Suggest versus act
Every guideline about dismissing, correcting and ignoring a suggestion assumes the suggestion has not already happened — which is the assumption an agent with tools breaks first.
A designer from one of the portfolio companies is walking through an agent that files expense reports. Someone asks what happens when it gets one wrong. The answer comes back fluent and cited: we follow the HAX guidelines — support efficient dismissal, support efficient correction. Heads nod. The meeting moves on.
Nothing said in that exchange was inaccurate. Both guidelines exist, both are numbered, and the designer quoted them in the right order. The problem is that the agent had already filed the report.
This lesson is about that gap, and about stating it in a way that survives someone opening the paper on their laptop while you talk.
The five guidelines a room will quote at you
Table 1 of the CHI 2019 paper groups five guidelines under the heading When wrong. Here they are with their qualifiers, verbatim, because the qualifier is where the assumption hides:
| # | Guideline | Qualifier, verbatim |
|---|---|---|
| G7 | Support efficient invocation | “Make it easy to invoke or request the AI system’s services when needed.” |
| G8 | Support efficient dismissal | “Make it easy to dismiss or ignore undesired AI system services.” |
| G9 | Support efficient correction | “Make it easy to edit, refine, or recover when the AI system is wrong.” |
| G10 | Scope services when in doubt | “Engage in disambiguation or gracefully degrade the AI system’s services when uncertain about a user’s goals.” |
| G11 | Make clear why the system did what it did | “Enable the user to access an explanation of why the AI behaved as it did.” |
Read the verbs. Invoke. Dismiss. Ignore. Edit. Refine. Recover. Disambiguate. Degrade. Explain. Not one of them is undo, reverse, compensate, roll back, or notify the other party. Every verb in that list describes something the user does to an offer that is sitting in front of them, still inert.
G9 is the closest miss, and it is worth being precise about why it misses, because “recover” sounds like it covers you. Table 1’s right-hand column gives each guideline a worked example, and G9’s are a voice assistant’s editable reminder and an autocomplete suggestion. Both are pre-execution edits. The paper is describing a user fixing a draft, not a user living with a consequence.
Do not oversimplify the phase label
A tempting shortcut is to say “the When wrong phase is about wrong answers.” It is not, and someone will call it. G7 and G8 — invocation and dismissal — are not about wrongness at all. Table 1’s own caption hedges the whole scheme as “roughly categorized by when they likely are to be applied during interaction with users”. The phase names are Microsoft’s rough buckets, not claims that every guideline in a bucket addresses the same thing. Treating the bucket as homogeneous is exactly the kind of mischaracterisation that costs you the room.
The one bullet in the entire canon about taking an action
Apple’s Generative AI page in the Human Interface Guidelines carries the closest thing anyone has published, under Best practices:
Consider consequences and get permission before performing irreversible or potentially problematic tasks. Avoid automating destructive actions, like deleting photos, and actions that are hard to undo, like making a purchase on a person’s behalf. Generally, ask for confirmation before performing a significant action on someone’s behalf.
That is real guidance for a system that acts, and it is one bullet. Notice its tense. Before. Get permission before, ask for confirmation before, avoid automating in the first place. The whole bullet is a gate at the front of a task. It has nothing to say about the moment after the gate opens.
Apple is a platform vendor and this page sits inside the guidance adjacent to App Review, which gives it a kind of force the others do not have and also gives Apple an interest in what it says. That distinction gets a whole lesson later in this module. For now, note only that the page was new on 9 June 2025 and revised on 8 June 2026, and the bullet is unchanged.
Microsoft’s newest agent page keeps the suggestion frame
The obvious objection is that the canon has moved on since 2019, and Microsoft has published agent-specific guidance. It has. And its first principle keeps the user in the invoking seat. Human-centered design for agents on Microsoft Learn (ms.date 2026-05-20, updated 2026-07-14) opens on Built for intent, which asks designers to keep the user as the one initiating action, and illustrates it with a phrasing preference: “Summarize with Copilot” rather than “Copilot, summarize”.
Read that as a design position and it is a per-action, user-initiated invocation model. In 2026, on a page with the word agents in its URL, Microsoft is still describing a thing you ask for one action at a time. Two caveats before you use it: Microsoft sells the agent platform this page documents, and the page carries ai-usage: ai-generated in its own metadata. Cite it as a published Microsoft position, not as a research artifact.
Saying the absence so that someone can check it
An absence is the weakest kind of claim to assert and one of the strongest to demonstrate. The difference is whether you can say what you looked at and what would have counted as finding something. Here is that for this lesson, and you should be able to repeat it:
- Table 1 of the CHI paper, read directly from the PDF, guideline text and example-application column both. A hit would have been any G7–G11 example describing an already-executed, side-effecting action being corrected. There is none: the examples are reminders, autocompletes and search results.
- Apple’s Generative AI page, fetched through Apple’s own DocC JSON rather than the client-rendered HTML, and confirmed unchanged since the 8 June 2026 revision. A hit would have been any sentence about a task already underway. The permission bullet above is the whole of it.
- Microsoft’s agent design page, full page text. A hit would have been any undo, rollback or compensation language. There is none anywhere on the page.
That is the shape of a defensible absence claim. Three named documents, one stated criterion for what would have falsified it, and a result you can reproduce in an afternoon.
The line this course draws, which no source draws for you
Everything above is characterisation: what the documents say, what they do not, and how that was checked. The conclusion is this course’s own, and it should be spoken as an argument rather than as a finding.
The assumption an agent with tools breaks first is not accuracy. It is reversibility. Dismissal, correction and explanation are all cheap because the offer is inert while the user considers it. Take away the inertness and the five guidelines do not become wrong, they become unpriced. G8 still says make dismissal easy, and dismissal of an expense report already filed is a different engineering problem, a different screen, and a different conversation with the person who received it.
No source says that. It is an inference from reading G7–G11 and Apple’s bullet side by side, and the honest way to say it in a room is here is what I think follows, not here is what the guidelines say. The citation buys you the right to be listened to. The inference is yours to defend.
Retrieval check
Someone answers: "G9 says support efficient correction. Our agent has an undo button. So we are compliant." What is the accurate response, and what is the trap in giving it?
Check your answer
Something close to: G9 is real and an undo button is a good idea. But look at Table 1’s examples for G9 — an editable reminder and an autocomplete suggestion. Both are edits before anything happened. The guideline was written for a draft, so it does not tell you what undo means once the booking exists. That part we have to decide ourselves.
The trap is overclaiming. If you say “the guidelines do not cover correction” you are wrong and easily corrected. If you say “nobody has written about undo for agents” you are wrong too, and the Smashing article is one search away. The claim that holds is narrow and specific: the guideline exists, its evidence base is pre-execution, and the institutional canon leaves the post-execution case unpriced.
Hands on
Write the suggest-versus-act row into the position document
Done when: The Microsoft row in POSITIONS.md has a filled citation cell copied verbatim from the paper, a position cell that states the reversibility argument as yours rather than as the paper’s, and an evidence cell that names the limits of what you shipped.
- Open
learning/agentic-ux-canon/POSITIONS.mdand go to the Microsoft row. Fill the citation cell with G8 or G9, whichever you actually want to argue with, copying the number, the title and the qualifier from the paper rather than from this page. - Write the position cell in two sentences. First sentence concedes what the guideline gets right. Second sentence names the assumption — the offer stays inert while the user decides — and says what changes when it does not.
- Fill the evidence cell from the review gate, and state its narrowness in the cell itself: an app-level gate, request and response, no resume, no multi-step run that can stop halfway. A cell that names its own limits is stronger than one that does not, and this one is easy to overstate.
- Add a line under the row naming the Smashing article and its date, so that when someone raises it you are the person who already had it rather than the person hearing it for the first time.
- Say the whole row out loud as one sentence, under thirty words, and time yourself. If it needs a preamble to make sense, it is not the sentence yet.
What this does not cover
This lesson establishes that the canon assumes a suggestion which has not yet fired. It stops at the moment of firing. What happens when the agent fires three tools, commits two side effects and then stops is the halfway-failure lesson, which walks three separate published failure taxonomies and shows where that case would go in each. The over-hedging lesson takes G10 — the one guideline in this table this course actively disagrees with — and argues it in both directions with the trust research on each side. And the reading habit behind all of it, telling a platform rule from a research finding, is the platform-and-consultancy lesson at the end of this module.
Read this next — primary source
Guidelines for Human-AI Interaction — Table 1, guidelines G7–G11 and their example applications (CHI 2019, camera-ready)Amershi et al., Microsoft Research — fetched 2026-09-05. Microsoft publishes design guidance about the category of product it sells; free PDF, 13 pages, Table 1 on page 3.
Read Table 1 with your finger on the right-hand column, not the left. The guideline text is what everyone quotes; the example applications are what the guidelines were actually written against, and that column is where this lesson’s argument lives. G7 through G11 are the five that a room will quote at you the moment you describe an agent, and every one of their worked examples is a thing the user could still walk away from. Ten minutes on one table, and the absence you are about to claim becomes something you saw rather than something you were told.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.