Introducing AI Agent 3.0.
Rebuilt at every step to resolve more of your customer support, on every channel you sell on.
)
)
)
Every part of the system rebuilt.
AI Agent 3.0 is the sum of changes to how the Agent thinks, what it knows about each conversation, and how it decides when your team should step in.
Intelligence
Newer models and a rebuilt knowledge engine, so the Agent understands more and finds the right answer.
Context
The Agent works from the full picture: the right order, the whole conversation, and everything the customer sends.
Control
Clear rules for when people take over, and a reason behind every decision the Agent makes.
The architecture behind AI Agent 3.0.
AI Agent 3.0 is multi-model: every stage of a ticket now runs on a newer model, the one that tested best for that job. Gains at each stage compound, so the reply at the end is noticeably better.
- Stage 1ClassifierFast classification modelReads each incoming ticket and classifies its intent and tone.
- Stage 2Knowledge readerKnowledge reading modelReads your articles, product data, and policies to find what answers the question.
- Stage 3Procedure decisionsProcedure reasoning modelPicks the right procedure and step, and decides what happens next.
- Stage 4Decisions outside proceduresAdvanced reasoning modelHandles questions that fit no procedure, including informational answers and hand-offs.
- Stage 5Reply writerAdvanced reasoning modelOne shared writer for every reply, with each fact tied to a source.
- Stage 6QAAdvanced reasoning modelChecks each reply against the same context it was written from, before it’s sent.
Technical specifications
- Architecture
- Multi-model pipeline with six stages, each on the model that tested best for its job
- Knowledge retrieval
- Cross-language search, exact SKU and part-number matching, second-pass search, live product data
- Grounding
- Every fact tied to your business data, an article, product data or a policy. Quotes checked before sending
- Quality assurance
- Same-context QA review of every reply. Plain-English reasons for every decision
- Order context
- Quoted-order focus across the conversation, order and invoice number recognition, carrier tracking, per-channel terms
- Attachments
- PDFs, images including HEIC, and video, analyzed across the whole conversation, and guarded with fraud detection
- Hand-off triggers
- Product safety, unauthorized orders or charges, failed self-service, requests for a person, strong negative sentiment, health concerns
- Channels
- 300+ native integrations, including Amazon, eBay, Walmart, TikTok Shop and Shopify
- Availability
- Live for all AI Agent customers from October 2026.
The right answer, in any language, down to the SKU.
An AI Agent can only answer as well as it can find. We rebuilt retrieval from the ground up, so 3.0 matches far more customer questions to the right article, product or policy, from whatever words or language the customer uses.
)
Every answer tied to the right order.
Repeat buyers often ask about an older order, not their latest one. 3.0 uses the order number the customer quotes and keeps it in focus for the whole conversation, together with its invoice, tracking and marketplace terms.
)
Handles customers who go off script.
Real customers don’t follow your procedures step by step. Halfway through a return, they ask how long refunds take. 3.0 answers the question from your policy, then picks the procedure straight back up.
Reads the photos, PDFs, and videos customers send.
A photo of a crushed parcel or a PDF receipt often holds the answer. 3.0 doesn’t just recognize attachments, it reads them, and uses what it finds in the reply.
“It has helped us get through customer emails much quicker by asking the right follow-up questions and gathering all the information we need… The AI Agent also recognizes when images are needed and requests these from customers, reducing the need for back-and-forth emails.”
){kind=logo}
Knows exactly when your team should step in.
When you sell physical products, a report of an injury, an electrical fault, or a fire isn’t a job for automation. Neither is a charge the customer says they never made. 3.0 recognizes these moments and hands them to your team, with everything they need to act fast.
Customer reports a charger (order 40512) overheating and a minor burn. One photo attached. Flagged for product safety review.
Every reply grounded, checked, and explained.
You can’t trust what you can’t check. 3.0 ties every fact to a source, reviews every reply before it’s sent, and explains every decision in plain words, so your team always knows why the Agent did what it did.
- Customer received a different version and wants the correct one
- The order is within the return window
- Your return policy allows a replacement for this item
AI Agent 3.0 vs 2.0.
We ran AI Agent 3.0 and 2.0 on the same real eCommerce conversations and measured what matters.
Measured on more than 11,000 live conversations the AI Agent replied to: previous version 18 May to 14 June 2026, 3.0 5 to 18 October 2026. A conversation counts as resolved when the customer didn't need to write back for 72 hours, or when the AI Agent completed the merchant's procedure.
Every change in AI Agent 3.0.
The complete list, filter by area.
- [Oct 2026]Content retrieval overhaulMajor upgradeFinds the right product knowledge in any language, down to exact SKUs and part numbers.
- [Oct 2026]Order handlingGreatly improvedThe order your customer quotes is the one the AI uses, for the whole conversation.
- [Oct 2026]Invoices and trackingImprovedInvoices from the right order, live carrier status and tracking links kept.
- [Oct 2026]Answers inside proceduresNewSide questions mid-procedure are answered, with every quote verified.
- [Oct 2026]Informational fallbackNewGrounded answers to questions that fit none of your procedures. Requires Extended Coverage.
- [Oct 2026]Follow-upsBackFollow-up replies return, and procedures continue when customers send what was asked.
- [Oct 2026]PDFs, photos and videoImprovedReads PDFs, photos, videos and iPhone photos across the whole conversation.
- [Oct 2026]Safety and risk hand-offsNewInjuries, electrical faults, fire, disputed charges and very upset customers go to your team.
- [Oct 2026]Human-agent continuityImprovedFollows the step a human agent has set on the case.
- [Oct 2026]One reply writerImprovedConsistent tone and structure, with every fact tied to a source.
- [Oct 2026]Smarter QAImprovedChecks each reply against the same context it was written from, with fewer unnecessary blocks.
- [Oct 2026]Plain-English decisionsNewA one-line summary and readable reasons for every decision and QA finding.
- [Oct 2026]Newer models at every stepUpgradedClassification, context, retrieval, writing and QA all run on newer models.
AI Agent 4.0 is coming soon.
- Released2.0Summer 2026Improved on the original eDesk AI Agent: adding Procedures to the way the Agent works.
- You are here3.0Autumn 2026New models at every step, rebuilt retrieval, order intelligence, and safe hand-offs.
- Coming next4.0Coming soonMore to come. Sign up to be the first to see it.