Vietnam Investment Review reports Khoi Le describing an ELSA Speak campaign that used a Meta Business Agent with click-to-Messenger ads. He described a 26% rise in qualified leads, an 18% drop in cost per qualified lead, and a 32% rise in average order value. Those figures are vendor-reported. They are not an independent incrementality test, and Integrated.Social has not reproduced them. The useful point is the architecture: the agent sat between the ad and the conversion, which is the job a landing page used to own alone.
A landing page is no longer always the destination. It can be a conversion layer: the agreed place where a conversation resolves into a booked call, a structured brief, a consented contact record or another accountable next step.
What was actually reported about ELSA Speak
A reported campaign result, not a universal rate card
Vendor-reported, not an independent incrementality test: Khoi Le described a 26% rise in qualified leads, an 18% drop in cost per qualified lead and a 32% rise in average order value for ELSA Speak. The report does not establish a Meta benchmark, a causal result for every advertiser or a result Integrated.Social has reproduced.
The reported percentage belongs to a specific product, audience, creative, market, offer and sales process. It does not prove that a conversational agent will outperform a form, raise revenue or lower acquisition cost for a different business. Any commercial interpretation should keep that boundary.
Meta describes its Business Agent as a customer-facing capability across Meta business messaging surfaces. The operational question is not whether a chatbot is present. It is whether it gathers the right context, asks permissible questions, secures consent and hands a buyer to a person or system that can act.
Where the agent sits between the ad and the order
A click-to-message format can reduce the gap between curiosity and a first question. It does not remove the rest of the funnel. Teams still need to define what counts as qualified, what information the agent may request, what it may promise and when it must stop.
The conversion layer should have a job description
A strong conversion layer does four things:
- Clarifies intent: it asks a short sequence that distinguishes a useful opportunity from a casual question.
- Routes safely: it matches the answer to a relevant page, calendar, human queue or support path without pretending to decide more than it can.
- Records consent and context: it passes approved data into CRM or a booking workflow with a clear source and timestamp.
- Preserves measurement: it lets marketing and sales reconcile conversation starts, qualified handoffs, attended meetings and downstream outcomes.
That is more valuable than a generic “talk to our agent” prompt. It makes the next action accountable.
Modi’s view: the agent may become the new landing page
Agent CRO is the practice of treating the AI agent as the landing page: the conversation has to qualify the lead, answer the objection, and move the order, because many of those people will never open the site.
I would treat the agent as a conversion layer only when the offer, permitted questions and human handoff are explicit. It should make the next decision easier, not make claims that an evidence page, sales owner or regulated review must still answer.
A conversational start can be efficient, but buyers often need evidence before they book. A concise landing page can supply the proof, service scope, privacy explanation, examples and calendar option that a message thread cannot carry cleanly. The page should not duplicate the entire conversation. It should resolve the decision the conversation is trying to make.
For a B2B service, that might be: “Is this worth a diagnostic call?” For a product, it might be: “Is this configuration available for my use case?” The answer needs current source data and a human escalation route, not persuasive improvisation.
| Job | The agent may do it when | The page or a person must still do it |
|---|---|---|
| Qualify | Company type, objective, challenge, and timeframe, as this page already allows | Infer budget, eligibility, health, or legal standing |
| Answer an objection | Only inside the approved offer | A claim that needs an evidence page, a sales owner, or a regulated review |
| Carry proof, scope, privacy, examples | Never as a substitute | The landing page |
| Move the next step | Read more, book, speak to a person, or leave | An irreversible commitment with no named owner |
| Count a result | Handoff, attendance, and sales acceptance | A conversation count, or ELSA’s 26/18/32 as if it were our result |
This page covers the conversation that replaces the click. It does not cover a catalogue completing a sale with no session: zero-click commerce [blocked].
A conversion operating model for agent-led campaigns
1. Start with a narrow promise
Match the advert, first question and landing-page headline. If the creative offers a PPC audit, do not make the agent pivot into a generic AI consultation. Consistency reduces confused handoffs and makes measurement less ambiguous.
2. Separate qualification from persuasion
An agent can ask about company type, objective, current challenge and preferred timeframe. It should not infer budget, eligibility, health status or legal standing from weak signals. Where a consequential claim or recommendation is needed, route to a qualified person.
3. Design a reversible handoff
Every flow should give the prospect a clear way to read more, book a meeting, speak to a person or leave. A human should be able to inspect the conversation context, correct a mistaken classification and prevent an inappropriate automated follow-up.
4. Reconcile marketing and sales definitions
A “qualified lead” must mean the same thing in campaign reporting and CRM. Record the definition, owner, channel, event date, consent state and outcome. Otherwise an attractive conversation-start count becomes an argument rather than a metric.
What to measure before you scale it
Measure the full sequence: paid impression, conversation start, completed qualification, consented handoff, booking, attendance, sales acceptance and commercial outcome. The outcome may be slower than the campaign dashboard. That is not a defect; it is the reason to join the measurement to the real sales process.
The approach complements the agentic attribution problem [blocked] and the governance of agent-led content and actions [blocked]. For teams looking at the other side of the journey, see how merchant agents are entering ecommerce workflows [blocked].
How this fits a performance and agentic programme
[Image blocked: Agentic conversion handoff diagram showing the route from an approved campaign promise through conversation, permitted qualification, consented CRM or booking handoff, sales ownership and outcome review.]
Decision diagram: use the control path as a planning aid, not as proof of a commercial outcome.
For a distinction between an interface and a conversion layer, open the Gemini connected-apps SaaS interface analysis. The zero-click commerce sibling covers the consumer purchase route; this article stays focused on the consented lead handoff. A B2B lead-generation guide explains the surrounding buyer journey, while the organic, paid and agentic CPA/LTV framework explains why dashboard activity must reconcile with commercial outcome. Paid teams can compare the operating controls in the Google Ads optimisation guide and PPC Performance Max service. Case studies are a useful standard for labelling evidence; they are not evidence that Integrated.Social ran the ELSA campaign.
The first decision is not whether a conversational agent can answer enquiries. It is whether conversation is the right mechanism for moving this buyer forward. Before committing media spend or integrations, apply four gates. The buyer’s intent must be discoverable through a small set of predefined questions, and the next action must be reversible.
Where those conditions hold, an agent can support short discovery and routing. Where intent is already clear, a focused page or form may create less friction. Where the decision is consequential, facts change frequently or the answer depends on judgement, the conversion layer should collect only the context needed to place the prospect with a person.
A useful commercial test is to state the transaction in one sentence: “After this interaction, the prospect will have received X and our team will do Y.” If the team cannot describe that outcome plainly, it has not designed a conversion process that an agent should run.
Build the layer as a controlled system
Pre-launch checklist
- Set one conversion objective. Choose the accountable next step: a qualified enquiry, a booked conversation or a configuration request. Do not ask one flow to educate, qualify, sell and support at once.
- Write an answer boundary. List approved topics, permitted questions, prohibited statements and escalation triggers. Marketing, sales and the relevant subject owner should be able to inspect it.
- Design the page around the unresolved decision. Supply the supporting scope, evidence, process, contact route and an alternative to continuing with the agent.
- Specify the handoff record. Map the minimum sales fields, campaign and conversation context, consent state, timestamp, route selected and owner. Test incomplete and duplicate enquiries before launch.
- Prepare the exception path. Decide who receives uncertain, frustrated or out-of-scope enquiries, what context they see and how the resolution is marked for review.
Release narrowly, then inspect the work
Use one offer, audience and route in the first release. Keep a version record for the advert, opening prompt, questions, page and handoff rules. A performance change can then be investigated: the team can separate a creative issue from a qualification change or a sales-capacity problem.
Review completed, abandoned and escalated conversations alongside their CRM records. It is to establish whether the system collects useful context, applies the stated route and creates work that sales recognises as worth acting on.
The strategic conclusion
An AI agent should not be treated as a replacement for conversion design. It is a new interaction layer that needs a constrained job, a truthful offer, a clear record of consent and a route to a human decision-maker. That is how a conversation becomes an accountable commercial handoff.
For an outside review of the entities, answers and handoffs that support discovery, request a free AI Growth Audit and, if a scoped conversion-layer review would help, book a discovery call.
About the Author
Modi Elnadi is founder of Integrated.Social, a London AI growth consultancy. Since 2014 he has combined performance media with answer-engine optimisation and agentic lead systems for B2B and B2C brands. These pieces are his working point of view for CMOs, not a vendor press release.
