Gartner surveyed 275 service and support leaders and released the results on 4 November 2026 at its Customer Service and Support Conference. Only 4% of customer service organisations said they had achieved the expected ROI from their AI rollouts. Christopher Sladdin, a Senior Director Analyst at Gartner, said leaders “must stop asking how they remove humans from interactions and instead ask where their involvement creates disproportionate value for customers.”
The claim is that AI in customer service has not paid off
The number is already being read as a verdict. Ninety-six out of a hundred teams spent the money and did not get the return they planned for. The easy conclusion is that the technology was oversold, and that the fix is to put people back into the conversation.
I don’t read it that way. Before I accept that AI failed, I want to know what each of those teams actually bought, and what they measured.
What most teams bought
Most customer service AI installed over the last three years does one thing. It answers questions. It reads the help centre, replies to “what is your return policy”, and the dashboard reports a deflection rate. Tickets that never reached a person. That number goes into the ROI model as savings.
But a deflected ticket is not a solved problem. The customer who asked “where is my order” got a paragraph about typical shipping times. They still don’t know where their order is. So they email, or message again, or call. The ticket comes back with a worse mood attached. The dashboard counted a win and the business paid for the same question twice.
Measured that way, the return was never going to arrive. The money went on replies, and the return was supposed to come from outcomes.
What changed, and what didn’t
What changed is real. A business can now reply instantly, in any language, at 11pm, to every customer at once. That is worth having.
What didn’t change is where the work sits. The order is still in Shopify. The parcel is still with the courier. The refund still needs a decision and a transaction. If the AI cannot reach any of those, every real request still ends with a person doing the work. The expensive part of customer service is still being done by people.
Gartner’s own figure points the same way. The article says AI-forward organisations invest nearly 1.8 times more in the foundations than in the AI use cases themselves. Read that from an operator’s seat: the money that pays off is spent underneath the chat, on what the AI can reach and do. The 4% built that. The rest bought the use case and skipped the plumbing.
Take an online shop
Picture a shop doing a few thousand orders a month. Most messages fall into a handful of types: where is my order, change my address, I want a refund, is this in stock. An answer machine handles none of them to the end. It explains the policy and hands over.
Resolved looks different. A customer asks where the order is. The system checks the order, sees it left the warehouse on Monday, and replies with the tracking link. A customer asks to change the delivery address. The system checks the order hasn’t shipped, updates the address, and confirms. A customer asks for a refund on a S$40 item. The shop set its refund limit at S$50, so the refund is issued in the chat. A S$300 refund request goes to the team with the order details attached.
Each of those is an enquiry closed with no person involved. That is where the return comes from, in hours actually freed.
This is the layer I build AeroChat on. You describe the workflow in your own words, it connects to your order system and backend processes, and the conversation checks the order, changes the delivery or issues the refund, within the limits you set.
Where I agree with Gartner, and where I don’t
Sladdin is right that humans and AI are not fit for the same job. People belong where a human conversation changes the outcome: a big-ticket purchase, a complaint that could lose a long-time customer, and the work of reading thousands of AI-handled conversations to decide what to fix. His line that AI “enables leadership to learn from customer interactions at scale” is the part of the report I would underline. My own view is that competent AI can already handle most enquiries, somewhere around 80 to 90% of them, and the people move to supervision and to the moments that matter.
Where I part ways is the diagnosis. The 96% have an automation that stops at the answer, and adding people back into order-status questions does nothing for the ROI. It reinstates the cost the AI was supposed to remove, and leaves the cause untouched.
What to do
If you are in the 96%, don’t start with the question of whether to bring humans back. Audit what your AI is allowed to do.
- List your ten most common customer requests. Next to each, write whether the AI can finish it or only explain it.
- For each one it can only explain, name the system it would need to reach and the limit you would set. Refund amount, address changes before dispatch, which orders it may touch.
- Change the measure. Count enquiries resolved with no person involved, and the hours that freed. Stop reporting deflection.
- Put people on the conversations where a human changes the outcome, and measure those separately.