# Experience and channels
What customer channels must provide for agents: a separate customer edge for disclosure, hand-over and evidence, on shared platform capabilities, measured by problems solved rather than calls avoided.
Author: Murali Sid (https://linkedin.com/in/muralisid)
Source: https://www.agenticarchitectureskills.com/layers/r09-experience-and-channels
Published: 2026-08-20
Updated: 2026-10-15
Licence: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/
Cite as: Murali Sid. Experience and channels. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/layers/r09-experience-and-channels
---
The experience and channels layer (R09) is where customers and staff meet the organisation. It covers phone, chat, apps, portals, text messages, letters and the screens operators use. It is one of the [fourteen layers](/layers) you own. Today it holds contact centre platforms, phone menus, portals, outage maps and a first wave of chatbots. When an agent talks to a customer, what it says is the company speaking, in law and in public. So this layer must provide a separate customer-facing edge. The edge carries the channels, the AI disclosure, the hand-over to a person and the record of what was said. Everything behind the edge is shared with every other agent. That includes knowledge, identity, tools, testing, monitoring and model access. These are shared platform capabilities, part of the [seven capabilities](/architecture/system-view) an agent platform adds. Measure whether the customer's problem was solved, not whether the conversation avoided a person.
## What sits in this layer today [#what-sits-in-this-layer-today]
Here is what a typical estate already runs in this layer. Vendors in brackets are examples, not a market claim.
| System category | Electricity distributor | Water utility | Hospital network | Generator, oil and gas, or miner |
| --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- |
| Contact centre, phone and video (Genesys, Amazon Connect, NICE) | Fault line and life support line. The retailer runs the bill, app and chat channels | Contact centre | Booking centre and contact centre. Telehealth (Healthdirect Video Call, Zoom, Teams). Interpreter services | Not named in the research |
| Portals and apps | Connections portal and outage map | Customer portal and app to pay, move home and give a meter read | Patient portal only where the record system has one (Epic MyChart). GP portals (The Viewer, HealtheNet, Orion Health Concerto) | AEMO Markets Portal and trading screens at generators |
| Notices and reminders | SMS outage alerts | SMS and web notices for outages and water restrictions | SMS reminders. Printed letters through a mail house | Crib-room screens at mine sites |
| Operator and field screens | Field mobility apps (SAP FSM, Salesforce Field Service, IFS) | Same | Not named in the research | Control system consoles. Historian clients (PI Vision). Mobile work apps. Tablets rated for hazardous areas |
| Chat and voice agents in use | Retail side: AGL's assistant contains 36 percent of about 125,000 monthly chats \[vendor]. Origin's voice agent serves over 100,000 customers (February 2026) | Customer chatbots | Healthdirect virtual triage, about 1.55 million encounters from April 2023 to March 2025 | Not named in the research |
| Vendor agents announced for these channels | Retail side: Kraken agents with Sierra, live at one unnamed utility (June 2026) \[vendor]. SAP's self-service agent, live since Q4 2025, no customers named \[vendor] | Hansen's AI Virtual Agent, no Australian or NZ deployment found \[vendor]. Salesforce utility subagents, which the vendor says may change \[vendor] | Epic's patient agent Emmie, first live at Sutter Health in March 2026. Not confirmed for Australia \[vendor] | Honeywell Experion Operations Assistant advises control room operators (March 2026) \[vendor] |
Standards: WCAG 2.2 for accessibility in every channel, and PCI DSS for card payments. Energy adds the Consumer Data Right in Australia. Hospitals add My Health Record conformance and recorded contact preferences. Plant screens follow ISA-101, the hazardous-area device rules (IECEx, ATEX) and EEMUA 191 for alarms. The full estates are on the [utilities](/use-cases/utilities), [healthcare](/use-cases/healthcare) and [energy](/use-cases/energy) pages.
## What agents add [#what-agents-add]
### What the agent says is the company speaking [#what-the-agent-says-is-the-company-speaking]
In Moffatt v Air Canada (2024), a Canadian tribunal rejected the claim that a chatbot is a separate legal entity. It found negligent misrepresentation. A German appeal court (OLG Hamm, 12 May 2026) held that a chatbot is part of corporate communication. Its statements count as the operator's, like an employee's, and general disclaimers do not protect the operator. So failures in front of customers can be pinned on the company, can bind it, and are public. Failures inside the company stay contained and can be put right. That legal exposure is the reason for a separate edge.
### Build only the edge separately [#build-only-the-edge-separately]
Five things belong to the customer edge, because nothing inside the company does the same job.
* Channels and telephony, including callers who interrupt the agent mid-sentence.
* Measures of containment and escalation, wired to workforce management.
* AI disclosure and consent.
* Guardrails hardened against anonymous attackers. A car dealership's chatbot was talked into agreeing to sell a vehicle for $1.
* A record of what was said, to whom and when, kept as legal evidence.
Everything else is shared with the rest of the organisation's agents. That means the knowledge base, customer identity data, tools, tests, monitoring and model access. Running a second copy of any of them is a defect. Each shared part has its home. Knowledge sits in the [data platform layer (R02)](/layers/r02-data-platform) and the [agent data engineering layer (R14)](/layers/r14-agent-data-engineering). Identity sits in the [security and identity layer (R10)](/layers/r10-security-and-identity). Tools pass through the [integration fabric layer (R03)](/layers/r03-integration-fabric). Tests sit in the [intelligence and learning layer (R06)](/layers/r06-intelligence-and-learning). Monitoring sits in the [observability and FinOps layer (R12)](/layers/r12-observability-and-finops). Model access sits in the [agent platform layer (R07)](/layers/r07-agent-platform).
### Tell customers it is an AI [#tell-customers-it-is-an-ai]
EU AI Act Article 50(1) became enforceable on 2 August 2026. People must be told they are dealing with an AI system, at the latest at first contact. The exception for when it is obvious is read narrowly (Commission guidelines, 20 July 2026). Penalties reach EUR 15 million or 3 percent of worldwide turnover. The duty applies outside the EU wherever the outputs are used in the EU. The US is a patchwork of state laws, with different triggers in California, Utah and Colorado. The [governance, risk and sovereignty layer (R11)](/layers/r11-governance-risk-sovereignty) tracks these dates. In the EU, the design question is now how to disclose, not whether.
Cursor's support bot (April 2025) invented a login policy to explain logouts caused by a timing bug. Its replies were not labelled as AI, and developers cancelled within hours. The fix labelled every AI reply in email support.
### Measure problems solved [#measure-problems-solved]
A conversation that ends without a person counts as contained, whether the customer was helped or gave up. That is false containment. So measure resolution and repeat contact as a pair. Price the cost per resolved issue, not the cost per contact. Klarna reversed course in May 2025 after automating 67 percent of chats \[vendor]. Commonwealth Bank cut 45 roles over a voice bot, then called it an error and offered them back. Both had set a containment or headcount target, and resolution did not follow. A real person must be reachable, and the customer should not have to start again. Hand-over with the conversation's state is part of launch, not a later addition.
### Read vendor claims against vendor tests [#read-vendor-claims-against-vendor-tests]
Salesforce's own research benchmark, CRMArena-Pro (May 2025), reads against its own marketing. The best model scored 58.3 percent on single-turn tasks and 30 to 35 percent on multi-turn ones. The same vendor reports 84 percent resolution in production \[vendor]. The task mixes and definitions differ, so this guide reports both. Test on multi-turn conversations, because single-turn tests overstate capability by about half.
### Voice is its own problem [#voice-is-its-own-problem]
A voice agent lives or dies on delay and on handling interruptions. A chain of speech-to-text, a language model and text-to-speech adds 1 to 3 seconds per turn. Speech-to-speech models cut that delay. Practitioners treat under half a second as natural and over a second as broken. These are thresholds in circulation, not measured results. McDonald's ended its IBM drive-thru partnership in July 2024 over acoustics, not language.
### Screens a model builds [#screens-a-model-builds]
Some agents reply with screens, cards and forms that a model builds at run time. That is a display choice, not a new kind of agent. The [four questions](/architecture/agent-types) about the agent still apply. Four rules apply to any such screen. A fixed check confirms its structure before the customer sees it. It shows the source of each fact. Its buttons come only from a fixed list of permitted actions, enforced outside the model. Personalising it by behaviour or emotion needs the customer's consent. No outside measurement of these screens was found.
### Customers' own agents are arriving [#customers-own-agents-are-arriving]
Customers' shopping agents now read company pages and buy through new protocols. The Agentic Commerce Protocol covers checkout, and the Universal Commerce Protocol covers the merchant side. The Agent Payments Protocol carries signed spending permissions, called mandates. Product pages are the least readable to these agents, at 66 percent in one scorecard \[vendor]. So make product data readable by machines.
### A worked example: a utility customer account [#a-worked-example-a-utility-customer-account]
Take a water utility, or an energy retailer, serving customers by chat and phone. The agent says it is an AI and offers a person. It checks identity with fixed rules, and asks for a one-time code before any change. For an energy retailer, it drafts a home move and sends the market steps through the existing gateway. When a customer mentions life support, a fixed rule registers it at once. For a payment plan, the agent gathers facts and shows options within policy. A trained person approves the plan. When it hears signs of hardship, the agent hands over warmly, with a summary. The hardship officer decides. Payments go through a compliant pay link, and the agent never sees a card number.
Hardship, payment plans and disconnection fall under [the rule for decisions about people](/architecture/deterministic-zones#the-rule-for-decisions-about-people). A named person decides them. Credits, refunds and account changes that move money sit in the money zone of the [deterministic zones](/architecture/deterministic-zones#movement-of-money). There a fixed rule or a named person decides. In Australia, the energy rule maker (AEMC) made a final rule on 10 September 2026. It strengthens payment difficulty and disconnection protections from 30 March 2027. A second rule, on payment plan standards, was still open on 15 October 2026.
## What must be true before an agent touches it [#what-must-be-true-before-an-agent-touches-it]
Before an agent faces a customer, these must hold.
* **It says it is an AI.** It tells people at first contact, in every channel, in a clear and accessible way.
* **It knows who it is talking to.** The customer is verified, with a step-up check before any change. The agent acts only within that customer's own entitlements.
* **It answers from approved sources.** It grounds answers in a governed knowledge base. It refuses rather than improvises on policy.
* **A person is reachable.** A staffed queue exists, and the hand-over carries the conversation's state. Containment is never made by closing the human queue.
* **Money and people decisions stay fixed.** Changes that move money pass the money zone. Decisions about a person go to a named person.
* **Card numbers never reach it.** Payments go through a PCI DSS path.
* **The record sits outside the agent.** Transcripts and traces are stored where the agent cannot change them.
* **It is tested like a public statement.** Adversarial tests run before launch and after each change. Its [risk profile](/architecture/risk-profile) is scored before release.
The enforcement point is the tool gateway and the policy decision point, which the edge shares with every other agent. Disclosure is fixed in each channel's configuration, not left to the model. Both follow [enforcement outside the model](/architecture/enforcement).
## Choices and their verdicts [#choices-and-their-verdicts]
Each choice is the common choice, tested against the simplest credible alternative. The full records are in the [decision catalog](/decisions).
* **[Containment or resolution as the target](/decisions#cd-20) (CD-20):** target resolution, and let containment be an outcome. Every documented reversal had set a containment or headcount target.
* **[A separate customer stack or a shared platform](/decisions#cd-24) (CD-24):** a separate edge on shared platform capabilities. The reason is legal exposure, not technology.
## Concerns this layer must honour [#concerns-this-layer-must-honour]
The ten cross-cutting concerns are set out in the [concern matrix](/architecture/concern-matrix). This layer carries a specific duty for eight of them.
* **Identity and access:** customer identity and entitlement checks come before any account action.
* **Observability:** resolution, repeat contact and escalation by trigger, with containment tracked apart.
* **Traceability and audit:** full conversation records are kept as legal evidence.
* **Grounding:** answers come from a governed knowledge base, and the agent refuses rather than improvises on policy.
* **Impersonation:** AI disclosure is mandatory, and no agent presents as a named person.
* **Safety and oversight:** a real queue for escalation, gated actions that cannot be undone, and paths for vulnerable customers.
* **Cost:** cost per resolved issue, not per contact.
* **Resilience:** a defined fallback when no person is available.
## Evidence and limits [#evidence-and-limits]
The legal case for a separate edge rests on court rulings from February 2024 and May 2026. No published case measures a unified customer and internal stack at scale. None credits success to deliberate separation either. So this guide recommends the separate edge on legal grounds, not measured outcomes.
Rollback is normal. In a vendor-commissioned survey of 2,527 leaders, 74 percent had rolled back an AI customer agent. Each rollback followed a governance failure (Sinch, January to February 2026).
Disclosure has a measured cost and a measured benefit. In a 2019 field experiment, up-front bot disclosure cut purchase rates by more than 79.7 percent. Customers who knew they were dealing with AI reported satisfaction 34 points higher (COPC, January 2026).
Many figures here come from vendors and carry the \[vendor] tag. Gartner figures reached this guide second-hand, so confirm the wording before citing them. No credible independent benchmark for escalation rates exists as of mid-2026. The voice delay bands are practitioner thresholds, not measured results.
## Next step [#next-step]
See this layer at work in the proposed [billing exceptions](/use-cases/billing-exceptions) example. Then read the next layer, the [security and identity layer (R10)](/layers/r10-security-and-identity), which gives every agent its own identity.
Explore this topic
* [Experience and channels findings](/library/layers/r09-experience-and-channels/findings) and [sources](/library/layers/r09-experience-and-channels/sources).
* [Products named for orientation](/architecture), on the one-page wall chart.
* Estate table: industry research for utilities, healthcare and energy, 15 October 2026. Rows are industry knowledge unless a source is named.
* Table sources: AGL on IBM watsonx (IBM case study, undated) \[vendor]. Origin half-year deck (12 February 2026, search excerpt). Kraken and Sierra (11 to 12 June 2026) \[vendor]. SAP self-service agent (ASUG, 21 January 2026) \[vendor].
* Table sources: Hansen AI Virtual Agent (18 February 2025; FY26 results, 19 August 2026) \[vendor]. Salesforce Agentforce for Energy and Utilities (product pages, 2026) \[vendor]. Healthdirect triage (Medical Republic, 2025).
* Table sources: Epic Emmie (HIT Consultant, 10 March 2026) \[vendor]. Honeywell Experion Operations Assistant (19 March 2026) \[vendor].
* Worked example: utilities research, customer account journeys, 15 October 2026. AEMC final rule on payment difficulty protections, 10 September 2026 (search excerpt). AER v Origin (Federal Court, 6 July 2022) set the bar for capacity-to-pay conversations.
* Court rulings: Moffatt v Air Canada, 2024 BCCRT 149, February 2024. OLG Hamm, 12 May 2026. LG Muenchen I, 28 May 2026, extended operator liability to AI Overviews about third parties.
* Market structure: Gartner tracks four customer experience markets. They are CCaaS (September 2025), Conversational AI (July 2026), CRM Customer Engagement Center, and knowledge management (5 August 2026).
* Microsoft and ServiceNow appear in none of the 14 Conversational AI or 9 CCaaS positions. Microsoft's Service Agent reached general availability in M365 Copilot on 15 July 2026 \[vendor].
* Disclosure law: Article 50(2) on marking synthetic content has a grace period to 2 December 2026. California SB 1001 (2019) is triggered by commerce. Utah's AIPA (2024) by a request and by sector.
* Colorado HB 26-1263, the Chatbot Safety Act, was signed 29 May 2026 and takes effect 1 January 2027.
* Disclosure experiment: Luo, Tong, Fang and Qu, Marketing Science 38(6), 2019, more than 6,200 customers. Undisclosed bots matched proficient human agents.
* Measurement: COPC, January 2026, 1,000+ consumers in six countries. 74 percent satisfied with their last AI interaction, above 90 percent when resolved. Net Promoter Score fell by up to 70 points on failure.
* COPC names three causes of failure: measuring deflection, optimising for sounding human, and weak data foundations.
* Escalation: Zendesk CX Trends 2026, 11,000+ respondents in 22 countries \[vendor]. 52 percent of Chinese respondents lost context at hand-over. 74 percent find retelling their story frustrating.
* Preference: Metrigy, 503 respondents, November 2025. 84.9 percent prefer human agents, and 80.1 percent still do when promised equal resolution.
* Pricing: $0.99 to $2.00 per automated resolution against a $6 to $12 human comparator \[vendor].
* Reversals: Klarna automated 67 percent of chats, 2.3 million in 30 days, the work of 700 staff \[vendor]. Commonwealth Bank called the cut an "error" on 21 August 2025.
* Gartner forecasts, reached second-hand: 80 percent autonomous resolution of common issues by 2029. Half of organisations that planned service staff cuts will drop the plans by 2027.
* Assist-side evidence: Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), May 2025. Across 5,172 support agents, 15 percent more issues were resolved per hour.
* Vendor test: CRMArena-Pro, arXiv 2505.18878, 4,280 tasks. Knowledge answers scored about 23 to 36 percent. Confidentiality refusal was 0 to 0.4 percent. Hardened prompts raised it to 24 to 62 percent, at a cost of 1 to 11 points of task success.
* Vendor production claim, September 2025: 84 percent resolution across 380,000 conversations \[vendor]. Support headcount fell from 9,000 to about 5,000, and costs fell 17 percent \[vendor].
* Voice: Taco Bell reported over 2 million orders across 500+ locations while slowing its rollout \[vendor]. In a Deepgram and Opus Research survey of 400 leaders, 21 percent were very satisfied \[vendor].
* Commerce: ACP (OpenAI and Stripe, September 2025, specification 2026-04-17, beta). UCP (Google and Shopify, 11 January 2026). AP2 for payment mandates.
* Commerce: Adobe Q1 2026 scorecard on page readability \[vendor]. AI-referred traffic converted 42 percent better in March 2026 \[vendor]. Salesforce credited AI with $67 billion of Cyber Week 2025 orders \[vendor].
* Access: the Ninth Circuit overturned Amazon's injunction against Perplexity's Comet browser. Whether agentic access can be refused remains unsettled.
* Rollback survey: Sinch, January to February 2026, vendor-commissioned, fielded by an independent institute. The share rose to 81 percent among organisations with mature guardrails.
* Excluded: a widely recycled pairing of two vendors' containment figures, not found in the article it cites. Also excluded: voice-agent use among large banks, voice growth rates and an average containment rate.
* Also excluded: every escalation-rate benchmark and a pilots-never-reach-production rate.
* Re-verify quarterly: Article 50 enforcement and first penalties, the Colorado start date, and ACP and UCP status. Also check agentic-access cases and any credible escalation benchmark.