Companies Using AI for Customer Service: 7 Setups a Small Firm Can Copy

Klarna, Lyft, Intercom's own customers and others use AI for customer service, and the pattern a small firm can copy is narrow: let AI answer the repeated, factual questions from your own policy text, send everything else to a person, and measure resolutions yourself. Klarna reported that its assistant handled two-thirds of chats within a month, in February 2024.

Each setup below comes with a date and a source, because big-company numbers are easy to repeat and hard to check. Several are vendor-published claims, and one company later changed course. After each setup there is a small-firm version with steps and sums. The examples for translation are a yoga studio, a removals firm, a tutoring centre and a landscaping company.

Follow me on Instagram@sagnikteaches

Setup 1: an assistant on the front line (Klarna, February 2024)

On 27 February 2024 Klarna announced that its AI assistant, built with OpenAI, had been live for a month. The company's press release reported 2.3 million conversations, two-thirds of its customer service chats, and said customers resolved errands in under two minutes against eleven previously. It also said repeat enquiries fell by 25% and that customers could still reach live agents.

Connect on LinkedInSagnik Bhattacharya

What the setup actually was: a chat assistant inside the company's own app, covering a defined list of topics (refunds, returns, payment problems, cancellations, disputes, invoice errors), with a route to a human always available. The scale is not copyable. The shape is.

Subscribe on YouTube@sagnikteaches

The small-firm version

1. List topics, not features. Export a month of enquiries and tag them. For a yoga studio, a realistic tally of 160 messages: class times and availability 48, cancellations and credits 37, pricing and passes 29, parking and what to bring 18, complaints 9, other 19.
2. Pick the factual ones. Times, passes, what to bring and parking are 95 of 160 messages (59%) and have one right answer. Cancellations depend on rules and customer history, so they come second.
3. Write the source text. One page per topic, in your words, that the assistant may use. Nothing outside it.
4. Keep the human door open. Every chat ends with "Want a person? Say so", as the Klarna release did.

The sums: if the assistant fully resolves, say, 60% of the 95 factual messages, that is 57 chats a month. At 3 minutes each, you save about 2.9 hours, or roughly $63 at a loaded $22 an hour. A tool that costs $60 a month is break-even at best, which is why the small-firm version starts with a free or cheap drafting step rather than an automated chat. The tutorial on what to automate first in small-business support goes through that ordering in more depth.

Setup 2: the same company stepping back (Klarna, May 2025)

This is the setup to read next, because it is the honest sequel. In a Bloomberg interview published on 8 May 2025, Klarna's chief executive said cost had been "a too predominant evaluation factor" and that "what you end up having is lower quality". As Entrepreneur reported the next day, the company began recruiting customer service staff again, in fully remote roles, so that customers could always reach a person. Bloomberg's report added that Klarna remained enthusiastic about AI.

The details are less dramatic than the headlines: the same report described a pilot of remote agents working in "an Uber type of setup", meant eventually to replace outsourced agents, while natural attrition kept shrinking the overall workforce. The reliable part is the lesson, which the chief executive stated himself: judging AI support mainly on cost leads to a drop in quality.

The small-firm version: add a quality measure beside the cost one

Track three numbers every week, not one:

• Cost per resolved enquiry (tool fees plus staff minutes).
• Re-contact rate: the share of customers who write again about the same thing within seven days.
• Escalation quality: out of ten chats passed to a person, how many arrived with the full context.

A removals firm using AI to answer "what do you charge for stairs?" might see cost per enquiry fall from $2.20 to $0.90, while re-contacts climb from 8% to 19% because the answer omitted the 3-floor limit. The cheap metric looked excellent. The second one shows the cost has moved onto the customer. The sampling routine in spot-checking AI support replies gives a weekly method for catching this.

Setup 3: an AI assistant tuned to a brand voice (Lyft)

Lyft's case study on Anthropic's site describes support in 2023 with wait times of 30 to 40 minutes and agents juggling several customers. Lyft says it tested several models on accuracy and on tone, and chose Claude partly for its personality. The page reports an assistant that reduced support resolution time by over 87% and improved decision-making accuracy by over 30%. It is a vendor-published case study on an undated page, which I read in November 2026, and those figures are the company's claims, not an audit.

The useful detail is how it rolled out: Lyft started with driver support, a narrower and more rule-driven area, and expanded to rider questions afterwards. It treated tone as a requirement to test, not an afterthought.

The small-firm version: test tone on twenty of your own replies

1. Collect twenty past replies you were proud of. A tutoring centre might choose its parent replies about missed sessions, fee queries and exam dates.
2. Paste five into the assistant as examples of voice and ask it to describe the style in three sentences.
3. Give it the other fifteen enquiries without the replies, ask for drafts using that style description and the policy text, and compare.

Here is how we write to parents: short, warm, no exclamation marks, always
name the child, always say what happens next. Use only the policy text
below. Draft a reply to this message. If the policy does not cover it,
reply with "NEEDS A PERSON" and say why in one line.

Policy: [paste]
Message: "Hi, [first name] missed Tuesday because of illness. Can we move it?"

An illustrative good output names the child as "[first name]", states the make-up rule from your policy (for example, one free make-up within fourteen days) and says who will confirm the new slot. A bad one invents a rule. Count how many of fifteen are usable without change. If it is 11 or more, you have a voice worth automating partly; if it is 6, fix the policy text first.

Setup 4: a 50% resolution claim, priced for a small queue (Robin)

Intercom's pricing page carries a quotation from the director of technical support at Robin, a workplace operations software company: "We're seeing a 50% resolution rate with Fin". On the same page, the chief digital officer of solidcore, a fitness brand with many studios, says it is saving "hundreds of thousands of dollars" with Fin as it opens more studios. Both are vendor-page testimonials that I read on 5 November 2026 (the page rotates its quotes), and what they mean depends on how Intercom counts a resolution.

That definition matters. Intercom prices Fin at $0.99 per outcome. Its help centre counts a confirmed resolution when the customer says the answer worked, and an "assumed resolution" when the customer leaves the conversation without asking for more help after Fin's last answer. Silence is not the same as satisfaction. The same $0.99 also applies when Fin completes a procedure you have set up to end in a hand-off to a person, so some chats a human finishes are billed too. Zendesk's help centre describes a similar time-based window plus an AI check before a conversation counts as automatically resolved; read its current article for the figures.

The small-firm version: price it with your own volume

Suppose the yoga studio gets 200 conversations a month and the tool really resolves half, or 100. At $0.99 each that is $99 a month in per-resolution fees. If a seat for the person handling the rest costs $29 a month (Intercom's Essential list price on annual billing; new customers were offered $19 when I checked), the total is $128. The time it displaces: 100 chats at 3 minutes is 5 hours, worth $110 at $22 an hour. So the tool costs more than the time it saves unless the chats are longer, the staff cost more, or the saved time earns bookings. At 400 conversations the sums change. If some hand-offs run through a billed procedure, add those to the fee column as well.

Notice what the table says: at short chat lengths the fee eats the saving, and the value comes from chat length and volume. Measure your own average handling time before choosing a per-resolution product. The cost-per-agent breakdown compares seat-based alternatives.

Setup 5: resolution time cut from days to hours (Synthesia)

Another quotation on Intercom's pricing page, from the vice president of customer support at Synthesia, an AI video platform for business, says that since launching Fin the resolution time went from five days and five hours to four hours and 37 minutes, which the page presents as a 96% decrease. Again a vendor-page testimonial, read 5 November 2026, with the vendor's own definition of resolution time.

The notable thing is the measure chosen: time to resolution, which is what a customer feels. A five-day wait for a support answer turns into churn.

The small-firm version: fix the waiting, not just the answering

For a landscaping company, "support" is quote requests and rebooking after rain. The pain is slow replies, not volume. A cheap fix is an auto-acknowledgement with the next step, plus an AI-drafted answer for the owner to approve at set times.

1. Measure the current gap: how long between a customer's message and your first useful reply? Sample twenty messages. Say the median is 19 hours.
2. Draft replies with AI for the three most common messages: "Can you quote?", "Can you come earlier?", "Was my visit cancelled by weather?".
3. Approve and send in two batches a day, at 9:00 and 15:00.
4. Re-measure after two weeks. Aim at a median under 6 hours.

This setup copies the outcome the Synthesia quote points at, using only an assistant and a template, with nothing automatic.

Setup 6: one busy workflow automated, with a human sign-off (Nuuly)

Nuuly is a clothing rental subscription in the URBN family of brands, alongside Urban Outfitters, Anthropologie and Free People. A customer story posted by Intercom's community team on 10 September 2026 reports that Fin, Intercom's AI agent, fully resolves 35% of the roughly 80,000 conversations Nuuly receives each month, that customer satisfaction held at 95% while subscribers grew 40% in a year, and that Fin saved $359,000 in salary costs over six months. The longer version on Intercom's Fin site, which I read on 5 November 2026, adds that the support team has stayed below 80 people against a five-year plan for 120. These are vendor-published claims, and neither page says what Fin itself cost, so the $359,000 is a salary figure, not a net saving.

What the setup actually was: Fin answers the simple, frequent questions from a knowledge base the team wrote, and it is connected to Nuuly's own systems so it can look up a subscriber's billing date, order status and account details. The busiest workflow, pausing, cancelling and refunding subscriptions, used to mean cross-checking several systems by hand. Fin now pulls the live subscription and order data, verifies a return against carrier tracking and applies the change. For more complex cases it gathers what an associate needs, and Nuuly keeps the sensitive final step with a person. Natalie Hurst, its senior director of customer success, says the team was reorganised "to allow for more career growth and more pay", and senior associates have moved into retention work.

Two details matter more than the headline figures. The resolution rate is 35%, well below the numbers most vendor pages lead with, and the story still credits it with keeping the team far smaller than planned. And the automation was built around one high-volume workflow, checked against real account data, with a person kept on the step that touches money, rather than around a general chatbot.

The small-firm version: one checked action, approved by a person

The yoga studio from setup 1 had 37 cancellation and credit messages in its month of 160. Suppose 20 of those are members asking to pause a monthly membership for a holiday or an injury. Each one currently takes about 8 minutes: find the member in the booking system, check whether they have paused before, reply, then change the billing.

1. Write the rule down. For example: monthly members may pause for up to four weeks, twice a year, with seven days' notice. If the rule exists only in the owner's head, the assistant cannot follow it and neither can a new member of staff.
2. Let the assistant gather and check, not apply. It collects the name and dates, compares them with the rule and prepares a one-line summary for the owner, who checks the pause history in the booking system.
3. Keep the billing change with a person. The owner approves it in the booking software, which is the small-firm equivalent of the sensitive final step Nuuly keeps.
4. Name the job the saved time goes to before you start. Nuuly moved senior associates into retention work. The studio's version is a short call or message to members who have not booked for three weeks.

When a member asks to pause their membership:
1. Ask for their full name and the dates they want to pause.
2. Check the policy: monthly members may pause for up to 4 weeks,
   twice a year, with 7 days' notice.
3. Do not apply the pause or promise it. Reply: "Thanks, I've passed
   this to the studio to confirm today."
4. Send the owner one line: name, dates, and whether the dates fit
   the policy.

The sums: 20 requests at 8 minutes is 160 minutes a month. If the owner's part shrinks to 2 minutes of checking and approving each, that is 40 minutes, a saving of 2 hours, or about $44 at $22 an hour. For a queue this size the saving is small. The larger gain is what the two hours buy: if the retention calls keep even one $90-a-month member who was drifting away, they are worth more than the time saved. Track both, as in setup 2. For setup and rollout in the first month, the first 30 days of AI support tutorial has a day-by-day plan.

Setup 7: an airline's chatbot promise it could not keep (2024)

The last setup is a warning. A tribunal decision dated 14 February 2024 found an airline responsible for what its website chatbot told a passenger about bereavement fares. The chatbot said the reduced fare could be claimed after travel, within 90 days; the airline's own policy page said refunds were not allowed for travel already taken. The airline argued it could not be held liable for information given by its agents or representatives, including a chatbot. The tribunal rejected that, found the airline had not taken reasonable care to ensure its chatbot was accurate, and awarded the passenger the fare difference plus interest and fees. This news report covers the decision. It came from an online civil-dispute tribunal, so it is not a ruling that binds other courts.

What generalises: whatever your assistant says in your name can be treated as your statement. The set-up failure was probably two sources of truth, a chatbot and a policy page that disagreed.

The small-firm version: one source of truth and a promise filter

1. Keep one policy document and have the assistant read only that. When the policy changes, change the document the same day.
2. Forbid money promises. The instruction "Never offer refunds, discounts, credits or exceptions; say a person will decide" removes the most expensive error class.
3. Date-stamp the policy and test the assistant with five tricky questions each time you update it.

Rules for the assistant:
1. Answer only from the policy text below. If the answer is not there,
   say "I'll ask a person to confirm" and stop.
2. Never promise a refund, credit, discount, price, date or exception.
3. If the customer mentions a complaint, injury, legal issue or says
   "cancel", pass to a person with a one-line summary.
4. Say you are an automated assistant in your first message.

A removals firm would test it with: "Can I get a refund if you damage my sofa?" The right reply is not a yes or a no, but "That needs a person, here is what to send them". If the assistant answers with a policy it made up, the policy text is missing a clause, and you now know before a customer does.

A mistake that came from copying the headline number

A plausible small-firm story. A tutoring centre reads that a company automates two-thirds of chats and buys a chatbot expecting the same. In its first month, 22% of conversations end without a human. The owner reads that as a failure and almost cancels. Looking closer, the centre's messages are mostly about individual students, which a bot cannot answer from general policy. Klarna's mix was refunds, returns and payment status, which suit automation.

The lesson: automation rate is a property of your enquiry mix as much as of the tool. Tag a month of messages before buying anything, so you know your ceiling. A mix of 60% individual-student questions puts a ceiling near 40%, and 22% may be a good result.

Five questions to put to any case study before you copy it

Every setup above comes from a company talking about itself, so use the same filter on the next one you read, whatever the source.

1. What does "resolved" mean? A customer leaving without asking for more help, or a customer saying thanks, are different measures. The first can be inflated by people who simply gave up.
2. What is the date? Klarna's two statements are fifteen months apart and point in different directions. A good result in month one says little about month fifteen.
3. What was the enquiry mix? Refund and payment-status questions automate well. Advice, disputes and anything involving a particular person do not.
4. Who sells the thing? A vendor page is marketing by design. That does not make it false, but it means failures are absent.
5. What is your own number? Replace every percentage with the figure from your last 100 messages. If you cannot, you do not yet know whether it applies.

Run a quick example. A study says a tool resolved 70% of chats. You count 100 of your own and find 35 are repeats of fixed-answer questions. Your ceiling is nearer 35% before any quality loss, so a quote of $0.99 per resolution implies about $35 a month at 100 chats, not $70. That figure, not the headline, is what you compare with the staff minutes saved.

Choosing which setup to copy first

Most small firms should start at the top of that table and move down only when the numbers justify it. If you are choosing between building something and buying a product, the piece on build or buy for an AI chatbot covers that decision, and the tutorial on AI against outsourced customer service compares the other alternative.

How to check it is working after a month

1. Pull 30 conversations at random, ten of them AI-handled, ten passed to a person, ten untouched by AI.
2. Score each as correct, partly right or wrong, using your policy as the answer key.
3. Compare the re-contact rate with the month before.
4. Read every complaint and every message containing the word "refund", "cancel" or "wrong".

If wrong answers exceed one in ten, tighten the policy text and the rules, then re-test. If the re-contact rate rose, the setup is saving your time by moving work to customers, which is not a saving. For a firm that wants a second pair of eyes on which of these setups fits its inbox, that is part of what my AI implementation consultation covers, and the contact page is the way in.

Questions about copying big-company AI support

Are vendor case-study numbers reliable?

Treat them as claims, not measurements. A vendor chooses which customer to feature and how to define a resolution. Look for the definition, the date and the sample size, then test your own number on a few weeks of real conversations before you rely on anything quoted from a brochure.

How many chats do I need before AI support pays off?

There is no fixed threshold, but the sums get thin below about 100 conversations a month, because a per-resolution fee or monthly seat can exceed the time saved. At several hundred a month, with repeated questions, the sums usually become worth testing. Measure your own repeat-question share first.

Can I be held responsible for what my chatbot says?

A tribunal decision dated 14 February 2024 held an airline responsible for what its website chatbot told a customer, rejecting its argument that it could not be liable for information from a chatbot. It came from an online civil-dispute tribunal, not a court whose rulings bind others, but it is a good reason to keep the bot's sources current and limited.

Should customers be told it is AI?

Say so plainly at the start of a chat and offer a human route. Some jurisdictions require disclosure, and it is also good practice: customers forgive a bot that admits its limits much more readily than one that pretends to be a person.

What is the cheapest setup to copy first?

An assistant that drafts replies for a person to send, using your own policy text and past answers. It needs no integration, carries little risk because nothing sends without a read, and shows you which questions repeat. Move to automatic answers only for the questions where the drafts were nearly always right.

Further reads

• How to Hire Your First Customer Service Person Alongside AI — When the right move is a person working beside the tool.
• WhatsApp Customer Service With AI: Setup, Costs, and Limits — Setup, costs and limits of AI replies on WhatsApp.
• Customer Service QA Software With AI: What to Compare — Tools for scoring the quality of support replies.
• Best AI Chatbots for Small Business Websites — Chat tools for a small firm's website, compared.
• AI Knowledge Base Options for Small Businesses Compared — Build the single source of truth the assistant reads.
• AI KPIs for Small Businesses: 12 Metrics Worth Tracking — A short list of measures to track after launch.
• Avoca AI for ServiceTitan Shops: What It Does, Who It Fits — Assess Avoca's ServiceTitan booking, AI CSR and lead follow-up against a small shop's actual missed-call problem.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Klarna press release of 27 February 2024; Bloomberg interview of 8 May 2025 (syndicated text) and Entrepreneur's report of 9 May 2025; Fox Business report on the tribunal decision of 14 February 2024; Lyft case study on Claude's customer pages; Intercom pricing page quotes and Fin outcomes help article (read 5 November 2026); Intercom community customer story on Nuuly (10 September 2026) and Nuuly's customer story on Intercom's Fin site (read 5 November 2026).