When the Website Visitor Is an Agent: Jeff Lunsford on Context, Identity and the Test of Commercial Value

1. Guest Background

Jeff Lunsford, CEO of Tealium, joins host Craig S. Smith for an approximately 56-minute episode of Eye on AI about the data infrastructure behind agentic commerce. The discussion asks what businesses must supply when software agents arrive on a customer's behalf: current behavior, understandable attributes, identity relationships and consent status. The title's 50% traffic figure is Lunsford's forecast, rather than an independently established finding about the web.

His relevant experience helps explain the emphasis on infrastructure. Lunsford recalls entering banking technology after leaving the Navy in 1994, encountering an industry whose overnight batch processes have proved durable. He describes Tealium's origins as a two-person consulting business in 2009, followed by funding and the development of a software company in 2011. A team with a web analytics background began with website tag management as its entry into the broader problem of fragmented omnichannel data.

Lunsford now positions Tealium as a neutral layer between customer experience systems and the warehouses and transactional systems behind them. He says it integrates with approximately 1,300 vendors. That gives the interview a specific operating perspective: collecting events, mapping attributes, resolving identity and distributing context across heterogeneous systems. It also establishes the commercial lens through which his argument should be read. He is explaining a problem his company sells infrastructure to address; the proposed mechanisms and the claimed financial outcomes therefore deserve separate evaluation.

2. What the Episode Covers

The first useful clarification concerns orchestration. Smith asks about a term often used for managing agents and their workflows. Lunsford explicitly says Tealium does not take control of an entire agent. Its job is to collect, label, filter and enrich data, then supply context to an agent or model, including the visitor's consent status. Fraud algorithms and next-best-action systems can use the same infrastructure. Supplying the inputs to a decision is a different responsibility from planning and executing the subsequent task.

In Lunsford's account, events arrive from websites, mobile applications, call centers and physical locations. The platform associates them with identity, enriches a visitor profile and distributes the result to destinations such as advertising platforms, email systems and warehouses. He contrasts this with successive batch ETL pipelines that assemble and pass along historical data. The proposed advantage is that a new interaction can influence another system while the interaction still matters.

He also gives real-time orchestration a reliability dimension. A receiving vendor may fail to catch an event, so the sender needs to detect the failure and resend it. Labels, routing, observability and delivery confirmation belong in the architecture alongside speed. The interview does not explain duplicate suppression or a complete delivery guarantee. Its description establishes the need to handle failure, without supplying enough detail to infer how every failure mode is resolved.

A retailer operating four brands supplies the clearest example. A shopper browses jeans on a phone in a mall and subsequently enters a store. An employee could see that recent interest on a tablet and use it in the conversation. The same profile could inform a call center, abandoned-cart messages, retargeting or loyalty activity. Lunsford describes both a store system pulling a profile when needed and rules pushing information after an event. He does not explain precisely how the arriving shopper is identified, an important dependency beneath the apparently seamless experience.

The architecture remains deliberately mixed. Tealium accepts batch files, and audiences can live in a warehouse, a customer data platform or a marketing cloud. Lunsford says the company mainly replaces point-to-point integrations. He also distinguishes behavioral experience data, such as activity around a loan application, from account balances supplied by transactional backends.

CloudStream illustrates the warehouse relationship. According to Lunsford, an inbound event can trigger extraction of a profile or an entire audience for activation elsewhere. Smith's follow-up establishes that data does leave the warehouse, so language about avoiding a shadow copy cannot mean that nothing moves. Caching and retention remain unexplained. Meanwhile, flexible attribute mapping provides a shared representation across systems with different schemas, allowing one channel to use what another has observed.

3. Core Views: Reasoning, Examples, and Limits

The strongest argument in the interview is about the relationship between information and an opportunity to act. Lunsford invokes airline passengers needing a gate or baggage claim during a journey, and a car buyer who has reserved Saturday for making a purchase. In each case, the usefulness of context depends on when it arrives. A model can reason competently and still respond poorly if its picture belongs to an earlier stage of the situation. This supports matching data freshness to a decision window. It does not establish that every dataset needs the same latency or deserves the same expenditure.

Lunsford extends that argument into performance claims: real-time experiences, he says, convert at rates 30% to 200% above delayed experiences, while an email sent four minutes after a website departure has roughly six times the open rate of one sent after 24 hours. The episode supplies no sample description, controlled comparison or consistent measurement framework for those figures. An email open is also a different outcome from a completed purchase. The figures are useful prompts for experiments about timing, but they cannot safely become a general forecast of financial returns.

The agent traffic discussion requires similar care with denominators. Lunsford says agents account for about 20% of customers' good inbound traffic, separately noting that bots represented about 40% of traffic before AI. He predicts an agent share of roughly 50% in a year and also uses the phrase mid next year. These are different categories and a future expectation, not interchangeable measurements. The supplied evidence does not establish the recording date or explain agent classification, so the upload date should not silently become the forecast's starting point.

His underlying growth thesis is more interesting than the headline percentage. Agents make searching for a trip, booking dinner and pursuing other tasks easier, potentially increasing total sessions. A larger agent share therefore need not imply that human visits have halved. Websites may remain the front door even when some interactions no longer require HTML rendering. Lunsford treats agents as another channel layered onto existing ones, with human interfaces, personalization and consent management still necessary. The implication is shared context across channels; whether additional activity produces profit remains a separate question.

For machine-facing interactions, he recommends enriching the website data layer and exposing relevant information through JSON, JSON-LD, APIs or supported protocols. His reasoning is that a compressed purchasing path still needs information that a person might otherwise discover across several pages. Making that information interpretable could reduce friction. It does not, by itself, establish permission to transact. Lunsford also claims agent sessions generate approximately 38% more revenue per session than human sessions and suggests agents will favor websites where tasks are easier to complete. Without classification methods, comparable samples or ranking evidence, the revenue comparison cannot establish incremental lift caused by agents, and the preference claim remains an expectation.

Identity is another unfinished part of the proposal. Lunsford offers two possibilities: attach an agent attribute to a human profile, or give the agent its own identifier and associate it with a person. He mentions budgets and authorized activities, but does not describe a complete verification mechanism. Recording who represents whom is valuable; it is insufficient to establish what that representative may do in a particular transaction.

Finally, the commercial examples combine several possible causes. Lunsford reports average improvements of about 25% in conversion-related advertising metrics after conversion API connections, and 30% to 40% conversion improvements in travel when replacing day-old data. His dealership example describes lead-to-sale conversion rising from 10% to 30% after data sources were cleaned and a shared layer established, followed by greater confidence in acquisitions. These are supplier-reported outcomes without reproducible experimental designs in the episode. The dealership changed data quality and coordination as well as timing. The defensible lesson is that those factors may work together; the interview does not isolate the contribution of real-time delivery.

4. Learning and Application

A practical starting point is one decision with an explicit time requirement. Choose the moment a store employee opens a profile, a service representative answers a call or an agent requests information. Then inspect when the underlying data was produced, when the profile changed, when the warehouse refreshed and when the destination acted. The interview describes both batch ingestion and event-triggered extraction. That combination makes an important boundary visible: retrieving a record immediately does not make its contents current. If the bottleneck is upstream, faster delivery may simply distribute stale information more efficiently.

Acceptance criteria should follow the decision through the complete path. Specify the attributes actually needed, verify the identity association and confirm that the destination receives usable information. Because Lunsford discusses retries, ask what happens when an event is processed twice. Repeated delivery could matter differently for a profile update and a triggered customer message. Pulling context when an employee requests it and pushing context after an event serve different moments; select between them according to the use case and failure consequences. Connector coverage helps establish reach, but cannot replace this verification or explain the missing store identification step.

For external agents, separate identity association from execution authority. The episode's ingredients suggest recording the person, the agent identifier, the budget, permitted activities and consent status. Turn those into questions that must be answered for a particular task. Reading a preference and completing a purchase should not inherit identical permissions merely because the same profile is involved. This is an implementation requirement derived from the discussion, rather than a finished specification supplied by it. Likewise, enriching a data layer should follow a decision about which recipient needs which attributes; access to a comprehensive profile does not justify exposing that profile wholesale.

Internal agent use deserves a distinct pilot. Lunsford says the MCP server introduced in June allows tools such as Claude and Codex to create rules, attributes, campaigns and audiences through natural language. That changes platform configuration, whereas a consumer agent visiting a website is a different workflow. Begin with a bounded change whose output can be inspected, establish the operating permissions and determine how to restore the prior state. The interview explains the attraction of a lower learning barrier but leaves approval and rollback mechanisms unspecified. Easier instruction is not evidence that change management has become unnecessary.

The financial test should include volume. Lunsford describes volume-based pricing, annual commitments and overage tiers, alongside returns from engineering efficiency, speed, agility and advertising performance. If agents generate more sessions, compare useful task completion and revenue with processing costs and staff time. Record concurrent data cleanup and integration changes so that a successful pilot still tells the organization where its next investment belongs.

Finally, model integration savings using the actual topology. Lunsford calls the connection problem factorial, but his example of fully connecting frontend and backend systems grows by multiplication: two on each side require four links. A shared layer may reduce maintenance, yet the benefit depends on which connections exist and how mappings are maintained. Testing that burden alongside freshness, permissions and usage costs turns the interview's architectural direction into a decision with observable success conditions.

Source

• Original episode: AI Agents Will Take 50% of Your Traffic — Is Your Data Ready? | Jeff Lunsford, CEO of Tealium