The Grok Bot overview, checked the same day, describes persistent work across tools and returning for approval. Those are product-specific behaviours. I'd test the intended action before relying on either product's delegation controls.

Open the tools directory and look for “Works on its own”. Scheduling a task is only one part of handing it over. Handling exceptions matters just as much.

What stays your call: my role sets the boundary in every mode

My role means the decisions, checks and accountability I retain. A human-in-the-loop is a person who reviews or intervenes at a defined point. “Someone checks it” is incomplete unless that person has the sources, time and authority to reject the work.

I use a three-question check:

1. What are the stakes if this is wrong? Consider the people affected and how hard the action is to reverse.
2. How quickly can I check the result? If checking requires specialist investigation, reduce the scope before delegating.
3. Does this need my name or authority? Keep the decision with the person authorised to make it.

For government work, the public-sector section below covers the shared questions of rules, records and data sovereignty, meaning which jurisdictions and authorities govern the data. A connection does not establish permission to use a record for every purpose.

By role

These are suggested starting tasks drawn from knowledge-work areas such as evidence gathering, deliverable drafting, reporting and research. They are workflow examples, not claims that a named organisation has adopted them.

Consultants

Typical starting mode: Augment (does several steps, you check). Prepare prospect research, draft a proposal or statement of work, and turn approved findings into a client presentation. Ask for assumptions and missing evidence alongside the deliverable.

What stays yours: making the recommendation and any commercial commitment.

Policy teams

Typical starting mode: Augment (does several steps, you check). Collect evidence, compare source documents and synthesise findings into an options brief. Keep dissenting evidence, and separate what a source claims from the team's interpretation. The final say sits with the responsible officials.

What stays yours: deciding the recommendation and advice that go to the decision-maker.

The public-sector section below covers the rules that apply.

Local government

Typical starting mode: Assist (helps you think), then bounded Augment. Start with document processing, knowledge retrieval and operational reporting: locate an approved procedure, summarise public consultation material or assemble a service report. These are proposed applications of those task areas, not verified council deployments. I'd consider Adapt only for a tested, recurring preparation task. The public-sector section below applies.

Executives

Typical starting mode: Augment (does several steps, you check), with selective Adapt. Assemble performance evidence, draft stakeholder updates and rehearse questions before a decision meeting. A recurring pack can become a bounded preparation task once sources and checks are settled.

What stays yours: choosing the organisation's position.

Researchers

Typical starting mode: Assist or Augment. Search literature, collect evidence, analyse data and synthesise a cited report. Keep a record of search limits and check primary sources before relying on a conclusion. Collecting material is one job; judging it is another.

What stays yours: deciding whether the evidence supports the argument.

Public sector: state and federal agencies, councils and policy teams

I'd begin a public-sector workflow with its authority and records, then choose the tool. A useful draft does not establish that a process is approved for government use. The examples below distinguish official evidence from my suggested mode mapping.

The DTA's Policy for the responsible use of AI in government, checked on 17 October 2026, is version 2.0, effective 15 December 2025. Its page was last updated on 1 December 2025 and says v1.1 took effect on 1 September 2024. It applies to non-corporate Commonwealth entities, with exceptions. Its mandatory requirements include accountable officials, transparency statements, use-case accountability, registers, training and impact assessment.

My mapping: Assist can help staff understand approved guidance. Augment can prepare evidence and draft material for the accountable person's review. Adapt needs explicit operating boundaries and escalation. This maps chiefly to What stays your call, supported by What it can connect to and How much it does on its own. These mode labels are my interpretation, not DTA policy categories.

The DTA's Microsoft 365 Copilot evaluation distinguishes licence distribution in January to June 2024 from evaluation engagement in January to July 2024. It describes a non-randomised trial and warns that self-assessed productivity benefits may overstate or understate actual effects.

The official findings, checked on 17 October 2026, identify summarisation and rewriting as the most used functions. Teams and Word met participants' needs; Excel functionality and Outlook access issues hampered use. Verification and editing reduced some efficiency benefits. Tailored training improved confidence.

My mapping: those summarisation and drafting uses principally illustrate Assist and Augment. They do not establish readiness for Adapt or evaluate today's Copilot Cowork. The relevant layers are The app you open, How it works with you and What stays your call. I'd test preparation and checking together, using representative records.

For state agencies and councils, confirm the applicable jurisdiction's requirements rather than treating Commonwealth policy as automatically binding. I have omitted named state and council deployments because none was officially verified for this guide. Nothing here replaces local policy or advice.

Ready to move your team from asking to doing? See how Craft and coaching help →

Starting safely

Choose one recurring task whose result you can check. Start read-only, meaning the tool can view approved inputs but cannot alter the originals or send anything. Use a limited source folder and save drafts separately. Name the reviewer before the first run.

Write the brief in ordinary language: “Prepare a draft weekly project update from these approved notes. Link each decision to its source. List missing information. Save the draft for my review. Ask before changing shared records or contacting anyone.” Test what happens when a source is missing or two documents disagree.

Check actual controls. Anthropic's Cowork page, checked on 17 October 2026, describes selected folders and tools, approval for deletion and configurable permissions for significant actions. Microsoft's Copilot Cowork page, checked the same day, says sensitive steps such as sending an email or updating a record pause for your approval. OpenAI's dot guide says you can pause a dot at any time, and that asking for a draft doesn't authorise sending it.

After each run, inspect the sources, the output and any actions taken. Record corrections to the procedure. Expand access only when the next task requires it, and keep a way to pause work. My test is whether a colleague could understand what happened and take over the review.

FAQ

What are Assist, Augment and Adapt?

Assist, Augment and Adapt are three levels of how much AI does for you. Assist helps you think while you decide. Augment carries several steps for your review. Adapt works on its own within your rules while you stay available. Choose the level for each task, rather than permanently labelling a person or tool.

How much should AI do for you?

AI should do as much as you can responsibly delegate and reliably check. The right boundary depends on the stakes, the effort needed to verify the result and whether the work needs your authority. I keep consequential decisions with the responsible person and expand delegation after a bounded task works consistently.

What's the difference between a chatbot and AI that does the work?

A chatbot mainly returns responses for you to use; AI that does the work can take actions towards an output. That might include reading approved files, comparing evidence and saving a draft. The practical distinction is the work completed between messages, together with the permissions and checks governing those actions.

How do I pick the right level for a task?

Use three questions: what are the stakes if this is wrong, how quickly can I check it, and does it need my name or authority? Start with less delegation when consequences are serious or checking is difficult. Allow more steps when inputs, output and review are clear, while retaining consequential approvals.

What should always stay my call?

Your authority, commitments and acceptance of consequential work should stay your call. AI can prepare evidence and alternatives, but the authorised person must decide whether to rely on them. Keep permission changes, sensitive disclosures and external commitments explicit. Name the reviewer and give them enough context to reject an inadequate result.

What does AI-native work look like in a team, and in an organisation?

In a team, AI-native work means shared inputs, repeatable procedures and a named reviewer for completed steps. In an organisation, it also means clear access rules, accountable owners and ways to handle exceptions. People can still use AI for thinking while selected recurring tasks run within agreed boundaries and remain open to inspection.

How do I start safely?

Start with approved information, read-only access and a small task whose output you can check completely. Define the deliverable and require approval before changes or messages. Inspect sources and missing information after the run. Practise stopping the task, then widen its scope only when the review process is working reliably.

Keep going

Use Start here to orient yourself, then browse the tools with one task in mind. Describe the output, allowed sources and decisions you retain before comparing features. A useful next experiment is one you can finish and review.

If someone forwarded this to you, subscribe to The Helix Loop weekly email for dated tool changes and what they mean for your work. Each week the items are organised by the 7 layers, each tagged Assist, Augment or Adapt, so you can see whether it moves the line. Agent Watch supplies the tool information; Craft and coaching help you put it into practice.

Ready to move your team from asking to doing? See how Craft and coaching help →