Delegating authority to AI Agents for rail ecommerce

The integration of Artificial Intelligence [AI] and online retailing for a sector such as transport isn't so much a technical jump as one of trust (and consent based upon that trust).

It's already possible to automate the manual ecommerce journey via scripts (as long as you always want the same journey and ticket purchased each time). This technique is used frequently on concert booking sites, where unscrupulous parties use automation tools to buy up rows or even sections for popular acts. Safe in the knowledge that they can be resold for a healthy profit.

But now the time has come for the normal user to use Agentic AI for themselves to book everyday things... Including transport tickets. However this means them placing their trust in a machine (albeit one made entirely in software). And delegating the responsibility to source the correct fare and purchase the required product... Just once.

Does Rail Agentic Commerce needs an online account?

To create a truly personalised rail ecommerce booking experience for an individual just by using Agentic AI means providing customer specific information at different steps in the typical online retail rail process.

Let's consider the first step... an Origin & Destination journey search.

Probably initiated by a user prompt such as "buy a rail ticket from X to Y". 

This is fine if the customer wants to travel now, but not for travel tomorrow at 8am.

To start this search they need to say "buy a rail ticket from X to Y, to depart around 8am tomorrow".

And what about if the person has a Railcard (say an over 50s discount)?

Are they then going to ask

"buy a rail ticket from X to Y, to depart around 8am tomorrow using my Club50 Railcard"?

This is already getting quite combersome and we haven't even got to the complex stuff yet...

Would it not be much easier to just state "buy my usual train ticket for tomorrow morning"?

Based on the assumption that: 

1. They do actually have a history of bookings for the same origin & destination and around time of day to incur a peak fare product

2. The Railcard has been associated / linked with them.

But to do this requires the AI Agent to access their previous purchases (regardless of whether these were made by a human or an AI) and to see where & when these tickets were for. 

In other words, I can't see how you get more personalised Agentic Commerce unless AI access to a customer account is granted. 

The Machine Economy - transport is next

I think the integration of AI & automation into the transport ecommerce process is about to gain significant momentum.  

Currently there's several rail journey planning projects at a decent technology level (TRL6 and above). So the next logical step is to enable a machine user (not a human user) to complete the entire online transaction process autonomously and at scale.

Sure there's a few hoops to jump through. But these aren't really technical, they are mainly trust and consent based. E.g. how is a Customer going to safely give permission to an AI to use their online account and their payment details?

Creating trusted AI Agents for Rail

Why would I trust an AI Agent to automatically book a rail ticket on my behalf (e.g. without my input beyond an initial prompt / request)?

The answer lies in using a combination of AI and factual data sources. A technique called Retrieval-Augmented Generation [RAG].

RAG thus relies on merging the usefulness of AI with trusted information from official sources (e.g. rail timetables published as Open Data and data purchased from a reliable supplier, such as local events)

This helps to eliminates hallucinations, by enabling autonomous agents to plan, recommend products, and potentially execute complex, multi-step transactions.

There's therefore no reason in my mind not to use RAG to also query customer specific data such as transport preferences and previous journeys made. Provided it's done securely.

Should we trust AI Agents for transport?

Generative AI still hullucinates e.g. makes stuff up. This is because Large Language Models (LLMs) predict the next likely word, rather than checking facts. So they often "guess" the answers.

As an example in the transport sector...

A customer-service chatbot on the Air Canada website told a traveler they could claim fares that did not exist (a Canadian tribunal later ruled the airline was legally liable for its chatbot's hallucination)

The landscape of AI hallucinations therefore highlights both progress and ongoing challenges. 

So why trust AI Agents for automating ecommerce transactions?

Transport Journey Planning is easy for AI, so is it innovation?

Public Transport Journey Planning using AI is now becoming easier and relatively commonplace for multiple modes and different use cases. 

(Perhaps because trusted transport Open Data from public sources is now so prevalent across the UK?).

This recent email I've received from OpenAI exemplifies what I mean.

But Agentic Commerce isn't about creating anonymous journey plans for everyone.

Agentic Commerce is specific to the individual and their transport needs. It's the addition of more user preferences and the storing of customer information to create a bespoke digital rail booking service for each user.

So next time someone says they are using AI for transport journey planning... I will have to question if it is really Innovation, or just what is expected from AI now. And perhaps the real innovation public transport is cracking how to fully automate the Agentic Commerce process.

Agentic Commerce and the rail booking process

The typical online rail booking process (roughly) breaks down into several well known steps:

1. Origin & Destination search, with additional date & time and Railcard / concession parameters to find a journey

2. Selection of a single fare, usually the cheapest, from a range selected - maybe done twice for a return

3. Addition of extras, such as seat reservations 

4. Payment and confirmation 

Currently a lot of AI Agent work I have seen focuses on the first step and maybe the second (if the rule is to always find the cheapest single). 

But the real purpose of Agentic Commerce is to complete all 4 steps automatically on behalf of a customer. Ideally without any human intervention.

Agentic Commerce and rail transport is a good fit

I think that the rail sector should be one of the quickest to adopt autonomous AI Agent-based online retailing. Before other modes of transport such as air travel. 

Why?

1. Because the price is comparatively low (e.g. around £10 - £20 for a shorter daily journey)

2. Because origin and destination points are fixed (making it much easier to agree on the station name or identifier - e.g. CRS code)

3. Because train routes and timetables are agreed & standardised, making Journey Planning easier (e.g. it is easier to trust the system and mistakes are less likely to be made by an AI making decisions on behalf of a human)

Agentic Commerce for Rail

Agentic commerce for Rail could mean many things to different people. But in my opinion it means using AI agents to autonomously: plan a journey, compare prices (fares), apply any relevant discounts (e.g. Railcard), and then purchase train tickets on behalf of a customer.

The concept isn't that much of a technical jump from the current human powered booking process. And could be one of the more regular and smaller value agentic transactions carried out.

The perils of sharing different types of data at once

Organisations of all sizes have data that is of benefit to users both inside and outside their boundaries (both in the private and public sectors). Therefore having the tools and processes to find and share this data should make things run more efficiently and effectively... hopefully.

In a recent client conversation, we discussed the creation of a data portal / platform for the easier discovery and sharing of data. This inevitably led to the discussion about the types of data that the organisation wanted to share and who they wanted to share it with. Leading to the realisation that they actually had the need for sharing data that ranged from the very secure (restricted customer or commercially sensitive information) through to Open Data (information that they wanted to share for free outside of the organisation). They also had requirements for sharing data that sat somewhere in the murky area between those two extremes (information limited either by license or by access / user) including some they wanted to monetize.

Most data platform projects I have worked on previously have focused on the sharing of particular types of data (e.g. just Personally Identifiable user information or a mix of limited /restricted and Open data), but not the need for sharing different data sources from across The Data Spectrum:

https://www.theodi.org/about-the-odi/the-data-spectrum/

So creating a single data sharing technology platform that can share any sort of organisation data creates a number of issues:

Privacy and security:
Balancing accessibility and protection is tricky. You want your service to be easy to use, but with more sensitive data, strict access controls and strong security measures are vital.

Data standardisation:
Different users across an organizations often format their data differently. Such a service needs to be able to handle these inconsistencies or offer tools to convert data to a common format for sharing.

Traceability & reusability:
Data consumers need to understand how the data was collected and what it can be used for. Your service should provide clear audit trails and data provenance, as well as ways of licensing and charging for the data - especially if the data consumer is an external user.