UKG’s 12,000 AI Agents Are a Management Test, Not a Tech Win
Most companies don’t have an AI strategy. They have a software shopping problem. UKG built 12,000 AI agents—and the hard part wasn’t the technology.
Most companies don’t have an AI strategy. They have a software shopping problem.
UKG has built more than 12,000 AI agents and launched 387 internal AI applications. That sounds impressive—and it is—but the number that should make every CEO and founder sit up is not 12,000. It’s 27%.
That is the estimated share of UKG customer calls now handled autonomously by AI-enabled voice and chat agents. That is where the bullshit stops. Not at a keynote. Not in a LinkedIn post about “transformation.” At the point where a customer gets helped, a human team has more room to solve harder problems, and the business can measure whether service, sales and retention improve.
Prakash “PK” Kota, UKG’s chief information officer, is making a useful case study out of an issue most executives are still dodging: AI is not an IT project. It is a management job.
UKG did the boring work before chasing the shiny thing
Kota joined workforce-management software company UKG in April 2025 after two decades at Autodesk, including seven years as CIO. He did not arrive, buy a pile of AI licences and announce that the future had begun.
His first major job was cleaning up unfinished technology integration work following the 2020 merger of Kronos and Ultimate Software. Within his first 90 days, he centralised responsibility for systems, data and AI. Within six months, UKG had consolidated its core enterprise-resource-planning software with Microsoft, its customer-relationship-management software with Salesforce, and combined multiple data warehouses.
That sequence matters.
Every founder wants to believe their business is special enough to skip the boring stuff. I’ve watched plenty of companies try it. They bolt an AI layer onto disconnected data, muddled ownership and half-broken workflows, then act surprised when the result is expensive chaos with a chatbot attached.
UKG’s approach was more sensible. Get control of the plumbing. Then let people build.
The company says its 14,000 employees have broad access to ChatGPT Enterprise and Google Gemini Enterprise, while product and engineering teams also use Anthropic’s Claude Code. More than 1,400 employee ideas have produced 387 internal applications. Kota has also avoided locking the company into long multi-year AI contracts because the tools are changing too quickly.
That last bit is particularly sharp. A five-year software agreement made sense when enterprise technology moved at the speed of a council permit. AI does not. Locking your company into a vendor before you know which workflows matter is not strategic commitment. It is paying upfront for the right to be wrong.
The leadership move was decentralisation with guardrails
The lazy version of AI leadership is one central team approving every use case. It feels safe because executives can point to a committee, a policy document and a bloke from legal.
It also kills momentum.
UKG has used an idea-to-implementation framework that lets employees submit ideas and relies on power users and job-specific personas to help prioritise them. The company’s logic is dead right: the people doing a job usually know which repetitive, irritating and low-value tasks are wasting their week. Senior leaders often do not.
But decentralisation without standards is just shadow IT in a nicer shirt.
The management trick is to separate idea generation from permission to deploy. Let anyone identify a problem. Give teams approved tools, clear data rules, risk limits and a practical route to get something tested. Then insist on an owner, a metric and a date for deciding whether the thing stays, improves or gets killed.
Kota has framed UKG’s internal AI effort around three things: talent, tools and tokens—the computing spend required to use these systems. That is a useful antidote to the usual fixation on tools alone.
You can give everyone access to the best model on earth. If they do not understand the work, lack confidence to use it, or have managers who punish experimentation, you have bought very expensive autocomplete.
The useful metric is not hours saved
UKG estimates its AI work is adding 8,500 hours of productivity each month. Fine. That is worth knowing, but it is not the finish line.
“Hours saved” is the corporate world’s favourite AI metric because it is easy to put in a PowerPoint deck. The problem is that saved time can become dead time. If a customer-support worker saves 30 minutes but spends it refreshing Slack, the company has not created value. It has simply digitised a break.
UKG appears to understand this better than most. For coding, Kota says the company looks beyond the volume of code produced and asks whether customers actually buy the features created. In customer service, it is tracking productivity alongside upselling and customer-sentiment scores.
That is the right standard: what changed in the commercial outcome?
Did response time fall? Did customer satisfaction rise? Did churn improve? Did conversion increase? Did errors reduce? Did the team redeploy capacity into higher-value work? If you cannot answer one of those questions, your AI project is probably a demo pretending to be a strategy.
The 27% autonomous handling rate matters because it connects deployment to a real workflow. UKG says its system also creates roughly 300 case studies from newly resolved service issues, which can then help both the AI and human teams handle similar matters. That is not merely automation. It is a feedback loop.
And feedback loops are where operational advantage comes from.
The overlooked risk: you can automate away your training ground
Here is the contrarian bit: not every efficiency gain is good management.
Entry-level work is often tedious, repetitive and underpaid. It is also where people learn context. A junior support rep learns what customers actually complain about. A junior analyst learns which numbers matter and which numbers are decorative. A junior developer learns why an apparently simple request breaks three other things.
Automate all of that carelessly and you may save money this quarter while starving your business of future managers, operators and product people.
This is why AI cannot sit solely with the CIO, CTO or a shiny new “head of AI.” The CHRO needs to be in the room, and so do the people running revenue, operations and finance. McKinsey’s 2026 HR Monitor found only 11% of surveyed organisations take a strategic, long-term approach to workforce planning. That is a fairly savage admission when work itself is being redesigned in real time.
The question is not, “Which jobs can AI remove?”
The better question is, “Which tasks should disappear, which skills become more valuable, and how will we teach people to make better decisions once the machine handles the basic work?”
UKG’s focus on frontline workers is worth watching here. The company serves organisations in industries such as retail, manufacturing and healthcare—places where AI is not just drafting emails for knowledge workers. It can affect shifts, scheduling, onboarding, service and the day-to-day experience of people who do not spend their lives in front of a laptop.
That is a bigger management challenge than installing a chatbot for the marketing team.
What this means for you
If you run a business, do not begin by asking your team which AI tool they want. That question gets you 40 subscriptions and no operating advantage.
Do this instead tomorrow:
1. Pick one painful workflow with a clear number attached. Customer-response time, proposal turnaround, no-show rate, stock discrepancy, support backlog, sales follow-up—something real.
2. Name one accountable operator. Not a committee. One person who owns the result, the risk and the decision to stop if it fails.
3. Set three measures before the pilot starts. One efficiency measure, one quality measure and one commercial measure. For example: calls resolved, customer satisfaction and upsell rate.
4. Give the people doing the work permission to suggest fixes. They see the friction. Your executive team mostly sees reports about the friction.
5. Keep your contracts short and your architecture flexible. The AI market will change faster than your procurement calendar. Do not confuse a vendor relationship with a competitive advantage.
6. Decide where saved time goes. Into more sales calls, better customer recovery, training, product improvement or fewer people? Be honest. Every option has consequences, and vague promises poison trust.
UKG’s 12,000 agents are not the lesson. The lesson is that Kota treated AI as a redesign of work: centralise the foundations, distribute the problem-solving, measure commercial outcomes and make people accountable for what happens next.
That is leadership. Everything else is just software with a flashy logo.