Think about looking for a business on Google Maps. Finding the name is only part of the job. You want enough useful information to decide whether to go there and a practical next step. The business does not need to persuade you to start your search somewhere else.
That is a useful starting point for thinking about AI chats. When customers become comfortable researching a purchase in a conversation, making a business useful there becomes a question of practicality. Can they get an accurate answer, see an available option and continue towards buying?
For a commercial director, this creates two separate investment decisions: how to make the business accessible through that channel, and whether a proprietary agent would add enough value to justify its development and operation. Those decisions can have different answers.
Being useful where a customer searches does not mean every company needs its own app. It does not make the website redundant. And launching an integration does not establish that more people will buy.
Answer in 60 seconds
A branded agent inside ChatGPT can give customers access to a company’s actual capabilities: finding available options, checking current prices, comparing a shortlist or preparing a booking or quotation. The business case is strongest when a useful customer task needs information or actions that the business controls.
First check whether an existing commerce platform, distribution partner or well-formed referral already provides the route you need. A proprietary app needs a reason beyond having your logo inside the chat.
Separate three possible benefits: reaching new customers, helping existing demand convert, and making repeat interactions more convenient. Measure each against its own baseline. A booking attributed to the app is not necessarily an additional booking.
At IZZY, we would scope the useful capability and a bounded test before promising a commercial outcome. We would not promise immediate conversion uplift simply because the integration exists.
In this article
- What a branded agent actually adds
- Where the business value can come from
- When to build your own channel
- Which other businesses should consider it
- What the business still needs to control
- How to judge the economics
- What to test before expanding
1. What a branded agent actually adds
In this article, a branded agent means a customer-facing integration that an AI assistant can use to retrieve business information or perform a defined task. It is more than a general answer about the company, but it need not be an autonomous salesperson allowed to spend money.
OpenAI’s current developer documentation packages integrations as plugins. They can connect to external systems through an MCP server, which exposes controlled tools. Results can be returned as structured information, with an optional interface when a customer needs to compare, edit or confirm something. A bespoke visual app is not required for every capability. OpenAI’s plugin architecture.
Consider a hypothetical hotel group. A traveller describes a destination, dates, party size and budget. The integration checks the group’s booking system and returns suitable hotels, available room options and current rates. The traveller can refine the request and continue to the group’s checkout.
This would let the business bring part of its booking service into the conversation, retrieving prices from an authoritative source. Payment could still happen on the merchant’s website. The exact booking handoff would be part of the product’s scope.
The visible experience may look simple. Behind it, the company still needs dependable availability, pricing rules, a booking destination and a way to handle failures. Building the agent means connecting those capabilities and making them usable in a new context. The AI platform supplies the model; the business supplies the service integration.
2. Where the business value can come from
There are several plausible commercial benefits, but they should not be combined into one promise of increased conversion.
Reach a customer during selection
A buyer who has not chosen a supplier may describe a need before searching for a particular brand. A useful integration could help the business participate at that point, returning an offer that meets the request.
This is an acquisition hypothesis. It depends on whether the target platform makes the integration available, whether the customer encounters or enables it, and whether it is useful for the request. Total platform audience is not the audience of an individual app.
Help an interested customer decide
Someone may already know the brand but struggle to compare suitable options. Current inventory, meaningful differences and a handoff that preserves the selection could reduce repeated searching or data entry.
The commercial question is whether the new route improves that journey. If the same customer would otherwise have booked on the website, the entire booking value cannot be counted as new revenue.
Make an existing relationship more convenient
A repeat customer might prefer to ask for an available service or compatible product within the conversation they are already having. Where supported and appropriately authorised, account-specific information could make the answer more useful.
Convenience can justify investigation before a sales uplift is visible. It still needs evidence: customers choosing the route, successfully completing the task, and experiencing fewer avoidable steps. A possible future retention benefit should remain a hypothesis until it is measured.
There is evidence that AI-assisted discovery produces real website visits, but the scale and relative quality depend on the period, category and method. A 2026 peer-reviewed study analysed 973 ecommerce sites; 49 countries each contributed at least ten million sessions to its data. ChatGPT referrals accounted for less than 0.2% of sessions in the August 2024 to July 2025 sample and performed better in more complex product categories. The study is descriptive and uses last-click attribution, so it supports monitoring and a complexity hypothesis rather than a current market-size or conversion forecast for an app. Marketing Science study.
3. When to build your own channel
The Google Maps comparison has a useful limit. Making your business accessible through an interface customers use is one decision. Commissioning a proprietary product inside that interface is a larger commitment.
Before discussing development, compare the available routes.
| Route | What to investigate | When it may be sufficient |
|---|---|---|
| AI search and referral to your website | Whether useful business information can be found and the destination answers the customer’s next question | The customer mainly needs information and a reliable route to an existing page |
| Existing commerce or distribution integration | Whether your platform or provider can supply current offers to the target channel, under suitable terms | Standard catalogue, availability or purchase handoff already covers the task |
| Proprietary branded integration | Whether your own data, account rules or service logic materially improves the in-chat experience | The customer needs a capability that the existing routes cannot provide adequately |
| Assistant on your own website | Whether visitors already on the site need conversational help choosing | The problem is product finding within your owned journey rather than access through an external channel |
These are distinct routes, not mandatory stages of a maturity ladder. OpenAI’s March 2026 commerce announcement describes product feeds, third-party delivery paths and merchant-owned checkout alongside the option of deeper branded app experiences. Check your provider’s current eligibility and capabilities before buying a separate integration. OpenAI’s merchant discovery announcement.
The distinction becomes clearer in the hypothetical hotel example. A group with several properties may have a useful selection problem to solve across its own inventory. A single hotel may have less reason to build a separate branded discovery tool if customers are still choosing between unrelated properties. It should first examine whether its booking or distribution provider can represent its availability effectively.
Neither size alone nor the presence of competitors decides the answer. A small specialist could have valuable booking logic that a generic provider cannot support. A large group could still struggle to give customers a reason to use its integration.
For technical preparation, use our AI shopping and catalogue-readiness test. If the problem is inside the store itself, the relevant decision is covered in our guide to on-site AI shopping assistants.
4. Which other businesses should consider it
The opportunity is broader than hotels. Look for a buying task where customers describe several constraints, current commercial data matters, and a useful next action can be defined.
The following are evaluation scenarios, not claims that every platform supports them or that these sectors have proven returns.
| Business | A useful customer request | Data and next action | Question before funding a proprietary agent |
|---|---|---|---|
| Retail with complex selection | Find a compatible product within budget that can arrive when needed | Product compatibility, variants, stock and delivery; a preserved product or basket selection | Does the brand’s own selection logic add value beyond existing product feeds? |
| Restaurants and appointment services | Find a suitable location and available time | Locations, capacity, service duration and booking conditions; a reservation step | Can the existing booking provider already offer this route? |
| Local and home services | Explain a job and find a provider able to handle it | Service area, scope and capacity; a qualified request for a quote | Can the team respond and deliver, and can the interface avoid presenting an estimate as a final price? |
| B2B suppliers | Identify an available part that meets a specific requirement | Verified specifications, compatibility and applicable account access; a quote or approved ordering route | Does restricted pricing or specialist knowledge justify a controlled integration? |
A simple catalogue with an easy purchase path may gain little from a long conversation. A complex service that cannot establish price or suitability without inspection may benefit from better qualification rather than automated selling. Start with the decision the customer needs help making.
5. What the business still needs to control
Moving an interaction into a chat changes where it starts. It does not remove the commercial responsibilities behind it.
Before commissioning the integration, assign ownership for six things:
- The offer: which system provides the current price, availability, restrictions and terms?
- The customer relationship: what identity or account information is available, with what permission, and what can the business actually retain or use?
- The commitment: who confirms a reservation, order or quote, and what happens if the underlying availability changes?
- The handoff: does the next screen preserve the selected option, or make the customer start again?
- The service afterwards: who handles amendments, cancellations, fulfilment and complaints?
- The evidence: can the company reconcile the interaction with a valid outcome in its own systems?
Customer identity, access to conversation data, checkout control and a right to contact someone again are separate questions. Do not hide them inside a promise that the business will “own the customer”. Resolve the actual data flows, platform terms and permissions for the chosen implementation.
There is also a release dependency. OpenAI’s current public publishing flow includes review, approval and a separate developer publication step. Changes to reviewed remote metadata can require another submission. Allow for that work in the operating plan; a directory listing does not establish how often the integration will be used. OpenAI’s submission guidance.
6. How to judge the economics
Start by naming what the business is buying. It may be a limited experiment in customer access, a service improvement for existing customers, or a new acquisition channel. Each needs a budget ceiling and an evidence standard appropriate to its purpose.
Do not use one conversion dashboard. Keep four ledgers that answer different questions.
| Ledger | What belongs in it | What it can establish |
|---|---|---|
| Attribution | AI referrals, assisted interactions and valid transactions tagged to the app | Which activity touched the channel |
| Incrementality | Total orders, contribution margin and new-to-brand customers created by exposure | Whether the business gained an outcome that would otherwise not have occurred |
| Substitution | Orders diverted from direct, organic, paid, affiliate, marketplace, staff-assisted or owned-app routes | Which existing channel gained or lost volume and margin |
| Operations and risk | Build and run cost, support work, latency, failures, refunds, cancellations, fraud and remediation | The full cost of providing the capability and keeping its promise |
This creates a useful discipline:
AI-attributed demand = new demand + accelerated demand + diverted demand + attribution error
Only the first component is unambiguously new. Accelerating a purchase can still be valuable. Diverting a transaction from a more expensive channel can improve contribution. Both need their own evidence and should not be reported as new customer acquisition.
For a sales-channel investment, compare the contribution the business earns with and without the new route. Contribution means what remains from a valid sale after the variable costs of delivering it.
Net channel value = incremental contribution + verified operating savings − build and run cost − error and remediation cost − lost margin from substitution
Use the same evaluation period throughout and prevent double counting. If a reduction in staff-assisted service is already included in per-order contribution, do not add it again as an operating saving. A one-off build estimate is not the total cost of the channel.
For the hypothetical hotel group, this separates three very different results: an additional stay, an existing booking made sooner through a different interface, and a booking shifted from a channel with different distribution costs. The app could record activity in all three situations while producing very different financial outcomes.
Use your own costs and evidence. Public referral growth does not supply a conversion forecast for your app, and a vendor’s demonstration does not establish demand. Without a credible comparison, report attributed outcomes and the remaining uncertainty rather than claiming incremental return.
7. What to test before expanding
A useful pilot begins with one customer task and an explicit route into the experience. Specify the target market, eligible users, data sources, next action, budget limit and the decision the pilot must inform.
Where traffic and risk allow it, use three comparable arms:
- Existing journey: the current website, app or assisted route.
- AI referral: conversational discovery with conversion on the existing merchant journey.
- AI action: the same discovery opportunity plus the bounded proprietary capability.
Randomise access where possible instead of comparing self-selected users of the new tool with everyone else. Measure total contribution per eligible customer across all channels, then inspect new-to-brand customers, paid and direct displacement, cancellations, returns, service contacts, stock or quote errors and the cost of operating the integration. If individual randomisation is impractical, a matched-market design may be possible, but it needs enough volume and a defensible comparison.
Test named-brand requests separately from unbranded requests. A customer who explicitly asks for your business is different from someone looking across a category. OpenAI recommends testing direct, indirect and negative prompts when evaluating tool selection. Those tests establish behaviour in the tested setup; developer-mode success does not prove discovery among customers who have not enabled the integration. OpenAI’s metadata evaluation guidance.
Follow the whole measurable journey: relevant exposure, invocation, useful result, next action, valid order or booking, and later cancellation or support outcome. Record which parts you cannot observe. Compare like-for-like journeys where possible, accounting for differences in customer intent, market, availability and promotions.
Agree what would change the decision:
- If customers use the capability and reach the intended outcome, assess whether the benefit warrants further investment.
- If the information or handoff is unreliable, repair that part before increasing exposure.
- If a limited pilot generates too little relevant use to judge, record insufficient evidence and revisit the access hypothesis within the agreed budget.
- If an existing provider route solves the task at acceptable cost and control, use it.
A missing immediate conversion uplift does not erase a demonstrated convenience benefit. Equally, convenience should not become a permanent excuse for costs nobody can justify. Decide which benefit matters, what would count as evidence and when the business will review it.
Decide which part of the customer journey is worth bringing into the chat
The practical ambition is to let a customer get something useful done where they choose to look. For one business, that may mean accurate information and a reliable link. For another, it may mean live availability, a personalised shortlist and a booking handoff supported by its own systems.
Choose the least complex route that delivers the required customer value. Fund a proprietary agent when the capability it adds, the control it preserves and the evidence it can produce justify the extra commitment.
Bring the task you want to support, the system that holds the relevant commercial data, and any evidence of how customers currently find and buy from you. We can examine the integration options and define what a bounded test would need to establish. The starting point is a useful customer capability and a measurable decision, without a promise of instant conversion. Scoping and building that bounded capability is the work of our AI Agents & LLM Products service.
Sources and scope
Platform documentation and linked IZZY pages were checked on 9 September 2026. Capabilities, review requirements and availability can change; confirm the chosen platform, market and merchant eligibility before implementation. The hotel and sector examples are illustrative scenarios, not client case studies. The four ledgers, channel-value equation and pilot design are IZZY’s analytical tools, not a validated return-on-investment model. The cited referral study is descriptive evidence about an earlier period, not the causal effect of a proprietary app.