صياغة التميز في البرمجيات
دعنا نبني شيئاً استثنائياً معاً.
اعتمد على شركة Lasting Dynamics للحصول على جودة برمجيات لا مثيل لها.
ميشيل سيمينو
سبتمبر 30, 2026 • 10 min read
The two best-known experiments in AI customer service both ended in a public reversal, and neither failed because the software did not work. They failed on sequence: the headcount went first and the evidence arrived afterwards. One line of disclosure before the argument: we build custom AI software, including for customer teams, and most of this article argues that you should start with features you already pay for.
Klarna is the case most people half-remember. In February 2024 it said its assistant could do the work of 700 representatives, having taken over two thirds of its service chats. Fifteen months later its chief executive, Sebastian Siemiatkowski, told Bloomberg that cost had been too dominant a factor and the result was lower quality, adding: "Really investing in the quality of the human support is the way of the future for us". Klarna kept the assistant on the high-volume work and started recruiting people again for the harder conversations.
Commonwealth Bank of Australia made the same mistake in under two months. In July 2025 it said 45 customer service roles were no longer needed because an AI voice bot had cut call volumes, and by 21 August it had reversed the decision, admitting, in a statement reported by ABC News, that it "did not adequately consider all relevant business considerations". The Finance Sector Union's account explains why: call volumes were rising, management was offering overtime, and team leaders were being pulled onto the phones.
Gartner expects more of this, predicting in February 2026 that by 2027 half of the companies that attributed headcount reductions to AI will rehire staff for similar work under different job titles. Kathy Ross, a senior director analyst at Gartner, adds an uncomfortable caveat: "Most recent workforce reductions were influenced by broader economic conditions rather than automation alone". The firm's survey of 321 service leaders in October 2025 found that only 20% had actually reduced agent headcount because of AI, and 55% reported stable staffing while handling higher customer volumes.
Gartner's forecasts also cut the other way: it predicts that by 2029 agentic AI will resolve 80% of common customer service issues without human intervention, cutting operational costs by 30%. Both forecasts hold at once: the volume moves before the team does, and the order of those two events decides whether AI customer service leaves you with better service or an apology to write.
AI customer service is three jobs with very different levels of risk, and only one of them ever speaks to your customers. Separating them is most of the decision, because the job that gets written about is the one to attempt last.
The first is answering the customer directly, through a chat or voice assistant. It is the most exposed, because a wrong answer arrives with your name on it and is yours to repair, not your supplier's. Since 2 August 2026 it also carries a legal duty in the European Union, where Article 50 of the AI Act requires that people be told they are interacting with a machine, clearly and at the first exchange, unless it is already obvious. Suppliers, meanwhile, count a conversation as resolved when nobody asked for a person, which is not the same as a customer who got what they needed.
The second is helping the person who answers. The software drafts a reply for your agent to check, summarizes a customer's history before a call, translates, or finds the paragraph of your policy that applies. The industry calls this agent assist, and the point is that your customer still talks to a person while your staff catch the mistakes before anyone outside sees them.
The third is sorting the work: reading what arrives, tagging it, routing it and pushing the urgent or angry messages to the front of the queue. Customers never see it, and a mistake there costs minutes rather than relationships. It is also work that older rule-based automation did badly, because a fixed script cannot tell an angry customer from a confused one. Customer operations is also where most companies meets this software first, a pattern that repeats across enterprise applications.
Jobs two and three are where most owners should start, for a reason that has nothing to do with ambition: a mistake there is caught by your own staff, on your own time, before a customer ever learns it happened. That is why this kind of customer service automation tends to survive contact with reality, while the customer-facing version ends up in the news.
دعنا نبني شيئاً استثنائياً معاً.
اعتمد على شركة Lasting Dynamics للحصول على جودة برمجيات لا مثيل لها.
The best independent study of AI in customer service followed 5,172 agents, and its headline finding is not the one usually quoted. The average matters far less than who the average is made of.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond published Generative AI at Work in the Quarterly Journal of Economics in 2025, tracking the rollout of an assistant that suggested replies to agents at a Fortune 500 software company selling business software to small and mid-sized American firms. Issues resolved per hour rose 15% on average, but among novice and low-skilled agents they rose 34%, while among experienced, highly skilled agents the effect was minimal. That distribution is the finding: the gain did not land on your best people but on the ones who had not yet become your best people.
The number that should hold an owner's attention, and the one that makes the business case for AI customer service, is attrition. Among agents with less than six months on the job, access to the assistant came with roughly ten percentage points less attrition, a fall of about 40% against a baseline of 25% in that group. The same paper cites industry estimates putting annual turnover in contact centers near 60%, at a replacement cost of ten to twenty thousand dollars per agent. Training people and replacing the ones who leave are two of the largest hidden costs of any support team.
The limits deserve equal prominence. This is one company, one tool, and a rollout that ran from November 2020 to February 2021 on a generation of models several years behind anything you would buy today, with most agents working outside the United States, chiefly in the Philippines, for subcontractors. The authors flag the attrition result as their least robust, since they cannot rule out that the tool went first to agents the company expected to stay.
One measure did not move at all, and it is the one most owners would check first. Customers wrote to agents more warmly, by about half a standard deviation on a sentiment scale, and asked for a manager roughly a quarter less often, yet the formal satisfaction score barely changed. As for the experienced staff who gained little, that pattern is not confined to support work, and it is the subject of a separate piece on where the effort actually goes.
Your customers have already told researchers what they are afraid of losing to AI customer service, and it is the part you would want to automate last anyway. Their answer is not really about accuracy.
A Gartner survey of 5,728 customers, conducted in December 2023 and published in July 2024, found that 64% would prefer companies did not use AI in customer service at all, and that 53% would consider moving to a competitor if they learned a company was going to use it. The largest concern was not wrong answers: it was that reaching a person would become harder, ranking ahead of worries about jobs and, after those, about accuracy. Keith McIntosh, who led the research, put it plainly, saying that once customers have exhausted self-service they are ready to reach a person, and many fear the software will become one more obstacle in the way.
Emily Potosky, a senior director of research at Gartner, was blunter about capability: AI "simply isn't mature enough to fully replace the expertise, empathy, and judgment that human agents provide". The conversations that should stay with people are easy to name: complaints, exceptions to your own policy, anything where money goes back to the customer, bereavement, health, and anyone already halfway out the door. They are low in volume and high in consequence, which is precisely the kind of work that automation handles worst.
The route to a person is part of what you sell, so a company that keeps its people on these conversations, and says so plainly on its contact page, is describing a feature rather than an overhead.
بدءاً من الفكرة إلى الإطلاق، نقوم بتصميم برامج قابلة للتطوير مصممة خصيصاً لتلبية احتياجات عملك.
شارك معنا لتسريع نموك.
The roles change before the headcount does, and if you do not decide in advance what the freed time is for, it will quietly disappear. This is the part vendor pages skip.
Someone has to own the answers, because when a price, a product or a policy changes, whatever the software reads has to change the same day or it will keep telling customers last quarter's terms with complete confidence. Someone else has to review the work, reading a sample of AI-handled conversations weekly the way a team leader reads a new hire's tickets. Both are jobs with names attached, and the reviewer has to be allowed to switch things off.
Then the agents move to the harder cases, which is the trade-off almost nobody puts in the business case. When the simple tickets go to the software, every conversation your people have left is a difficult one, so each shift gets harder even when the queue gets shorter, and training, breaks and pay may need to reflect that before it shows up in your turnover figures.
The hiring data suggests this is under way. In that same October 2025 survey, 42% of organizations said they were hiring specialized roles such as AI strategists, conversational AI designers and automation analysts, and Melissa Fletcher, who reported the findings, advises leaders to stop framing AI customer service programs around headcount reduction and to plan for the new roles instead. So plan the new jobs before you plan the cuts, and if headcount does fall, let it fall through ordinary turnover once the numbers justify it. A hiring freeze can be lifted next quarter; a redundancy cannot be taken back.
The number your supplier puts first can improve while your team gets busier, which, according to its union, is what happened at Commonwealth Bank. Any scorecard you accept from a supplier will be built out of what that supplier can measure.
That first number is usually deflection or containment, both of which mean conversations the software kept away from a person. A contained conversation can be a solved problem, and it can equally be a customer who gave up and told a friend instead. Suppliers who charge per resolution have an obvious interest in the broader definition, so read theirs before signing.
Six numbers avoid that problem, because the supplier defines none of them:
Take a month of those six before you switch anything on, because without a baseline you will be arguing from memory against a dashboard, and the dashboard wins that argument every time, including the times when it is wrong. If repeat contacts and complaints rise while the AI dashboard improves, believe the complaints.
The sequence matters more than the choice of tool, and the first two steps of AI customer service usually cost nothing beyond licenses you already hold. Each one is reversible, which is what makes this order worth following.
نحن نصمم ونبني منتجات رقمية عالية الجودة ومميزة.
الموثوقية والأداء والابتكار في كل خطوة.
Start by helping your staff, since reply drafting and conversation summaries are now included in most established help desk products, which makes the first move a matter of finding what you already pay for and switching it on for one team, with drafts still checked by a person before they are sent. Then sort the work, because automatic tagging and routing is invisible to customers and can be undone in an afternoon, which makes it the cheapest place to discover how messy your categories really are.
Only then let the software talk to customers, and only for one kind of conversation, the dullest and highest-volume one you have. Tell people they are dealing with a machine, keep the route to a person one click away rather than buried, and read the transcripts yourself for the first month.
Staffing comes last, judged on six months of those six numbers rather than one quarter of enthusiasm. Most readers should not build anything custom yet. Custom work earns its cost when the software has to reach inside your own order, booking or claims systems, or when customer data cannot leave your infrastructure. That is what we sell, and it is not step one.
The companies now getting value out of AI customer service used it to make their team better before they asked it to make the team smaller. That is the whole finding, and it is as available to a twenty-person company on a help desk subscription as it was to a Fortune 500 company with 5,172 agents.
Does your support team need AI that works inside your own order, booking or claims systems? Tell us what it has to handle.
This article was written by Michele Cimmino at Lasting Dynamics, an EU-based custom software development company. His other writing on this blog covers the Cyber Resilience Act, enterprise AI agents and the move from rule-based automation to systems that reason. The view taken here comes from building this software for clients rather than from selling a platform, which is why the recommendation starts with the tools you already pay for and treats a custom build as the last step rather than the first.
In three distinct jobs. It answers customers directly through a chat or voice assistant; it helps the person who answers, by drafting replies, summarizing history and finding the right policy; and it sorts incoming work by tagging, routing and prioritizing messages. Only the first speaks to your customers, and it carries most of the risk.
Not in the way the headlines suggested. Klarna and Commonwealth Bank both reversed staff cuts they attributed to AI, and Gartner predicts that by 2027 half the companies that cut headcount for AI will rehire for similar work under different titles. Its October 2025 survey found only 20% of service leaders had actually reduced agent staffing. The work shifts toward harder cases rather than disappearing.
AI agent assist is software that helps the person answering rather than replacing them. It suggests a reply for the agent to check, summarizes the customer's history, translates, and locates the policy that applies. The customer still talks to a person, so mistakes are caught internally. It is the safest form of AI customer support to start with.
Usually yes for the two jobs your customers never see, drafting help for staff and automatic sorting of incoming work, which are often already included in the help desk you pay for. Be slower with a customer-facing assistant, which carries legal disclosure duties in the EU and reaches customers directly.
With six numbers a supplier cannot control: repeat contacts about the same problem, time from first message to the problem being fixed, customer satisfaction, complaints, staff turnover and time to train a new hire, and support cost per case genuinely resolved. Take a baseline month first, and do not judge the program on deflection alone.
حوّل الأفكار الجريئة إلى تطبيقات قوية.
لنصنع معاً برمجيات تُحدث تأثيراً.
ميشيل سيمينو
أؤمن بالعمل الجاد والالتزام اليومي كوسيلة وحيدة للحصول على النتائج. أشعر بجاذبية لا يمكن تفسيرها للجودة وعندما يتعلق الأمر بالبرمجيات فهذا هو الدافع الذي يجعلني وفريقي نتمسك بشدة بممارسات أجايل والتقييمات المستمرة للعمليات. لديّ موقف تنافسي قوي تجاه كل ما أتناوله - بطريقة لا أتوقف فيها عن العمل، حتى أصل إلى القمة، وبمجرد أن أصل إلى القمة، أبدأ العمل للحفاظ على مكانتي.
