Agentrolux
Example · Customer service24/7 resolution agent

Answer every customer, even at midnight, and let your team handle what matters most.

An AI agent that answers the routine questions from your approved information, takes the small actions you allow, and hands anything about money, complaints or safety to a person.

See a sample conversation Talk about your business
Ten simulated conversationsAnswers only from your sourcesA person, whenever asked

Overnight inbox · 07:00

Simulated example

  • WhatsApp · “Where is order 1042?”

    23:41 · Answered from the order system

    Answered

  • Chat · “Can I move to Friday 11am?”

    00:07 · Rescheduled, logged

    Answered

  • Chat · Refund on a ripped sofa

    02:14 · Over the £50 limit · with a person

    Person

  • Email · Ingredient question

    03:02 · No approved source · with a person

    Person

2 answered · 2 with a person · nothing left waiting

Every message gets a reply before morning

01 — The problem

Three small failures that cost customers.

None of them look dramatic on a quiet Tuesday. Together they decide whether someone comes back.

01

The reply arrives tomorrow

A customer asks at 9pm whether the item arrives before the weekend. They buy elsewhere at 9:20. You answer at 8am.

See what slow service costs
02

The same questions, every day

Opening hours, returns, delivery, prices. Your team answers the same six questions dozens of times, and the complex cases wait behind them.

See the repeat-contact benchmark
03

The morning backlog

The queue is sorted by arrival time, not urgency. A refund request and a question about a birthday gift sit in the same pile.

See the checkpoints
02 — How the agent works

Seven stages, two human checkpoints.

Select any box for a plain-language explanation and an example. The dashed boxes are where a person takes over.

Human checkpoint

Check handover rules

Before anything final is said, it checks the stop list. Refunds above the limit, complaints, legal, medical and safety messages go to a person. So does any request for a person.

A £180 refund stops here and goes to a person, with no promise made.

See a conversation
Text version of this flow

Four channels — website chat, email, WhatsApp and the social inbox — feed one agent. The agent works through seven stages, and two human checkpoints sit on the branch to a person.

  1. 1. Receive message. Any channel, any hour. The customer writes on the channel they already use. The agent notes the time, the language and the conversation so far.
  2. 2. Identify need. Order, booking, product, policy, refund, complaint. It works out what the customer actually wants before it writes anything. If the need is unclear, it asks one short question.
  3. 3. Check approved sources. Help centre, policies, catalogue, order system. It looks for the answer only in the information the owner approved. It never fills a gap from general knowledge.
  4. 4. Answer or act. Reply, reschedule, resend, update, small refund. It replies in the customer's language with the next step, or takes one of the small actions the owner has pre-approved.
  5. 5. Check handover rules. Money, complaints, safety, a request for a person. Before anything final is said, it checks the stop list. Refunds above the limit, complaints, legal, medical and safety messages go to a person. So does any request for a person.
  6. 6. Confirm resolution. Did that actually answer the question?. It checks whether the customer has what they need. If not, it offers a person and hands over instead of trying again.
  7. 7. Log conversation. Sources, actions, handover reason. Every conversation records which sources were used, what action was taken, and why a person took over, so the owner can audit it and improve the help pages.
  8. Human checkpoint: a person reviews money. Refunds over £50, compensation, complaints. Money decisions stay with a person. The agent can approve a small refund inside the written limit; anything above it stops and waits.
  9. Human checkpoint: a person takes over. Any request, safety, legal, or no approved answer. A person can join any conversation, and the customer can ask for one at any time. The request is honoured immediately, without trying to keep them in the chat.

Outputs: a reply in the customer's language, a booking change inside the allowed hours, a handover summary for the person, and an audit log of every conversation.

The diagram matches the specification and the evaluation set. On a narrow screen the text version above carries the same information.

Check handover rules: Before anything final is said, it checks the stop list. Refunds above the limit, complaints, legal, medical and safety messages go to a person. So does any request for a person. Example: A £180 refund stops here and goes to a person, with no promise made.
01

Receive the message

Website chat, email, WhatsApp or the social inbox. The time, language and earlier messages come with it, so nobody starts from zero.

02

Identify the need

Order status, booking change, product question, policy, return, complaint — or something else. If it is unclear, it asks one short question.

03

Check the approved sources

Your help centre, policies, price list, catalogue, and read-only order or booking records. No source means no answer.

04

Answer or act

A reply in the customer's language with the next step, or one of the small actions you allowed: reschedule, resend, update an address before dispatch, refund up to your limit.

05

Check the handover rules

Money above the limit, complaints, legal, medical, safety, or any request for a person. It stops, writes a summary, and passes the thread on.

06

Confirm the answer landed

If the customer is still stuck, it offers a person rather than repeating itself. A correct handover counts as a success.

07

Log the conversation

The sources used, the action taken, and the reason for any handover. That record is how the help pages get better.

03 — See it run

Ten conversations, step by step.

A made-up shop, clearly labelled. Every reply, action and handover comes from the same rules described on this page. The AI disclosure appears before the first message, every time.

Ten conversations

Simulated example

No real customer data. The business, orders and names are invented for this page.

Where is my order?

WhatsApp · 23:41

Simulated example

A customer asks after hours about an order that shipped yesterday.

Before the first message · AI disclosure

You are talking to an AI assistant, not a person. I can pass you to someone at any time — just ask.

Customer

Hi, where is my order 1042?

1 message to go

Step 1 of 2

04 — You stay in control

The agent answers from your information. Money and safety go to your team.

Automation is useful only if you can see it, limit it and stop it. These are the parts a person keeps, in this workflow.

The agent can approve a refund up to your written limit. Above it, it cannot act — and it never promises a refund, a discount or compensation on its own authority.

  • You set the sources

    Which help pages, policies and price lists the agent may use. A page past its review date is treated as unavailable.

  • You set the limits

    The refund ceiling, the booking window, and whether an address can still be changed. Enforced in code, not in a prompt.

  • You can take over

    A person can join any conversation, and the customer can ask for one at any point. The agent stops immediately.

  • You can pause it

    For one channel or all of them, or restrict it to after-hours only. Messages wait for a person instead.

05 — Documented failures

What happens when service is left to chance.

Two public company cases and two labelled industry findings. Each one says what this workflow would do differently — not that it would have saved the company.

Automation done badly

The agent was trusted with more than it could support, or its owner stopped watching the quality.

01 · 2024

Automation done badly

Air Canada: The chatbot gave the wrong fare rule

CAD 812.02

awarded, including interest and fees

A customer asked Air Canada's website chatbot about a bereavement fare. It told him he could apply after travelling. The airline's real policy did not allow that, and it refused the claim. The British Columbia Civil Resolution Tribunal held the airline responsible for what its chatbot said.

Decision 2024 BCCRT 149, 14 Feb 2024 CanLII copy

An agent with this workflow could only quote the published fare policy, and would hand over rather than improvise a rule. The tribunal's point stands either way: the business owns what its agent says.

02 · 2024–2025

Reported, not measured

Klarna: Volume first, quality second

2.3M

conversations in the assistant's first month

Klarna said its AI assistant handled 2.3 million conversations and the equivalent work of 700 agents. In May 2025 its CEO said making cost the dominant measure had hurt service quality, and the company was recruiting human agents again. The quality decline is reported by the CEO and the press, not an audited measurement.

Klarna, 27 Feb 2024 CNBC, 14 May 2025

This workflow is designed for the opposite order: a correct answer or a person first, volume second. The handover rate is reported openly as a signal to improve, not hidden.

No automation, or self-service that stalls

Documented findings, not company cases. They show the cost of leaving customers to finish the job alone.

01 · 2024

Survey finding, not a company case

Self-service, across industries: Answers are easy to find and hard to finish

14%

of issues fully resolved in self-service

In Gartner's survey of 5,728 customers, only 14% of service issues were fully resolved in self-service, and only 36% even for issues customers called very simple. This is an industry finding, not a company case: it shows what happens when a business hides the last step behind a help page.

Gartner, 19 Aug 2024

The agent answers the question and then checks whether it is actually solved. If not, it offers a person and hands over with the thread, instead of sending the customer back to the same page.

02 · 2025

Benchmark, not a company case

Call centres, benchmark: Three in ten issues need a second contact

69%

first-contact resolution across 500+ call centres

SQM's 2024 benchmark of more than 500 North American inbound call centres put first-contact resolution at 69%. That means about three in ten customers had to come back. This is a benchmark, not a company case, and it measures phone calls rather than chat or email.

SQM Group, 2024 FCR benchmark

The agent closes routine questions on the first reply, and every conversation that cannot be closed is handed over with the context — rather than ending in “we'll get back to you”.

Each company case is a public report of what happened, not a claim that this workflow would have prevented it. Causes were mixed, and the sources say so. The last two cards are a survey finding and a benchmark, and are labelled as such.

06 — Facts about customer service

The bigger picture, with the source attached.

Every number links to its publisher, shows the sample or method, and names who paid for the research. We do not quote a figure we have not opened.

01
Source

MarketsandMarkets, AI for Customer Service Market

2025

Market forecast: USD 12.06B in 2024 to USD 47.82B by 2030, 25.8% annual growth. North America forecast to lead. Covers chatbots, agent assist, analytics and platform software.

Market revenue, not adoption. A forecast from a research firm. The release body reverses the two dollar values in one sentence; we quote the headline and report listing.

Open the source

$12.06B

The AI customer service market in 2024

A research-firm estimate, forecast to reach $47.82B by 2030 at 25.8% a year. Market revenue, not adoption by small businesses.

MarketsandMarkets · 2025
02
Source

Brynjolfsson, Li & Raymond, NBER Working Paper 31161

2023

Staggered rollout at one Fortune 500 software company: 5,179 customer-support agents. Access to a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled agents.

One company, one 2023-era tool. It assisted human agents; it did not replace them. Not a small-business study.

Open the source

14%

more issues resolved per hour with an AI assistant

In a study of 5,179 support agents at one large software company. Gains were largest for newer agents: 34%.

Brynjolfsson · 2023
03
Source

Gartner, customers and AI in customer service

2024

Survey of 5,728 customers, fieldwork December 2023. 64% would prefer that companies did not use AI in customer service.

Gartner does not publish the exact question or the sampling frame. It measures preference, not behaviour.

Open the source

64%

would prefer companies did not use AI for service

Gartner's survey of 5,728 customers. Preference, not refusal — which is why a person must always be one message away.

Gartner · 2024
04
Source

Gartner, agentic AI forecast

2025

Forecast: by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention, with a 30% reduction in operational costs.

A forecast, not a measured result. No method or sample is published. It does not describe this workflow.

Open the source

80%

of common issues resolved without a person, by 2029

Gartner's forecast, not a measured result, and not a description of this workflow. Treat it as a direction, not a target you have to hit.

Gartner · 2025
05
Source

SQM Group, Call Center FCR Benchmark 2024

2025

More than 500 North American inbound call centres. First-contact resolution measured with two standard customer-survey questions within one business day of the call; at least 400 surveys per centre. Aggregated FCR: 69%.

Call centres, not chat or email. “Resolved” is a survey answer.

Open the source

69%

of call-centre issues were resolved on the first contact

SQM's benchmark of 500+ North American call centres. About three in ten customers had to come back.

SQM Group · 2025
06
Source

HubSpot, What's Customer Responsiveness?

2025

Consumer survey: 66% expect a response from customer service in five minutes or less.

Vendor survey with no published fieldwork date, sample size or method. Directional only.

Open the source

66%

expect a reply within five minutes

A vendor survey, not a universal standard. It is still a useful reminder that “within one working day” is not what many customers expect.

HubSpot · 2025
07
Source

Salesforce, self-service customer service

2024

61% of customers would rather use self-service for simple issues.

Vendor-reported research; no sample size, fieldwork dates or method published.

Open the source

61%

would rather use self-service for simple issues

Salesforce's figure, with no published method. It supports a help centre — as long as a person is easy to reach when the page fails.

Salesforce · 2024
08
Source

Gartner, self-service resolution

2024

Survey of 5,728 customers, December 2023. Only 14% of issues fully resolved in self-service; 36% even for issues customers described as very simple.

About self-service as it exists, not about AI agents.

Open the source

14%

of issues are fully resolved in self-service

Same direction, different lesson: finding an answer is not finishing the job. The agent checks, then hands over if the customer is still stuck.

Gartner · 2024

These findings come from other businesses and other researchers. They do not promise what will happen in yours. The four-week pilot measures your own baseline before the agent answers a single message.

07 — Numbers that help you plan

Three numbers worth knowing before a call.

Each comes from a published source, with the sample beside it. Where no reliable average exists for your business, we ask you instead of inventing one.

01

69%

of call-centre issues were resolved on the first contact

About three in ten customers had to come back. A workflow that closes routine questions on the first reply is not chasing a perfect score — it is removing the repeat contact.

SQM Group · 2025
Source

SQM Group, Call Center FCR Benchmark 2024

2025

More than 500 North American inbound call centres. First-contact resolution measured with two standard customer-survey questions within one business day of the call; at least 400 surveys per centre.

Phone calls, not chat or email. “Resolved” is a customer's answer, not an operational check.

Open the source
02

66%

of consumers expect a reply within five minutes

Most small businesses cannot staff that expectation at 11pm. The honest fix is a correct immediate answer for routine questions, and a person with the context for the rest.

HubSpot · 2025
Source

HubSpot, What's Customer Responsiveness?

2025

Consumer survey. 66% say they expect a response from customer service in five minutes or less.

Vendor survey with no published fieldwork date, sample size or method. Directional only.

Open the source
03

14%

of issues are fully resolved in self-service

A help page answers a question; it rarely finishes the job. The agent must check whether the customer is actually done, and hand over when they are not.

Gartner · 2024
Source

Gartner, self-service resolution survey

2024

Survey of 5,728 customers, fieldwork December 2023. 14% of service issues fully resolved in self-service; 36% even for issues described as very simple.

About self-service as it exists, not about AI agents. Gartner does not publish the sampling frame.

Open the source

No reliable public figure exists for the share of your messages that arrive out of hours, so we do not print one. That number is yours to enter.

Use your own numbers
08 — Before and after

The same week, two different jobs.

Drag the handle to compare a week of catching up with a week of handling exceptions. This is an illustration, not a measurement of time saved.

BeforeAfter
Illustrative week

A week of catching up

  1. Mon 08:00Forty-one messages waited overnight. The team starts with the oldest, not the most urgent.
  2. Mon 11:20The same three questions — opening hours, returns, delivery — for the ninth time this week.
  3. TueTwo customers who asked on Saturday have already bought elsewhere.
  4. ThuA refund request sits in a shared inbox because nobody is sure who can approve it.
  5. FriThe owner spends the afternoon catching up on chat instead of on the shop.

With the resolution agent

  1. Mon 02:14A WhatsApp question about an order is answered with the real tracking link.
  2. Mon 08:00Nine conversations need a person: two refunds over the limit, one complaint, six no-answer questions.
  3. Mon 08:10The team reads the summaries, not the whole thread, and starts with the complaint.
  4. TueThe weekend's routine messages are already closed. Nobody starts the week behind.
  5. FriThe owner checks the week's audit log for ten minutes: sources used, actions taken, reasons for handover.

Drag the handle, or use the arrow keys. This is an illustration of how the week changes, not a measurement of saved time — the pilot measures your own baseline.

09 — Your version

Four questions about your own week.

Keep your answers. They are the brief for the first version we would build for you.

  1. How many customer messages arrive in a normal week, and how many arrive outside working hours?

  2. Which questions could be answered word for word from your own help pages or policies?

  3. Which decisions must always reach a person: refunds, complaints, or anything about safety?

  4. What is your usual reply time today, and what would good look like?

Next

Put your own figures into the calculator below. It runs in your browser, and nothing is sent or stored.

Put your week into the calculator

Bring the four answers to the free one-hour call. We will use them to sketch the first version, and you keep the written plan either way.

Talk about your business
10 — Plan the numbers

Move the sliders. See the shape of the saving.

A planning tool, not a quote. It shows how the arithmetic works and which inputs matter most. Every formula is printed below the result.

260

Count every inbound message, including the ones a person answers.

120

Customer questions only, not invoices or supplier mail.

140

Inbound customer messages in the business number.

30

Direct messages plus public comments that need a reply.

62%

Opening hours, order status, prices, policy, booking changes. Your estimate is fine.

4 min

Reading, checking, replying. Include the time spent looking things up.

£22

Pay plus employer costs. An estimate is enough.

34%

Evenings, weekends and holidays.

9 h

Hours. Use 12 if most messages wait until the next morning.

Planning estimate

Not a forecast

Staff hours back each month

59.1–83.7

about £1,299–£1,841 of staff time

Out-of-hours messages answered the same day

648–769 a month

of 810 messages arriving outside opening hours

First reply time

4.3 h–5.7 h

from 9.0 h today, for 37–53% faster

An estimate from your inputs and the assumptions below — not a quote, a promise or a forecast. Nothing is stored or sent.

How we calculated this
How we calculated this — every formula and assumption

Messages and hours

260 + 120 + 140 + 30 = 550 messages a week × 4.33 weeks = 2382 a month.

62% routine = 1477 routine messages. At 4 minutes each, the agent handles 60–85% without a person: 59.1–83.7 hours a month. That is our assumption, not a measured benchmark.

Staff cost: hours × £22 = £1,299–£1,841. It counts only the minutes per message, not hiring, tools or setup.

Reply time

Today’s first reply is 9.0 h. If the agent closes 60–85% of routine messages with a one-minute reply, the average becomes 4.3 h–5.7 h. The rest still wait for a person, so the time never reaches zero.

After hours

2382 messages × 34% = 810 outside opening hours. We assume 80–95% of those are answered the same day: 648–769. A person may still need to finish the job the next morning.

What this is not

Not a promise of savings, not a quote and not a forecast. It assumes customers accept an AI answer for routine questions; some will ask for a person, and the calculator still counts their message as handled time saved. The four-week pilot measures the real numbers before anything changes.

Inputs held where?

In this browser tab only. Nothing is written to storage, sent to a server, or kept when the page is closed.

11 — Build this next

Other workflows we can build for a service team.

Start with the one that costs you the most. These use the same pattern: read, decide, draft, ask, report.

Returns and refund triage

Read the request, check the order and the policy, approve small refunds, and send a clear status.

Ask about this workflow

Order-status chat

Answer “where is my order?” with the real tracking status, and hand exceptions to a person.

Ask about this workflow

Review replies for approval

Draft replies to public reviews, flag the ones that need care, and wait for your yes.

Ask about this workflow

Complaint routing by urgency

Sort complaints by what went wrong and how long they have waited, then route each to the right person.

Ask about this workflow

WhatsApp appointment reminders

Send the reminder, handle the reschedule inside your rules, and tell you who cancelled.

Ask about this workflow

Call summaries in the customer record

Turn a recorded call into a short note, with the promises made and the next step.

Ask about this workflow

Dedicated pages for these workflows are not built yet. Each card starts the conversation with the right context.

11 — Questions businesses ask

Straight answers, no sales fog.

If something is missing, ask on the call. We will tell you honestly whether this is a fit.

Book a free 1-hour call
Will customers know they are talking to AI?

Yes. The disclosure is the first thing in every conversation: it says this is an AI assistant and that a person is available at any time. It is not hidden in a footer, and it is not repeated as a warning on every message.

What if the AI gives a wrong answer?

It can only answer from information you approved: your help pages, policies, price list and order system. If nothing matches, it does not guess — it offers a person and hands over. Every answer records the source it used, so a wrong answer can be traced to the page that needs fixing.

Can a customer always reach a person?

Yes. Asking for a person ends the agent's turn immediately, in any language and at any point. The person receives the full conversation and the customer's preferred contact method, so nobody has to repeat themselves.

Do I need to change my phone or chat system?

No. The agent connects to the tools you already use: your website chat, your email inbox, WhatsApp Business and your social inbox. If a connection does not exist, we start with email and one chat channel. Phone calls are not part of this first version.

Who can read the conversations?

Only the staff accounts you name. Conversations are encrypted, kept for a period you set (30 days by default), and deleted on request. They are not used to train a model, and the agent never asks for card details, passwords or identity documents.

How long does setup take?

Two to three weeks for a first version. Most of the first week is spent fixing the help content, because the agent can only be as accurate as the sources behind it. Then a four-week pilot runs alongside your normal process before anything is switched on fully.

What does it cost?

Every business is different, so there is no published price. We scope it on a free 1-hour call and give you a written plan you keep, whether or not you go ahead.

What happens to messages that arrive at 2am?

The agent answers what it can from your approved sources, takes the small actions you allowed, and hands anything else to a person with a summary. Your team starts the morning with the exceptions, not the backlog. You can also limit the agent to after-hours only.

12 — Start here

Tell us which questions your team answers most. We'll show you what the agent can handle first.

One hour, no slides. Bring last week’s inbox and the three questions you are tired of answering. You leave with a written plan you keep.

Book a free 1-hour call Email us instead

hello@agentrolux.com

13 — Sources

Every number on this page

Checked October 2026. Vendor-sponsored research is marked as such in the source panel next to each figure. Sample sizes and limits are recorded in the page notes.

For technical teams

Channels, sources, handover rules, guardrails and the full specification

Channel integrations

  • Website chat: embed or existing provider, same-origin only, disclosure before the first message.
  • Email: IMAP/SMTP or Microsoft 365 / Google Workspace APIs, one reply per inbound thread.
  • WhatsApp Business Platform: inbound webhooks, template messages outside the 24-hour window.
  • Social inboxes: Instagram and Facebook messaging APIs; public comments move to a private reply.
  • Helpdesk and CRM: Zendesk, Intercom, Freshdesk, HubSpot — read conversation, write note and handover.

Approved-source configuration

  • Each source has an ID, an owner, a public URL or internal reference, and a review date.
  • Answers cite the source ID; a conversation can be replayed against the source version used.
  • Expired sources are removed from retrieval automatically rather than answered from stale text.
  • No source accessible → handover with no_approved_answer.

Handover and guardrails

  • Handover triggers: customer request, safety, legal or medical, refund above £50, complaint, no source, disallowed action, six-message loop.
  • Refund ceiling and booking window are code constants, not prompt text; changing them is a reviewed deploy.
  • Actions are idempotent, logged before the customer is told, and never retried silently.
  • Prompt-injection attempts are refused and logged; instructions are never echoed.
  • Personal data is redacted on ingest where it is not needed; retention is per-tenant.

Audit log format

conversation_id, at, channel, direction,
intent, source_ids[], confidence,
action { type, before, after, result },
handover { reason, summary_id },
disclosure_shown, language, retention_expiry
Download the workflow specification (.md)

Includes the decision rules, action limits, handover triggers, guardrails, metrics and the 46-scenario evaluation set.

Figures checked October 2026. This page describes a simulated workflow, not a client result.

Sample limits shown with every figure