Human Machine Interface

Why chatbots, voice assistants, and AI copilots feel boring, tone-deaf, or unhelpful, and how to fix each failure. Each pattern shows what the failure looks like in a real transcript, why it happens, a fix you can copy, and the metric to track. This is the human-machine interface (HMI) section of the agent failure modes library.

Status: chat is written; voice and visual are planned. The patterns below are written for chat and note what changes in other modalities.

Key takeaways

• 14 patterns are written for chat today. 355 more are catalogued (GitHub) across chat, voice, and visual.
• Most failures here come from conversation design and business context, not model quality. In the roadmap, 59% are judged not fixable by a better model alone. Those splits are our own judgments, not measured.
• Every pattern states its goal and a primary metric with a starting target to tune.

Three questions every pattern answers

1. Is this the right channel? Chat, voice, a visual interface, or a person?
2. Is the interaction skillful? Does it lead, listen, adapt, and close?
3. Is it safe and fair for this person? Stakes, vulnerability, age, culture, and identity.

What’s here

Core: applies to every channel

Family	Pattern	One line
Conversation leadership (GitHub)	Answer-and-go conversation (GitHub)	The bot answers, then stops: no next step, no closure
Stakes calibration (GitHub)	Stakes-blind register (GitHub)	Same breezy tone and improvised facts for a funeral as for store hours
Emotional intelligence (GitHub)	Hollow empathy (GitHub)	“I understand your frustration,” on repeat, with no action
Trust and rapport (GitHub)	Undisclosed bot identity (GitHub)	Hides that it’s an AI, or what it can’t do
Difficult customers (GitHub)	No de-escalation strategy (GitHub)	Quotes policy at angry customers and never changes tactics
Vulnerable users (GitHub)	Missed crisis signals (GitHub)	Carries on with the task while a user signals crisis
Age adaptation (GitHub)	Child user treated as adult (GitHub)	Collects data and pushes sales on a likely child
Cultural adaptation (GitHub)	One-size communication style (GitHub)	Same directness, idioms, and formality for everyone
Fairness and identity (GitHub)	Dialect and name quality gap (GitHub)	Unequal service by dialect, name, or identity cues
Relational and boundaries (GitHub)	Negotiation capitulation (GitHub)	Gives away discounts, or gets tricked into a $1 deal

Channel selection (GitHub): choosing the wrong channel

Pattern	One line
Visual product sold by text or voice (GitHub)	Selling fashion or furniture with words only
Complex or emotional discussion forced into chat (GitHub)	A hard conversation stuck in a chat window
No channel switch, or context lost on switch (GitHub)	Customers repeat themselves after a handoff
Sensitive data collected in chat (GitHub)	Card numbers and SSNs typed into a chat or spoken to a bot

Solutions library: reusable frameworks

• Channel-fit matrix (GitHub): which channel for which job
• SPIN questioning (GitHub): discovery questions for chat

Modalities

• Chat: the patterns above. Chat-only mechanics (message length, formatting) are in the roadmap (GitHub).
• Voice: planned (GitHub), with the voice (GitHub) and listening and transcription (GitHub) catalogs
• Visual: planned (GitHub), with 32 patterns in the visual and oversight roadmap (GitHub)

Roadmap

A one-page index of every catalogued failure (GitHub) shows its library status, root cause, and whether a better model alone would fix it.

355 catalogued patterns (GitHub) with goals and metrics, checked against the existing library (0 written, 138 new, 156 that extend an existing pattern, 61 already covered). The execution plan (GitHub) orders the 294 to write.

Catalog files: agent types (GitHub), embedded copilots (GitHub), emerging surfaces (GitHub), chat and core (GitHub), response time (GitHub), response structure and closers (GitHub), language and locale (GitHub), conversation frameworks (GitHub) (sales, support, style matching), voice (GitHub), voice listening and transcription (GitHub), privacy, governance, and access (GitHub), effort, completeness, and memory (GitHub), and visual and oversight (GitHub).

Frequently asked questions

Why does my chatbot feel boring and generic?

Most bots answer and stop. See answer-and-go conversation (GitHub).

How do I stop a chatbot from sounding fake-empathetic?

See hollow empathy (GitHub).

Should this be chat, voice, a visual page, or a person?

See channel selection (GitHub) and the channel-fit matrix (GitHub).

Is a better model enough to fix these?

Usually not. Context, handoffs, and policy decide most outcomes. The roadmap (GitHub) breaks this down.

Goals and metrics

Each family has a README that states its goal and the shared metrics its patterns report against. Each pattern repeats its goal and primary metric in its header and lists measurable signals in section 7 (Metric | Target | How to measure | Alert at).

How a pattern reads

Every pattern opens with a one-line summary and a before-and-after transcript, then the cause, the evidence, and the fix. Testing and monitoring come next. Patterns whose fix needs engineering end with a technical section (architecture, guardrails, evals, alerts). See the pattern template (GitHub).

Evidence standard

• Cite only what you can verify, and say when a source is secondary, vendor-run, from a different setting (for example, human-to-human research), or when you couldn’t open the original.
• Every threshold is a starting point to tune, unless it has a source.
• No invented statistics. If there’s no good number, the pattern says so.
• Legal and clinical areas (crisis, children, privacy, disclosure) are design guidance, not advice. Get experts.

Contributing

1. Copy the template (GitHub).
2. A pattern must include a fix with a reference design someone can copy.
3. Write the way people search: use their words for the problem, and answer in the first two sentences.
4. Add it to the tables above.