An AI chatbot client handover is the moment you prove, with evidence, that the assistant answers correctly, refuses what it should, sends leads to the right inbox and can be rolled back. The proof is a short pack of six documents the client signs, followed by a review of the first week of real conversations.
Key takeaways
- A demo is not proof. Hand over a test report run against the exact version going live, plus the release record that version came from.
- A lead that never reaches anyone is a silent failure: nobody sees an error. Prove lead delivery with a real conversation in Production.
- Settle ownership before you build: who owns the account, the prompts, the knowledge and the conversation data, and how the client leaves.
- Sell quality as two lines, a setup and acceptance fee and a monthly review, not as an hourly afterthought.
- The first-week review is where trust is won. Book it at handover, not after the first complaint.
If you have not built your test set yet, start with how to test an AI chatbot before launch, then come back for the handover.
Why handover is where agency projects go wrong
A common way an agency chatbot project goes wrong is quietly, after launch. Picture the pattern: the client approves a demo, and a few weeks later the front desk asks why nobody has heard from the website. The causes are usually ordinary:
- The version that was approved is not the version that is live. Someone fixed one more prompt the night before launch.
- Leads go to the agency, not the client. Or to a webhook still pointed at a test URL.
- Nobody agreed what "ready" meant. The client expected the assistant to quote prices; the agency scoped it to answer opening hours and book calls.
- The client is responsible for what the assistant says. In Moffatt v. Air Canada (2024 BCCRT 149), decided on February 14, 2024, British Columbia's Civil Resolution Tribunal held the airline liable for a bereavement-fare policy its website chatbot described incorrectly, and rejected the argument that the chatbot was a separate party responsible for its own statements. Your client carries the same risk, so they deserve evidence.
A handover pack turns "it looks good" into a record the client can sign.
What goes in an AI chatbot client handover pack
Six artifacts, each with a named sign-off. Keep them in one shared folder per client, and link every number to the run or record behind it.
| # | Artifact | What it proves | Evidence | Who signs off |
|---|---|---|---|---|
| 1 | Scope sheet | What the assistant should and should not do | One page: topics, refusals, hand-off rules, lead fields, channels | Client owner |
| 2 | Test report | It answers the agreed questions correctly | Test run against the release candidate, pass rate, failures and fixes | Agency lead, then client |
| 3 | Red-team summary | It holds up when someone tries to break it | Attack run results, what got through, what changed | Agency lead |
| 4 | Lead-routing proof | Leads reach the right people | A real Production lead in the client's inbox, with Reply working | Client's front desk |
| 5 | Release record | The approved version is the live version, and it can be undone | Release ID, who created and activated it and when, rollback target | Agency lead |
| 6 | First-week review plan | Someone will look at real conversations | Date, attendees, the metrics you will show | Both |
1. Scope sheet
Write it before you build and get it signed. It is the contract the tests are written against. Include:
- In scope: the topics the assistant answers, each with its source (a page on the site, a PDF, a price list).
- Out of scope: what it must decline, such as medical, legal or pricing advice beyond the published list.
- Hand-off rules: when it should offer a person, and how.
- Lead fields: exactly what it collects (name, email, phone, the service they want) and where each lead goes.
- Tone: three adjectives and one example reply the client likes.
2. Test report
Every scope line becomes at least one test case. In Rylvo, Tests are scripted conversations run in Test against the Test version of each bot: rules are checked by code, and quality is scored by a judge model you choose. Test kinds include happy path, guardrail, escalation, tone, policy compliance, hallucination probe, PII check and multi-turn. Tests are on every plan, with a cap on cases per bot (5 on Free, 50 on Lite). Owning the scoring, meaning your own pass bar, weights and dimensions, is on Pro and Enterprise.
The rule that matters: run the tests against the exact candidate you are about to release. Freeze the candidate, switch on Hold releases under Tests > Release gate (any plan, admins and owners), and run the tests while the candidate waits. Put the date and the candidate ID on the report. For a deeper treatment of nightly regression runs, see automated chatbot regression testing.
Here is a one-page report you can copy:
ASSISTANT TEST REPORT
Client: ________________ Assistant: ________________
Version tested: Release candidate ID ________ Date: __________
Judge model: ________________ Tested by: ________________
RESULTS BY KIND Cases Passed Notes
Common questions ____ ____ Sources match the scope sheet
Out of scope / refusals ____ ____
Hand-off to a person ____ ____
Lead capture ____ ____ All agreed fields collected
Privacy (PII) ____ ____
Multi-turn ____ ____
TOTAL ____ ____ Pass rate: ____%
FAILURES AND WHAT WE DID
1. Case: ____________ Reply: ____________ Fix: ____________ Re-run: pass / fail
2. ...
KNOWN LIMITS (accepted by the client)
- ________________________________________________
Signed (agency): ______________ Signed (client): ______________
Use "known limits" for the questions it still answers imperfectly, so the client hears it from you first.
Agree the pass bar in the scope sheet, not after the run. A workable default: every refusal, hand-off, privacy and lead-capture case passes, and any failure in common questions is either fixed and re-run or listed under known limits and accepted in writing. Set a percentage only if the client has asked for one.
A filled-in entry, for illustration:
- Failure: "Do you offer weekend appointments?" The reply said yes.
- Cause: an old page in the knowledge base still listed Saturday hours.
- Fix: remove the page, re-run the case, pass.
- Known limit: it cannot quote prices for custom work, so it offers a callback instead.
3. Red-team summary
Tests check what you expect. A red-team run checks what a hostile or careless visitor might try. Rylvo's Red Team, inside Edge Cases, writes attacks for the agent from its own instructions across 14 attack kinds, such as prompt injection, jailbreak attempts, PII extraction, hallucination bait and billing bypass, and sends them in Test. You choose the model that writes the attacks and the judge. Attacks that get through are filed as Edge Cases with their evidence.
Red Team and Edge Cases are on Pro and Enterprise, and the 7-day Pro trial includes them. Map the summary to a public reference the client's IT contact may recognize: the OWASP Top 10 for LLM Applications (2025) lists prompt injection as LLM01 and sensitive information disclosure as LLM02, as of November 2026. For the prompt-injection side, see how to protect a customer service chatbot from prompt injection.
One honest line belongs in the summary: by default nothing is changed without a person. Rylvo diagnoses each Edge Case and prepares a fix, and a person clicks Fix in Test. If you switch on "Fix in Test by itself", Rylvo may change a setting in Test and keeps it only if the replays pass. Either way, a fix reaches customers only when someone releases it. The exception is Edge Case rules, which apply to Test and Production at once.
4. Lead-routing proof
This is the easiest step to skip and the cheapest to prove, so it has its own section below. The proof is a real lead from the live assistant sitting in the client's inbox, answered with one press of Reply.
5. Release record
The client approved a version. Prove that version is the one answering customers. In Rylvo every change is made in Test and reaches customers in a release: bots, prompts, flows, skills, guardrails, knowledge and automations frozen together. Each release records who created it and who activated it, and an admin or owner can roll back to any release that was live before in one step. Connectors, channels and settings are shared by both environments and are not part of a release, so note any change to them separately. See Test and Production for chatbots for the full model.
Record the release ID, who activated it and when, and the ID of the release you would roll back to.
6. First-week review plan
Book a 30-minute call for seven days after launch, and put the date in the pack you hand over. The agenda is below.
Lead delivery to the client's inbox (with a reply path)
Rylvo can email each new lead from an assistant straight to the client's staff, the front desk or the owner:
- Up to 5 addresses per assistant. Each address receives nothing until its owner confirms it through a signed link, so you cannot sign up a client's colleague by typo.
- Reply goes to the lead. The email's Reply-To is the visitor's address, so the client answers with one press of Reply.
- The client's look. The email comes from the assistant's name and carries the client's logo and color.
- Production leads only. A lead captured in Test never leaves the workspace. So the proof has to come from the live assistant.
- Paid plans only, with a daily cap: up to 100 lead emails per assistant per day (20 on Starter). A lead past the cap is still captured in the Leads inbox; only the email stops.
- You keep a copy. The workspace owner and admins still get every lead by email unless they turn lead emails off in their profile, and every client email carries its own unsubscribe link.
For a CRM, send leads by webhook to Zapier, Make, n8n or the CRM itself through a connector. Connectors serve Test and Production alike, so each one has a Test button, and a Test reply only calls a connector after an owner or admin verifies its destination for Test.
The proof, step by step:
- Release the approved version to Production.
- Have the client's front desk confirm their address from the confirmation email.
- From a phone that is not on the office network, open the live widget and leave an enquiry with a real email address you control.
- Check that the lead arrived in the client's inbox with every field from the scope sheet.
- Press Reply and check that the reply goes to the visitor's address, not to the agency.
- If you use a webhook, check the lead reached the CRM. Note the time of each step in the pack.
For lead-capture design itself, see the lead generation chatbot guide.
Who owns the assistant after handover
Decide this before you build, and write it into the contract. There are three workable models in Rylvo:
| Model | How it works | Client gets | Fits when |
|---|---|---|---|
| Client owns, you advise | The client signs up and pays; they invite your team as members when they want hands-on help (a Pro seat each) | Full control and billing from day one | The client wants to own the asset; you sell a build and a retainer |
| You build, then transfer | You build in a workspace, the client joins as a member (a second person, so Pro), and you transfer ownership to them; you become an admin and lose billing rights | Ownership of the workspace; agree who pays for the plan before you transfer | You want to build fast without waiting on the client's sign-up |
| You run it for them | The assistant lives in your own workspace; you operate it as a service | Lead emails, read-only analytics links, a monthly report | Small clients who never want to log in |
Things the table does not say:
- A second person in a workspace needs Pro. Free, Starter and Lite hold one person. On Pro ($100 a month), each teammate is $40 a month, so the first two models cost a seat for whoever is not the owner.
- Roles are workspace-wide. Rylvo has four fixed roles, owner, admin, operator and viewer. Do not invite a client, even as a viewer, into a workspace that holds other clients' assistants.
- Ownership transfer is one step but needs a member. Only the owner can transfer, and only to an active member. The old owner becomes an admin, and the new owner can remove them later.
- In your own workspace, releases cover every assistant in it. You can hold back changes so a release for one client leaves another client's assistant exactly as it is, but check the release contents every time. Pro holds up to 25 agents and 50 deploy places, and its about 2,650 included AI replies a month (counted on the default model; a stronger model uses more) are shared across all of them.
- Write the exit down. Who owns the prompts, the knowledge files and the conversation data, how long you keep them after the contract ends, and what you export for the client.
The first-week review: what to show the client
Keep the review to five numbers and three conversations.
Before the call, open the week's traces: one trace per reply, with each search, tool call and model call. Read every conversation that ended in a hand-off and every one the visitor left early. On Pro, also check Edge Cases: with "check every reply as it lands" on, each kind of problem is filed once with its evidence.
On the call, show:
| Show | Why it matters | Where it comes from |
|---|---|---|
| Conversations and leads | Is the assistant earning its place? | Analytics room (Lite and above), Production only |
| Questions it could not answer | The next content to add | Knowledge gaps in analytics, traces |
| Hand-offs to a person | Are the hand-off rules right? | Analytics, traces |
| Issues found and what you did | Proof you are watching | Edge Cases (Pro) |
| Two good conversations, one bad | Real examples, honestly chosen | Traces |
Analytics rooms are on Lite and above (not on Free or Starter) and are split by environment, so test chats stay out of the numbers you report. To give a client a view without a login, an admin can share a read-only link to a room: it shows the room's numbers as they were when shared, never reply text, expires, and can be revoked. Daily or weekly email digests go to workspace members, so a client who is a viewer in their own workspace can get one each Monday. The chatbot analytics guide covers which metrics to put in the room.
End with a decision: which questions to fix, which test cases to add, and when the next release goes out. Repeat monthly.
How much to charge for chatbot QA
Quality is easy to under-price when it is bundled into the build. Split it into two lines the client can see:
- Setup and acceptance (one-time). The scope sheet, test set, red-team run, lead-routing proof and handover pack. Price it from your hours; the pack is the deliverable the client signs.
- Monthly review (recurring). The first-week review, then a monthly report, new test cases from real conversations, and one release a month. Price it from the hours per client per month plus your platform cost, then add margin.
An illustrative example, with your own numbers in place of these: 10 hours of setup and acceptance at your hourly rate, plus 3 hours a month of review at the same rate. If you run the assistant in your own Pro workspace, add the $100 a month platform cost to the monthly line, then your margin. This is arithmetic, not a market rate.
Your platform cost depends on the ownership model. If the client owns the workspace, they pay Rylvo directly and your fee is pure service. If you run assistants in your own Pro workspace, $100 a month covers up to 25 agents and about 2,650 included replies across all of them; extra replies are a one-time pack of 1,000 for $5. Prices checked on 2026-11-17; current plans are on the pricing page.
We found no public benchmark rate for chatbot QA we would stand behind as of November 2026, so anchor on your hours and the value of a lead to the client.
If you are comparing platforms for agency work, our Stammer AI alternative page and Rylvo for AI agencies set out the differences. If your clients need a fully branded client portal today, a tool built around that is a better fit than Rylvo.
Frequently asked questions
How do I prove a chatbot is ready for a client?
With evidence, not a demo: a signed scope sheet, a test report for the exact version going live, proof that a real lead reached the client's inbox, and a release record you can roll back to. The full pack has six documents.
What should a chatbot test report include?
The version tested, the date, the number of test cases by kind (common questions, refusals, hand-offs, lead capture, privacy), the pass rate, every failure with what was done about it, the judge model used, and who signed it off. One page is enough if every number links to the run behind it.
How much should I charge for chatbot QA?
Price it as two lines: a one-time setup and acceptance fee that covers the tests, red-team run and handover pack, and a monthly review retainer that covers the first-week review and a monthly report. Build the price from your hours and your platform cost, then add margin. There is no reliable public benchmark rate, so do not anchor on one.
Who owns the chatbot after handover?
Whoever owns the account it runs in, unless your contract says otherwise. Decide it before you build: either the client owns the workspace and invites your team as members, you build it and transfer ownership to the client, or you run it in your own workspace as a managed service. Write down who owns the prompts, knowledge and conversation data, and how they leave.
