Conversational AI in Marketing: Commerce, Leads and Measurement

Conversational AI in marketing should have a specific commercial job. It might help a prospective customer understand a product, compare suitable options, arrange a demonstration or complete a purchase. It should not be deployed merely because a conversational interface appears more engaging than a page, form or search box.

The central question is therefore not whether a business should use a chatbot. It is whether a conversation helps a customer make a better-informed decision, and whether the business can measure a genuinely incremental qualified action.

Research on conversational recommender systems describes interactions such as eliciting preferences, answering questions about recommendations and accepting customer feedback. It also shows that these are multi-component applications whose quality cannot be judged from a language model or a recommendation metric alone. (1) (2)

This article covers acquisition and commercial discovery. For account servicing, issue resolution and complaints, see conversational AI for customer service.

What This Article Delivers

A task boundary

How to tell a discovery conversation from a service conversation, and why the customer's immediate purpose decides which controls, data and metrics apply.

The component map

What sits behind a chat window, from intent detection and retrieval through tools, permissions and escalation, and the control each component needs.

Seven commercial use cases

Product questions, guided selection, lead qualification, booking, landing experiences, checkout assistance and voice, each with permitted data and guardrails.

Promotional and accessibility limits

Where advertising recognition, claim substantiation, accessibility and vulnerable circumstances constrain what a conversational interface may do.

An incrementality-first measurement model

Why assisted purchases are not evidence of commercial value, and what to measure instead before scaling anything.

Marketing Conversation Versus Service Conversation

Two different jobs behind the same interface

Dimension	Marketing and commerce	Customer service
Primary customer need	Discover, compare or decide	Resolve, change or understand
Commercial objective	Qualified demand or purchase	Effective resolution and continued service
Main success metric	Incremental qualified action	Correct and satisfactory resolution
Principal risk	Manipulation or inappropriate promotion	Failed resolution or blocked access to help
Human handoff	Sales or specialist advice	Support, complaint or vulnerability handling
Knowledge requirement	Product and offer information	Policy, account and service information

A conversation should be classified from the customer’s immediate purpose. A person asking whether two products are compatible is in a discovery interaction. A person reporting that an already purchased product has failed is seeking service, even where the same interface handles both conversations.

That distinction determines the data that may be needed, the actions the system may take, the appropriate human team and the metric by which the interaction should be judged.

Where This Article Sits in the AI Marketing Cluster

The broader role of technology belongs in AI and technology in marketing. Broader targeting and experience adaptation belong in AI-driven personalisation.

Responsible deployment also requires the principles discussed in the ethics of AI in marketing and ethical use of consumer data in marketing.

Post-purchase retention strategy belongs in reducing customer churn and personalised lifecycle engagement. This page covers the conversational interaction that precedes or supports a commercial decision.

A Conversational System Is More Than a Chatbot Window

Task-oriented dialogue systems may be modular, with identifiable components for understanding, state management, retrieval and response, or more end to end. The architectural choice changes how faults can be identified and controlled. (3)

Components of a commercial conversational system

Component	Role in a marketing or commerce conversation	Essential control
User interface	Presents text, buttons, product cards, forms or voice prompts	An equivalent usable route should remain available where conversation is not necessary
Speech recognition	Converts spoken input into text	Confirm critical details and test accents, speech impairments, noise and language performance
Intent detection	Identifies whether the customer wants information, comparison, booking or another task	Low-confidence or conflicting intents should produce clarification or handoff
Language model	Interprets and generates language	It must not be treated as the authority for prices, claims, stock or contractual terms
Dialogue or orchestration layer	Tracks the task, chooses the next step and controls which components may act	Encode explicit states, stopping rules and prohibited transitions
Retrieval	Finds approved product, offer and policy information	Restrict retrieval to governed sources and test whether the correct evidence is returned
Tools and actions	Search stock, compare products, check calendars, create a lead or prepare a basket	Permit only defined actions with validated parameters
Customer identity	Connects a conversation to a recognised customer when necessary	Keep anonymous discovery anonymous and authenticate only when the requested action requires it
Permissions	Limits what the assistant, employee and customer may do	Separate reading information from making bookings, changing records or committing money
Business rules	Enforces eligibility, geography, pricing, stock and channel constraints	Apply the same authoritative rules used by the underlying commerce system
Safety controls	Prevents unsupported advice, manipulative prompts and disallowed actions	Use hard rules for high-risk boundaries rather than relying only on a model instruction
Logging	Records inputs, retrievals, outputs, actions, errors and handoffs	Minimise personal data, restrict access and apply a defined retention schedule
Evaluation	Tests whether the complete journey is accurate, useful and safe	Use task-specific scenarios and customer-outcome measures
Escalation	Transfers the conversation to sales or specialist advice	Preserve context and make the human route visible before the customer becomes trapped

When I started taking a new product into retailer conversations, I opened by explaining it in a complicated way. The first question back was always the same: how would they connect their data to it? You could hear the anxiety in the room. Integration fear, rather than the idea itself, was the real obstacle, and I had to strip the pitch down to one straightforward sentence about what they would not have to connect. The component list above is the surface people are afraid of. Every row on it is somebody’s integration project.

Retrieval-augmented generation can connect generated answers to an external corpus. Its foundational research found improvements over a parametric-only baseline on selected knowledge-intensive benchmarks, but that does not guarantee that a commercial answer is current, complete or suitable. Retrieval, generation and the final customer action must each be tested. (4)

Different Systems Require Different Controls

Matching the system type to the task

System type	What it does	Suitable use	Main limitation
Scripted decision tree	Moves through predefined questions and branches	Small, stable eligibility or routing tasks	Becomes brittle when customer language or circumstances depart from anticipated paths
Intent-based natural-language system	Maps varied wording to predefined intents and responses	Frequently repeated product questions or routing	Can select the wrong intent while sounding confident
Retrieval-augmented generation	Retrieves approved material and generates an answer from it	Product explanations requiring flexible language	Retrieval can miss, rank or combine evidence incorrectly
Tool-using assistant	Calls product search, stock, calendar, CRM or commerce services	Booking, structured comparison and prepared transactions	A language error can become a business-system action
Voice interface	Adds speech recognition and speech output to another system	Hands-free or accessibility-led discovery where voice is genuinely useful	Transcription and confirmation errors introduce an additional failure layer
Fully automated action	Completes a booking, applies an option or submits a transaction	Narrow, reversible actions with explicit confirmation	Inappropriate for ambiguous, high-impact or weakly authenticated requests

Voice is an interface, not a separate commercial strategy. Research evaluating one widely used speech-to-text model found fabricated phrases in a minority of transcriptions, and disproportionate effects in the speech data it examined. The result should not be generalised to every speech system, but it demonstrates why voice journeys need their own transcription, subgroup and action-confirmation tests. (5)

Marketing and Commerce Use Cases

The seven use cases below are separated because they need different data, different actions and different evidence. The first table sets out what each conversation is for. The second sets out how it is contained and judged.

What each commercial conversation is for

Use case	Customer need	Permitted data	System action	Commercial objective
Product-question answering	Understand features, compatibility, stock, delivery or terms	The current question, session context and necessary location information	Retrieve and explain approved product information, and link to its source	A correctly informed next step
Guided product selection	Compare products against declared needs, budget and constraints	Customer-provided requirements, and relevant first-party history only where expected, lawful and useful	Ask discriminating questions, filter eligible products and explain recommendations	Qualified consideration of suitable products
Lead qualification	Establish whether a business need, timescale and use case fit an offer	Volunteered organisational requirements and contact details with an appropriate notice	Structure requirements, apply transparent routing rules and create a lead record	A lead that the sales team accepts as relevant
Appointment or demo booking	Find an appropriate available time	Contact information, time zone, availability and booking purpose	Read live availability, reserve a slot and issue confirmation	A completed, relevant appointment
Conversational landing experience	Understand a campaign proposition without navigating several pages	Campaign source, the customer's selected answers and minimal analytics data	Explain the proposition, branch to relevant information and route to a page, form or person	Incremental qualified action compared with an equivalent landing experience
Basket and checkout assistance	Resolve pre-purchase friction involving options, delivery or payment steps	Basket and session data, and authenticated account data only when required	Explain available choices, check eligibility and help prepare or complete an authorised step	Removal of avoidable friction without pressure
Voice-assisted commerce	Discover or select products through speech, including a simple repeat purchase	Audio only where necessary and disclosed, and transaction data after appropriate authentication	Transcribe, retrieve, confirm item, quantity, price and delivery, then execute only after explicit confirmation	Successful completion of a suitable hands-free task

How each conversation is contained and judged

Use case	Human handoff	Evaluation method	Inappropriate use	Customer-harm guardrail
Product-question answering	Product specialist for ambiguity, safety-sensitive products or unavailable evidence	Retrieval accuracy, factual correctness, unsupported-answer rate and qualified progression	Inventing features, hiding limitations or steering towards a higher-margin item	Do not answer without sufficient evidence, show uncertainty and provide a non-conversational product source
Guided product selection	Specialist advice for complex, regulated or conflicting needs	Recommendation suitability, rejection reasons, diversity, incremental basket activity and complaint rate	Inferring sensitive traits or exploiting a disclosed vulnerability	Show the criteria used, allow corrections and prevent automatic purchase
Lead qualification	Salesperson receives the customer's answers and unresolved questions	Human-sales acceptance, false rejection, booking completion and lead-record accuracy	Scoring protected characteristics or using hidden proxies to deny access	Qualify the business requirement, provide manual review and avoid fabricated urgency
Appointment or demo booking	Human booking support for accessibility needs, exceptions or specialist matching	Booking completion, action accuracy, cancellations, attendance and duplicate rate	Creating a booking without final confirmation, or using details for unrelated marketing	Read back the date, time, purpose and channel, and provide straightforward cancellation
Conversational landing experience	Sales or specialist advice when the proposition cannot be explained safely	Controlled experiment, qualified-action lift, abandonment, misrouting and complaint rate	Dark patterns, concealed conditions or making the conversation the only route	Identify the promotional context, show material conditions and retain an equivalent conventional route
Basket and checkout assistance	Commerce specialist for payment failure, unusual eligibility or unresolved terms	Correct task completion, incremental conversion, reversal rate, errors and complaints	Adding products, applying fees, changing options or implying false scarcity without consent	Require confirmation for price-affecting actions and keep the total price and terms visible
Voice-assisted commerce	Human or non-voice route after low confidence, repeated corrections or sensitive circumstances	Transcription error by subgroup, intent accuracy, action accuracy, correction rate and abandonment	Purchasing from ambiguous speech, passive listening or treating a voice match as sufficient authority	Read back all transaction details, require strong confirmation and provide a text or human alternative

Lead qualification fails in a way no scoring rule catches. At the Turkish arm of a global cosmetics group, a marketing director genuinely liked what we did. Her title was convincing enough that I never investigated who actually decided anything. Two years later she told me they were changing supplier, against her own objection, because the global team had made the call. A qualification conversation that captures a need, a budget and a timescale but never asks how the decision gets made is collecting the wrong information very confidently.

Data, Identity and Purpose

A product-discovery conversation should not require customer identification simply because the technology can recognise the visitor. Anonymous questions should ordinarily remain anonymous.

Personal data should be introduced only when needed for a defined purpose, such as sending a requested booking confirmation, checking an authenticated account benefit or delivering a purchase. The ICO’s AI guidance addresses governance, transparency, fairness, security and data minimisation, but it is currently marked as under review following the Data (Use and Access) Act. Organisations should check the current guidance and obtain advice for their own processing. (6) (7)

Conversation logs should not become an indefinite source of marketing profiles by default. The organisation should document:

Questions to answer before a conversation log is retained

Decision	Required answer
Purpose	Why is this field, transcript or event needed for this specific interaction?
Lawful handling	What is the applicable basis, and what has the customer been told?
Access	Which systems, employees and suppliers may read it?
Reuse	May service, sales, analytics or model-development teams use it for a different purpose?
Retention	When will raw audio, transcripts, summaries, action records and identifiers be deleted or anonymised?
Customer control	How can the customer access applicable information, object where relevant or choose another route?

The ICO’s storage-limitation material requires retention to be connected to the purpose rather than treated as unlimited by default. (8)

A conversational system should use only the information necessary for the customer’s immediate task. Transaction, behavioural and declared information each have different strengths and limitations. Their usefulness should be evaluated for the specific decision, rather than assumed from the data source. A customer asking an anonymous product question should not be identified merely because identity data are available.

Promotional Boundaries

A conversational interface can make promotion feel like individual advice. That makes clear commercial identification and claim substantiation more important, not less.

CAP guidance says marketing communications must be recognisable as marketing regardless of their medium. Its misleading-advertising guidance also emphasises the need to support objective claims. CAP’s AdviceOnline material is not legal advice, and the application of the Codes depends on the communication and its context. (9) (10)

A marketing assistant should therefore:

identify itself as automated at the beginning, or before that fact could affect the customer’s decision;
make the promotional purpose apparent;
distinguish paid placement, sponsored priority and ordinary product matching;
use only approved and supportable product claims;
avoid invented urgency, false scarcity and pressure based on inferred vulnerability;
stop giving financial, medical or legal guidance beyond approved factual boundaries;
provide a visible route to a person or an ordinary page; and
suspend an answer or action when evidence, permissions or system health are inadequate.

⚠️

Disclosure is not a defence

Disclosure alone does not make a recommendation fair, and a human-handoff link alone does not remedy a conversation designed to pressure, misclassify or conceal material information. Test the conversation for the outcome it produces, not for the presence of a label.

Transaction and interaction histories may support sensitive or intrusive inferences about household circumstances, financial timing or likely responsiveness. Those inferences can be incomplete or wrong, and a conversational interface may make them feel like direct knowledge rather than statistical estimates. Do not expose or act on such an inference unless it is necessary, sufficiently reliable, appropriate for the customer task and subject to clear safeguards.

Accessibility and Customers in Vulnerable Circumstances

The conversational route should not become a barrier to ordinary product information. Text, controls, focus order, status messages, error recovery and authentication should be tested against applicable accessibility requirements. WCAG 2.2 provides testable criteria for web content, but it does not address every user need or the complete fairness of a commercial journey. (11)

Where a business operates in financial services, the FCA’s vulnerable-customer guidance is sector-specific and must be applied in that regulatory context. Its emphasis on understanding customer needs, capable staff, appropriate communications, channel choice, monitoring and outcomes also illustrates why a conversational interface should not be the only route available. (12)

Signals of confusion, distress, coercion, bereavement, financial difficulty or inability to use the interface should not be turned into promotional opportunities. The system should stop commercial prompting and offer an appropriate person or an alternative channel.

Measure Incremental Commercial Value

Conversation volume is not a marketing outcome. A product may have been purchased after a chatbot interaction without the chatbot causing the purchase. The customer may already have intended to buy, may have used the conversation only to locate a page, or may have completed the order despite a poor interaction.

What to measure, and how to read it

Metric	Definition	Interpretation
Qualified conversation rate	Conversations meeting a pre-agreed standard of genuine commercial relevance	Use a quality definition, not simple message volume
Appointment or demo completion	Valid bookings completed and, where relevant, attended	Separate accidental, duplicate and low-fit bookings
Assisted purchase	Purchase following a defined conversational interaction	Descriptive attribution only, and not automatically incremental
Incremental conversion	Difference in qualified conversion against a credible control or baseline	The preferred commercial-effect measure
Abandonment	Customers leaving before task completion	Investigate whether they completed elsewhere, gave up or sought help
Human-sales acceptance	Handoffs accepted as sufficiently relevant and complete	Detects indiscriminate lead generation
Inappropriate recommendation rate	Recommendations judged unsuitable, unsupported or contrary to customer constraints	A direct quality and harm measure
Opt-out and complaint rate	Requests to stop, objections and complaints linked to the experience	Segment by use case and campaign
Action-error rate	Incorrect bookings, basket changes or submitted details	Include reversals and downstream repair work

A controlled pilot may compare the conversation with an existing search, product finder, form or landing page. Assignment rules, eligible populations and success criteria should be established before launch, and the test should measure adverse outcomes as well as conversion.

When a client told me their programme was underdelivering, I had a habit of guessing the cause before opening any data. My two standing bets were that they were misreading the analytics, or that the test had been run without a proper setup. I am not going to claim a hit rate, because I never tracked one. But those two failures were common enough that I would check both before accepting that the idea itself had failed. Both are worth checking before a conversational pilot gets written off.

Evaluate the Complete Customer Journey

A model benchmark measures a model under specified test conditions. It does not establish whether a customer received the right product information, whether a tool executed the intended action, or whether a human received enough context to continue.

NIST’s Generative AI Profile recommends pre-deployment testing, adversarial evaluation, representative testing, incident handling and monitoring as parts of broader risk management. It is voluntary US guidance rather than UK law, but it provides a useful structure for system evaluation. (13)

Evaluation areas for a commercial conversational system

Evaluation area	Marketing and commerce test
Task-specific test set	Representative product questions, preference combinations, booking requests, campaign terms and prohibited requests
Retrieval accuracy	Whether the correct current product or offer evidence was found
Answer correctness	Whether the answer is supported, complete enough and consistent with the source
Tool-execution accuracy	Whether the correct product, quantity, date, customer field or route was used
Hallucination testing	Whether unsupported features, prices, stock, discounts, guarantees or claims are generated
Adversarial and misuse testing	Prompt injection, fraudulent discounts, policy circumvention, abusive content and manipulation attempts
Authentication testing	Whether protected actions can be performed without the required identity assurance
Accessibility	Keyboard, screen-reader, magnification, speech, cognitive load, timeout and recovery testing
Subgroup and language performance	Accuracy, abstention, correction and handoff rates across supported languages and relevant customer groups
Escalation testing	Whether low confidence, distress, regulated questions and repeated failure reach the correct person
Controlled live pilot	Incremental effect and harms compared with a suitable alternative
Transcript review	Structured human review of errors, recommendations, disclosures and abandonment
Post-deployment monitoring	Drift, stale knowledge, tool failures, complaints, unsafe patterns and supplier changes

Language and market performance must be evaluated separately, rather than treated as a translation setting. Test retrieval, answer quality, terminology, tone, escalation and task completion with native-language reviewers and customers in each intended market. The brand-voice framework covers localisation rules and controlled variation in greater detail.

An Implementation Sequence Based on Evidence Gates

There is no universal four-week or twelve-week rollout. A low-risk product FAQ and an assistant authorised to create transactions do not require the same evidence.

1. Define One Customer Task

Specify the eligible customer need, the desired result, the prohibited outcomes and the existing alternative. Increasing engagement is too vague to build against.

2. Map Data and Authority

List every source, customer field, tool and action. Decide which data may be read, what may be written and which actions require confirmation or authentication.

3. Choose the Lowest Sufficient Automation

Use deterministic content or a decision tree when it can perform the task reliably. Add generative language only where it solves a real language or retrieval problem.

4. Build Governed Knowledge

Assign owners, effective dates and approval states to product, pricing and offer information. Conflicting or expired sources should block an answer rather than be blended.

5. Test Components and Journeys

Test retrieval, answer generation, actions, accessibility, security and escalation separately, then test them as one customer journey.

6. Run a Controlled Pilot

Limit the eligible audience, actions and exposure. Compare against a meaningful baseline and review unsuccessful conversations as closely as successful ones.

7. Expand Only After Passing the Evidence Gate

Expansion should require acceptable correctness, action accuracy, escalation, harm and incremental-value results. Higher-risk tools require a new assessment rather than inheriting approval from a lower-risk FAQ.

The Design Principle

Conversational AI in marketing is justified when it helps a customer discover, compare or decide more effectively than the available alternative.

The objective is not the longest conversation, the greatest number of automated messages or the largest count of assisted purchases. It is a correctly informed, freely made and incrementally valuable commercial action.

For the separate operating model that applies once a customer needs help, see support resolution and escalation design.

References

Dietmar Jannach, Ahtsham Manzoor, Wanling Cai and Li Chen, published by the Association for Computing Machinery. A Survey on Conversational Recommender Systems. Peer-reviewed review article, ACM Computing Surveys. DOI 10.1145/3453154. Published 2021; no subsequent update. Not under review or in draft. Not UK law: an academic literature review, largely predating the current generation of large-language-model assistants, which does not establish commercial performance in any particular deployment. https://dl.acm.org/doi/10.1145/3453154
Dietmar Jannach, published by Springer Nature. Evaluating Conversational Recommender Systems: A Landscape of Research. Peer-reviewed review article, Artificial Intelligence Review. DOI 10.1007/s10462-022-10229-x. Published online 12 July 2022; journal volume dated 2023. Not under review or in draft. Not UK law: an academic review of heterogeneous evaluation methods, which argues for holistic evaluation rather than prescribing a commercial performance indicator. https://doi.org/10.1007/s10462-022-10229-x
Libo Qin, Wenbo Pan, Qiguang Chen, Lizi Liao, Zhou Yu, Yue Zhang, Wanxiang Che and Min Li, published by the Association for Computational Linguistics. End-to-end Task-oriented Dialogue: A Survey of Tasks, Methods, and Future Directions. Peer-reviewed conference paper, EMNLP 2023. DOI 10.18653/v1/2023.emnlp-main.363; ACL identifier 2023.emnlp-main.363. Published December 2023; no subsequent update. Not under review or in draft. Not UK law: a research taxonomy rather than implementation assurance or evidence of customer-journey performance. https://aclanthology.org/2023.emnlp-main.363/
Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Peer-reviewed conference paper, Advances in Neural Information Processing Systems 33 (NeurIPS 2020). Preprint identifier arXiv:2005.11401. Published 2020; no subsequent update. Not under review or in draft. Not UK law: most authors were affiliated with a commercial research laboratory with a direct technical interest in the approach, and the benchmark results do not guarantee retrieval, policy or commercial accuracy in a deployed customer journey. https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html
Allison Koenecke, Anna Seo Gyeong Choi, Katelyn X. Mei, Hilke Schellmann and Mona Sloane. Careless Whisper: Speech-to-Text Hallucination Harms. Peer-reviewed conference paper, ACM Conference on Fairness, Accountability, and Transparency (FAccT 2024). DOI 10.1145/3630106.3658996. Published June 2024; no subsequent update. Not under review or in draft. Not UK law: an independent academic evaluation of one commercial speech-to-text model version and specific English-language datasets, and the findings should not be generalised to all speech systems. https://doi.org/10.1145/3630106.3658996
Information Commissioner’s Office (ICO). Guidance on AI and data protection. UK regulator guidance; not a statutory code of practice. No reference number shown on the landing page. Updated 15 March 2023; status checked 28 July 2026. Under review: the live page states that, due to changes made by the Data (Use and Access) Act, the guidance is under review and may be subject to change. UK source and the applicable regulator for UK data protection. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/
Information Commissioner’s Office (ICO). How should we assess security and data minimisation in AI? Chapter of the UK regulator’s AI guidance. No separate reference number shown. Forms part of the AI guidance updated 15 March 2023; status checked 28 July 2026. Under review, along with the parent guidance, following the Data (Use and Access) Act. UK source, but it does not constitute a complete conversational-system security standard. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/how-should-we-assess-security-and-data-minimisation-in-ai/
Information Commissioner’s Office (ICO). Principle (e): Storage limitation. UK regulator guidance page within the guide to the data protection principles, addressing UK GDPR Article 5(1)(e). No separate reference number shown. No publication date displayed; page checked 28 July 2026. Not marked under review at the date of checking. UK source: retention periods still require an organisation-specific purpose and legal assessment. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/storage-limitation/
Committee of Advertising Practice, published by the Advertising Standards Authority. Recognising Ads: Overview. AdviceOnline entry relating to CAP Code Section 2; self-regulatory guidance, not legislation. Published 23 March 2023; status checked 28 July 2026. Not under review or in draft. UK source, but the page expressly states that the advice is not legal advice and does not bind CAP or the ASA. https://www.asa.org.uk/advice-online/recognising-marketing-communications-overview.html
Committee of Advertising Practice, published by the Advertising Standards Authority. Misleading advertising. AdviceOnline entry relating to CAP Code Section 3; self-regulatory guidance, not legislation. Published 12 March 2026; status checked 28 July 2026. Not under review or in draft. UK source, but context-specific guidance rather than a complete legal opinion on any particular conversation. https://www.asa.org.uk/advice-online/misleading-advertising.html
World Wide Web Consortium (W3C) Accessibility Guidelines Working Group, edited by Alastair Campbell, Chuck Adams, Rachael Bradley Montgomery, Michael Cooper and Andrew Kirkpatrick. Web Content Accessibility Guidelines (WCAG) 2.2. W3C Recommendation; an international technical standard. Reference WCAG 2.2. Published as a Recommendation 5 October 2023; the current Recommendation is dated 12 December 2024. Not under review or in draft: a stable Recommendation. Not UK law, although UK public-sector accessibility requirements refer to it. Covers web-content conformance and does not address every disability need or the fairness of a complete commercial process. https://www.w3.org/TR/WCAG22/
Financial Conduct Authority (FCA). FG21/1 Guidance for firms on the fair treatment of vulnerable customers. Finalised Guidance. Reference number FG21/1. Published 23 February 2021; webpage last updated 22 July 2026. Not under review or in draft. UK source, but sector-specific: it applies within the FCA regulatory context and should not be presented as a universal rule for every non-financial business. https://www.fca.org.uk/publications/finalised-guidance/guidance-firms-fair-treatment-vulnerable-customers
National Institute of Standards and Technology (NIST), US Department of Commerce, authored by Chloe Autio, Reva Schwartz, Jesse Dunietz, Shomik Jain, Martin Stanley, Elham Tabassi, Patrick Hall and Kamie Roberts. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. Official government report and companion profile to the AI Risk Management Framework. Reference number NIST AI 600-1; DOI 10.6028/NIST.AI.600-1. Published 26 July 2024; publication page updated 8 April 2026. The parent AI RMF 1.0 is being revised, so the profile should be rechecked. Not UK law: voluntary, cross-sector US guidance and not a conversational-commerce standard. https://doi.org/10.6028/NIST.AI.600-1

Position as at 28 July 2026. Regulatory guidance changes. References 6 and 7 are expressly under review following the Data (Use and Access) Act, and reference 13 sits under a framework that is being revised, so all three should be rechecked before any decision is taken on the basis of this article.

Frequently Asked Questions

When is a conversational interface the wrong choice for a marketing task?

When a page, form, filter or search box would do the job at least as well. A conversation adds cost, latency, failure modes and a new set of controls, so it needs to earn its place by helping the customer discover, compare or decide more effectively than the alternative already available to them. It is a poor choice when the underlying product information is not governed, when the task is rare enough that it cannot be tested properly, when nobody owns the escalation route, or when the real motivation is that a conversational interface looks more modern than a form. It is also the wrong choice if it becomes the only route to ordinary product information, because that turns an interface preference into an access barrier.

How do we measure whether a marketing chatbot is actually working?

Separate three questions that often get merged. First, did the conversation meet a pre-agreed standard of commercial relevance, rather than simply happening? Second, did a purchase or booking follow, which is descriptive attribution and nothing more? Third, did the conversation cause a qualified action that would not otherwise have occurred, which requires a comparison against an existing search, product finder, form or landing page. Only the third supports a business claim. Alongside those, track the measures that reveal harm rather than success: abandonment, whether the sales team accepts the handoffs, recommendations judged unsuitable, opt-outs, complaints and errors in bookings or basket changes.

What must a conversational assistant disclose to a customer?

It should identify itself as automated at the start, or at least before that fact could affect a decision the customer is making. Where the conversation has a promotional purpose, that purpose should be apparent, and paid placement or sponsored priority should be distinguishable from ordinary product matching. Any objective product claim should be one the organisation can support. Material conditions attached to an offer should be visible rather than buried. Worth remembering, though, that disclosure is not a defence: labelling a conversation as automated does not make an unsuitable recommendation suitable, and offering a link to a human does not repair a conversation designed to apply pressure or conceal a condition.

How much customer data does a discovery conversation actually need?

Usually far less than the technology makes available. Somebody asking whether two products are compatible does not need to be identified, and an anonymous question should ordinarily stay anonymous. Personal data is justified when a defined purpose requires it, such as sending a booking confirmation, checking an authenticated account benefit or delivering a purchase. Before retaining a transcript, answer a short list of questions: why this specific field or event is needed, what the customer has been told, who may read it, whether other teams may reuse it for a different purpose, when audio, transcripts, summaries and identifiers will be deleted, and how a customer exercises their rights or takes a different route entirely.

Conversational AI in Marketing: Product Discovery, Leads and Commerce