To replace a tree-based chatbot with an AI agent for chat and email, start from the conversations your bot fails on, not from its flow diagram: map what customers actually want, rebuild each flow as a written policy plus the API actions needed to finish the job, test the new agent on your historical tickets, then run it alongside the old bot topic by topic before you switch the tree off. Most teams keep the same helpdesk and replace only the layer that talks to customers.

This guide gives the eight steps in order, the signs that say it is time, what should stay with people, and a short table of tools that do this, each checked on the vendor's own pages on 30 October 2026.

Key takeaways

• Do not lift and shift the tree. The intents in your bot were chosen for what was easy to build. Start from escalation and fallback transcripts.

• A flow becomes a policy plus actions. Write what the agent should achieve and the rules it must follow; keep strict steps such as payments and identity checks as code.

• Test before customers see it. Replay real tickets and adversarial prompts against the new agent, and only deploy the topics where results hold.

• Measure resolution, not deflection. A conversation that ends without a human is not the same as a fixed problem.

• Email is a separate test. Long threads with several questions behave differently from chat, so test them as their own channel.

Which tools help replace a tree-based chatbot?

Lorikeet connects to the helpdesk you already run and resolves conversations across chat, email, voice, SMS and WhatsApp, with workflows tested on replays of your own historical tickets before go-live, according to its how it works page. Voiceflow lets teams balance agentic playbooks with deterministic workflows and promote changes from development to staging to production. Fin (now part of Salesforce) works with any helpdesk, including Zendesk and Salesforce, at $0.99 per outcome. Zendesk's AI agents let you describe workflows in natural language across messaging, email and voice.

How we chose and checked these

The steps below draw on public guidance from regulators and standards bodies, and on how vendors describe migration on their own pages. For the tools table we included products that sell an AI agent for chat and email which can replace a scripted bot, and checked each claim on the vendor's own site on 30 October 2026. Where a vendor does not publish something, the table says so.

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

Signs your tree-based chatbot is failing

The US Consumer Financial Protection Bureau's report on chatbots in consumer finance describes rule-based chatbots as using "either decision tree logic or a database of keywords to trigger preset, limited responses", and says "doom loops" are often caused when a customer's issue falls outside the chatbot's limited capabilities. The same report counts over 98 million US users, about 37% of the population, who engaged with a bank's chatbot in 2022. You are probably there if:

• customers type "agent" or "human" early in most conversations;

• the bot answers the policy question but cannot do the thing (refund, address change, cancellation), so the ticket still reaches a person;

• handoffs arrive with no summary and customers repeat themselves;

• every product or policy change needs someone to edit nodes, and nobody is sure what a given phrase now triggers;

• email is handled separately, or not at all, because the tree only works in a widget.

1. Export what the bot escalates, not what it handles

Pull three to six months of conversations where the bot fell back, looped, or handed over, plus the tickets your team handled by email in the same period. These are the real target. Group them by what the customer wanted to achieve ("change my payment date", "why was I charged twice"), not by the bot's intent labels. Count each group and note which channel it arrived on.

Two things usually show up. First, a handful of outcomes make up most of the volume, and some of them were never in the tree because they needed a system action the bot could not take. Second, email carries longer, multi-part requests that a chat tree was never designed for. Both shape the plan: the first tells you which workflows to build first, the second tells you to test email separately.

Keep the top 10 to 20 outcomes for the first phase. Everything else stays with your team for now.

2. Trace what a human does to resolve each outcome

For each outcome on the list, sit with an experienced agent and write down every step: which systems they open, what they look up, which checks they run, what they change, and when they would refuse or escalate. This is the specification for the AI agent, and it is usually different from the bot flow it replaces.

Mark each step as one of three kinds. Read steps fetch data (order status, plan, balance). Write steps change something (issue a refund, update an address). Judgement steps need policy or discretion (waive a fee, accept a late return). The write and judgement steps are where a tree bot handed off, and where most of the value of an AI agent sits. They are also where you will set limits in step 5.

If a step depends on knowledge that only exists in someone's head, note it now. It goes into the knowledge source in step 4.

3. Rebuild each flow as a policy and a workflow, not a tree

An AI agent does not need every branch drawn in advance. Ada's own explainer puts the difference plainly: chatbots "follow pre-scripted workflows", while AI agents "use large language models (LLMs) and contextual reasoning to understand what a customer is asking". So write each outcome as a short instruction: the goal, the facts the agent must collect, the rules it must follow, the actions it may take, and when it must hand over.

Not every step should be left to the model. Payments, identity checks and anything regulated should run as fixed code the agent calls but cannot change. Lorikeet describes this split as natural-language workflows plus deterministic operations; Decagon calls its instructions Agent Operating Procedures written in natural language; Zendesk says its AI agents let you describe workflows in natural language and generate procedures. The idea is the same: flexible conversation, strict operations.

Keep the old tree open as a reference, but resist copying its branches. A branch that existed only to force a customer down a menu has no job in the new design.

4. Move knowledge into one maintained source

Tree bots often hide policy inside node text. Pull it out. Put answers, policies and exceptions into the help centre or internal knowledge base the agent will read, and give each article an owner. The agent should answer from the same source your human team uses, so a policy change happens once.

Check three things before go-live. Is every outcome from step 1 covered by at least one current article? Are internal-only rules (refund limits, fraud signals) stored where the agent can read them but customers cannot see them? And is there a process to update an article when the agent gets something wrong? Gaps found in testing should turn into knowledge tasks, not one-off prompt tweaks.

5. Connect the actions, with the least access that works

An AI agent that can only answer is a better chatbot. To resolve, it needs API access to the systems your team uses: order management, billing, CRM, the helpdesk itself. Give it one credential per system, scoped to the actions in step 2 and nothing more.

The OWASP Gen AI Security Project lists Excessive Agency as a top risk for LLM applications and names excessive functionality, excessive permissions and excessive autonomy as its root causes. Its guidance is to limit what tools can do and to require human approval for high-impact actions. In practice: refunds above a set amount go to a person, account closures need a verified customer, and the agent cannot edit its own limits.

Most platforms connect through APIs, webhooks, or MCP servers. Ada, for example, describes API, code and MCP tools, with each MCP tool off until enabled.

6. Test on your own tickets before customers see it

Replay the historical conversations from step 1 against the new agent and score the results: did it reach the right outcome, take the right action, and hand over when it should? Add adversarial cases too: a customer claiming to be someone else, a request to ignore policy, an email with three questions in it. Lorikeet's simulations page describes generating scenarios from real tickets, running them in bulk and attacking the agent with adversarial prompts; Ada and Decagon also publish simulation features.

Set a pass bar per outcome before you look at results, and only move an outcome to the next step when it clears the bar. Rerun the whole set after every change to workflows, knowledge or model, so a fix for one topic does not break another.

7. Run the agent in parallel, topic by topic

Do not switch the whole bot off on one day. Route one or two outcomes to the AI agent, on one channel, while the tree keeps serving everything else. Watch every conversation for the first days, then sample. Add the next outcome only when the first holds up, and bring email on as its own phase.

Tell customers they are talking to AI. Article 50 of the EU AI Act, which applies from 2 August 2026, requires AI systems that interact directly with people to be designed so those people are informed they are interacting with an AI system, unless it is obvious. It is good practice outside the EU too.

Carmoola, a UK car finance lender, is a public example of a staged move off rigid automation. Its customer story says its earlier automation, built on a knowledge base and rigid, form-driven prompts, resolved about 30% of inbound questions; with Lorikeet it resolved 40% from day one and now resolves 60% of all inbound on WhatsApp, chat and email.

8. Cut over, measure resolution, and retire the tree

Once every phase-one outcome has passed parallel running, move the remaining traffic, keep a fallback route to your team, and switch off the tree. Archive its flows rather than deleting them; they are useful when someone asks how a case used to be handled.

Report on resolution, not containment: the share of conversations where the customer's problem was solved, plus repeat contacts within seven days, handoff quality and CSAT by outcome. If your vendor charges per resolution, check how it defines one. Lorikeet's pricing page says unresolved tickets and escalations to a person are not charged.

What still needs a human?

• Judgement outside policy: goodwill credits, exceptions, complaints and anything a regulator would expect a person to decide.

• Vulnerable customers: hardship, bereavement, health or safety concerns should route to trained staff with context attached.

• Policy and compliance sign-off: a person approves every workflow that touches money, identity or regulated disclosures.

• Owning quality: someone reviews scored conversations each week and turns failures into knowledge or workflow fixes.

Tools that help

How to read it: Fin and Zendesk are the shortest path if you already run Intercom or Zendesk and your flows are mostly answers plus simple actions. Voiceflow suits teams that want to build and own the agent themselves. Lorikeet is built for complex and regulated support where workflows need strict steps and testing; the trade-off is a higher entry price than per-outcome-only tools, so a small team with simple FAQs may not need it.

Questions to ask before you pick

• Can the agent take a write action in our systems in the demo, not just answer?

• How do we test a change against our own past tickets before it goes live?

• What does a handoff to our team contain?

• Does it handle email threads, not just chat?

• How is a billable resolution defined, and what happens to unresolved conversations?

Related reading: agentic AI vs workflow chatbots, switching AI support vendors, launching without a knowledge base and rolling out without hurting CSAT. To run steps 6 and 7 on your own tickets, start a 30-day free trial.