Tata Steel put AI agents on customer complaints. Turnaround halved

Tata Steel says agents that read complaints, spot defects in photos and route them cut average turnaround by 50%. The baseline and the cost are not published.

Tata Steel, the Indian steelmaker with 35 million tonnes a year of capacity and more than 76,000 employees, says AI agents that read customer complaints, detect the defect in the customer's photos and send each case to the right team have cut its average complaint turnaround time by 50%. The figure is in the company's own press release of 22 April 2026, and Google repeated it on stage at its Gemini at Work event on 9 October. Tata Steel has not published the turnaround in days, its complaint volumes or what the system cost.

This week a question is going round the technology press: if AI is as capable as its makers say, where is all the automation? The answer at Tata Steel is unglamorous. It is not in the blast furnace. It is in the queue of complaints, invoices and HR tickets that sits between a customer, a plant and an office, and the work it took over is the sorting, not the solving.

What the complaint agents actually do

When a steel customer complains, the complaint usually arrives as a bundle: an email, a test certificate, a delivery note and photographs of the coil, sheet or bar that is wrong. Someone has to read it, work out whether this is a surface defect, a dimensional problem, a paperwork error or a late delivery, and pass it to the people who can investigate. In a big mill that triage step can be where a complaint sits longest.

Tata Steel's release describes agents that do exactly that step. They analyse the complaint material, detect the intent of the complaint, identify defects from the images, and route the case to the relevant resolver group. The company uses Google's Gemini models, including its vision models, for the mix of text and pictures.

Note what the description leaves to people. The agent does not decide whether the customer is right, agree a credit note or change a process on the line. Resolver groups still do the investigation. If the agents handle only that step, the 50% cut in average turnaround most likely comes from removing waiting time at the front of the queue, which is also where most companies lose days without noticing.

The automation was in the sorting, not the solving.

How big was the programme?

The complaint agents are one of more than 300 specialised agents Tata Steel says it deployed in nine months. Most were built on Zen AI, an internal low-code platform that lets developers and frontline managers build and test their own agents. A second internal tool, the Tata Steel Digital Assistant, answers questions across company systems and documents; the release says it helps the HR helpdesk resolve more than 70% of routine employee tickets on its own.

Other agents handle invoice processing, the classification of Goods and Services Tax as creditable or not, and contract analysis. In its 2025-26 integrated report, Tata Steel lists an award for the adoption of agentic AI at Tata Steel Business Delivery Centre, its shared-services company. The pattern across all of these is the same: documents in, a classification or a routing decision out, a person downstream.

Jayanta Banerjee, Tata Steel's chief information officer, is quoted in the release naming reduced customer response times and predictive maintenance as the uses that matter. The two figures the company chose to publish, complaint turnaround and HR tickets, are both service measures, not production ones.

What did it cost?

Tata Steel has not said. The release gives no contract value, no running cost and no staffing change. It does not say how many complaints a year go through the agents, which products or plants are covered, or what the average turnaround was before. A 50% cut from ten days to five is a different story from a cut of two days to one.

Three more cautions belong next to the figure. It comes from a joint release with the vendor, so it is the company's statement, not an audit. "Over 300 agents" counts things built, not things in daily use. And "routine" HR tickets is the company's own category; we do not know what share of all tickets that is.

Does this apply to a 200-person manufacturer?

Partly. You do not need 300 agents, a low-code platform or a contract with Google. The idea worth copying is narrower: the slow part of complaint handling is often the first hour of reading and routing, and that step can now be automated with off-the-shelf tools. Whether it is worth doing depends on volume and on how long complaints wait before anyone starts.

Our own rough arithmetic, not Tata Steel's: a firm that gets 30 complaints a month and spends 20 minutes triaging each spends about 10 hours a month on the step. That is not a business case for software. The real gain, if there is one, is calendar time. If complaints sit two or three days in a shared inbox before reaching quality, a simple routing agent, or even a better form, can cut the customer's wait sharply at almost no cost.

What it does not mean is that the investigation gets faster. Root-cause work, the 8D report, the credit decision and the conversation with an angry buyer stay with people, at Tata Steel as everywhere else.

What to do on Monday

Pull the last 50 complaints and, for each, record two dates: when it arrived and when the person who resolved it first saw it. If the gap is short, an agent will not help you; spend the effort on investigation. If it is days, fix the intake first.

Ask customers to submit complaints with a fixed set of fields and at least one photograph. Tata Steel's agents work because the complaint arrives as material a machine can read; a phone call noted on a sticky pad does not.

Write down your routing rules: which defect types go to which team, and who decides when it is unclear. If you cannot write them down, an agent cannot follow them either.

Then, if volume justifies it, pilot an agent on routing only, with a person checking every decision for the first month, and measure the same two dates again. If your supplier cannot show you that before-and-after, treat any 50% claim, including this one, as a promise rather than a result.