When Kyndryl spun out of IBM in 2021, it inherited best-in-class platforms and a fragmented buying process. A panel with ORO Labs and Exiger reveals what it took to fix that — and where AI fits in the sequence.
Kyndryl builds, runs, and transforms mission-critical technology for 4,000 customers in highly regulated industries — from government agencies to financial institutions like American Express and airlines like Delta. When the company spun out of IBM in late 2021, it deployed a full enterprise stack: SAP, Workday, ServiceNow, Okta, M365. Each platform worked well as a system of record. But the experience of buying something — of a Kyndryl employee requesting a third-party product or service — remained fractured across 12 enterprise functions, 7 to 50 separate processes, and up to 40 people per transaction. The scale compounded the problem: 5,000 active suppliers across 12,000 legal entities in 50 countries, with 400 procurement professionals fielding requests, and 85% of all purchasing tied directly to customer engagements.
In a webinar moderated by Exiger’s David Morgenstern, Ben Shanahan, VP of Strategy and Transformation at Kyndryl, Sudhir Bhojwani, Co-Founder and CEO of ORO Labs, and Skyler Chi, SVP and Head of Product and Innovation at Exiger, worked through the problem and the fix. Three themes emerged: start with the buying experience, not the risk function; bring supplier risk into the buying decision at the point of commitment; and build the workflow before you add the AI.
What This Means for Your Organization
1. Map the buying experience before you buy another tool. If your platforms are strong individually but the employee path from request to order is fractured across functions, the gap is orchestration. Startby standing in the shoes of a business requester and mapping every step, everyhandoff, every approval — then look for a single front door that wires thosesteps together.
2. Move risk assessments to the point of commitment. Periodic snapshots at onboarding miss changes that happen daily. Embed risk scorecards at the moment a buyer selects a vendor, route high-risk domains tothe right SME automatically, and log every decision with its supporting evidence so the audit trail is built into the workflow.
3. Reduce alerts by adding context, not by filtering signal. A risk alert is useful only when a buyer can tell whether it matters to this purchase. Connect each signal to the correct supplier relationship, add operational context (criticality, substitutability), and make the next action clear. The goal is fewer, better-interpreted alerts — not more data for buyers to sort through.
4. Sequence AI after the process work, not before. Build the risk controls, process engineering, data architecture, and orchestration first. Then identify AI candidates: repetitive, structured tasks where the outcome is bounded, the context is reliable, and the guardrails are clear. Plug agents into the existing workflow so employees encounter them naturally rather than switching to a separate tool.
5. Pick one workflow, measure it, and iterate from there. Choose a single consequential process where the problem is repeated and measurable. Embed engineers alongside the mission owners who will use it every day. Validate continuously. Expand only after that one workflow is working — large-scale rollouts that try to solve everything at once tend to confuse activity for progress.
The Problem: Fragmentation Between Systems
Shanahan described the situation Kyndryl inherited after the spinout. The company had over 5,000 active suppliers across 12,000 legal entities in 50 countries. About 85% of its purchasing tied directly to customer engagements — managed services, consulting work that required sourcing hardware, software, and services, then bundling them into Kyndryl package offerings. Risk management was essential, and speed mattered.
But the buying experience itself was the bottleneck. An employee who needed to purchase something faced a maze of processes spanning procurement, finance, legal, cybersecurity, data privacy, and HR. The platforms were strong individually. The connections between them were not.
“The problem wasn’t that we didn’t have systems, processes, and controls. It was that connective tissue between those elements that we needed to ensure a seamless end-to-end buying experience.”
The solution was ORO Labs, which became the orchestration layer — a single front door with guided routing, approvals, and data flow connecting every core platform. Bhojwani, who spent eight and a half years at SAP Ariba before founding ORO, described the core thesis: no single tool can handle the full buying process across heterogeneous environments.Orchestration brings teams, systems, and data into one coherent experience so users get what they need without bouncing between platforms.
ORO solved the routing problem — getting a request through the right approvals across the right systems. But routing alone does not tell you whether the supplier at the other end of the request is one you should be doing business with. That is where Exiger came in. Kyndryl needed third-party risk intelligence embedded at the point where a buyer selects a vendor, so the decision about whether to proceed carries the supplier’s current risk posture, a documented review, and an auditable trail. The two capabilities —orchestration from ORO, risk intelligence from Exiger — were designed to work together inside a single buying flow rather than as separate tools a user has to switch between.
Bringing Risk Into the Buying Decision
Historically, Kyndryl assessed third-party risk at two points: supplier onboarding and periodic reviews for critical suppliers. But supplier risk changes daily — financial health, cyber posture, ESG commitments, sanctions and adverse media can shift well after the onboarding snapshot. Some suppliers Kyndryl inherited from IBM had been onboarded years earlier.
The question Shanahan wanted answered at the point of any new commitment: Do we understand the supplier’s current risk posture, and do we have the right mitigation or acceptance in place before we proceed?
In practice, when a request enters the ORO flow and reaches vendor selection, Exiger generates a scorecard across several risk domains — including cybersecurity, ESG, financial, FOCI, operational, and compliance— on a one-to-ten scale. Anything scoring a nine or ten triggers an automated review by the relevant domain SME: cybersecurity issues go to the CISO’s team, data privacy to legal, business continuity to corporate security. The SME can review the Exiger executive summary and deep-dive insights, understand the root cause, and advise the business on whether to mitigate, accept, or choose a different supplier. Every decision and its supporting evidence is logged in the workflow for auditability.
Signal Is Not a Decision
Chi framed the design challenge at the center of the partnership. Exiger makes over three trillion data points available across 30 million underlying sources. The goal is to reduce what each buyer needs to interpret so the next action becomes clear — to distill that volume down to the specific context a buyer needs at the moment of a decision.
“Risk signal is not a decision.”
For signal to become actionable, three things must be true: the evidence must be traceable to source provenance through an audit log; it must be connected to the correct supplier or relationship (resolving complex ownership and sub-tier connections); and the business exposure must be understood — whether the supplier supports a critical service, whether it is easily replaceable, whether alternatives exist. A financial concern on are placeable supplier may merit light monitoring. The same concern on a supplier supporting a critical Kyndryl customer may require immediate escalation.
“Our objective is not to expose any of Kyndryl’s 400 procurement professionals to more noise or more alerts. That just drives analysis paralysis.”
The system is designed so a buyer does not need to become a TPRM expert. Risk levels appear across the risk domains in the buying flow. When a domain triggers high risk, the right SME is pulled in to explain the why behind the what — the root cause, the findings, the semantic summarization Exiger provides from its sources. The buyer sees the risk score and the explanation. The specialist sees the underlying evidence. The workflow records the decision.
Where AI Belongs — and Where It Does Not
Kyndryl’s approach to AI followed a deliberate order —one that runs counter to the instinct many organizations have when they encounter a problem. The team spent 18 months on what Shanahan called the boring work: a risk controls structure, process engineering, end-to-end process mapping, data architecture. Then it wired everything together with ORO. Then it injected intelligence. Only then did it look at where agents could take on structured, repetitive tasks — gathering, comparing, validating, summarizing, routing — plugged directly into the existing workflow rather than sitting as a separate tool employees had to swivel-chair into.
“We started with the boring work first before we got to the shiny AI stuff.”
For any AI candidate, Shanahan’s team asks a series of questions: Is the target outcome clearly defined and bounded? Does the agent have reliable context and data? Can we give it clear permissions and guardrails? How do we trace and verify what it did and the logic applied? Where does a human need to intervene? And where in the workflow should it plug in?
Bhojwani connected this to the accountability question.The RACI matrix — who is responsible, accountable, consulted, informed — does not disappear because agents are involved. An auditor will not accept “the AI agent made a mistake” as an answer.
“Building an AI agent that can independently do a small task is actually quite easy. The hard part is stitching it all together in a RACI model that is SOX compliant.”
Bhojwani’s framing: the right person in the loop, not just a human in the loop. The orchestration layer brings the right person to engage at the right risk level, with the right escalation path — and the data context to support the decision.
“If we want full agentic execution, there’s a very immutable audit trail attached to that agentic flow so that we can continue to be a supervisory human on the loop.”
Chi described how Exiger handles AI-assisted risk decisions more broadly. Every recommendation is cited to the sources the AI used, with an audit trail into how the risk model arrived at its output. Users who want step-by-step guidance get it, all traced to source provenance. The design principle: AI can accelerate work, but it can never hide the evidence orblur accountability across the RACI.
Starting with Outcomes, Not Features
Capabilities move from requirement to production in days, not months. Exiger deploys what it calls experience builders — engineers who are also domain experts in third-party and supply chain risk management, embedded close to the customer mission rather than working from Confluence tickets.
The cadence: a requirements session to land on a prototype on Monday. Engineering industrializes what was validated with users Tuesday through Thursday and Friday. The capability quickly ships to production the following Monday, thanks to the Exiger AI-product development lifecycle. The team works from Exiger’s existing data and AI stack, building workable prototypes that expose unknowns early — before committing to a lengthy engineering process that might produce something users say does not work.
Shanahan described an example at Kyndryl. The team initially designed the system so only risk domain experts could see the Exiger executive summaries and root-cause explanations. But business requesters started asking for more context — they wanted to understand why cybersecurity flagged as a nine out of ten. So Exiger and ORO are now building the ability for any Kyndryl employee viewing a supplier profile in ORO to click through to the Exiger executive summary and the reasoning behind the scores. The requirement came from watching users use the system.
The Visibility Payoff
With the workflow wired together, Kyndryl launched an analytical command center that shows every procurement-related request from start to finish — every step, every team, every queue, every bottleneck, every request for more information. When the business says procurement is too slow, the conversation shifts from perception to evidence.
If a risk review took two weeks, Shanahan’s team can open the transactions, see whether the request sat in a queue for five days before someone picked it up (a capacity problem), or whether it bounced back to the requester for more information (a form design problem). The command center pinpoints where the friction is and drives focused improvement to that specific point.
Advice for Leaders Starting Out
The three panelists drew different lessons from thesame experience — and the differences are instructive for any procurement orTPRM leader weighing a similar path.
Shanahan’s guidance: zoom out. Focus on the end-to-end buying experience from the perspective of the person trying to buy something, not on your functional scope. Do the process work — scrutinize every element, simplify, eliminate anything that does not drive business value or manage risk.Then connect it, make it visible, and progressively embed intelligence and automation.
“Don’t focus on your scope. Focus on the end-to-end buying experience by standing in the shoes of a business requester.”
Bhojwani emphasized that these projects are organizational changes, not feature deployments — iterative, multi-stakeholder, and requiring an agile platform and approach. You will not solve every subprocess in one shot. You take multiple cuts over time. Chi’s advice was the most concise:
“Start with one. Don’t boil the ocean.”
Behind that brevity was a sharper observation. Chi has watched Exiger’s customers attempt this kind of deployment, and the pattern is consistent: the ones who succeed start with a real operational problem. The ones who struggle arrive with a predetermined feature specification and get wrapped around those features instead of the outcome they are trying to achieve. The distinction matters because it determines what you build first —and whether the people who will use it every day have any reason to adopt it.
His recommendation: choose one consequential workflow where the problem is repeated and easily measurable. Embed forward-deployed engineers directly into that process alongside the mission owners who will use it every day, so you are building with them rather than handing them something they had no part in shaping. Validate that one workflow continuously. The improvement comes quickly when the feedback loop is short — and it keeps the team from confusing activity for progress, which is what large RACI-driven projects can become when they try to solve everything at once.
What’s Next
For Kyndryl, the next 12 to 18 months focus on three areas: reducing elapsed time and unnecessary touch points without weakening controls; extending proactive risk intelligence beyond the purchasing decision into the life of the engagement (so a supplier three years into a managed service contract generates an alert when its financial liability spikes); and progressively injecting AI agents to handle administrative tasks so procurement professionals spend their time on commercial judgment, supplier relationships, and customer bid support.
The webinar’s throughline, in Shanahan’s words, was sequence: lay the foundation, wire it together, inject intelligence, then automate. The connective tissue comes first. The AI comes after. And the buyer never leaves the workflow.


.jpg)









