Should AI & Automation CoE's Merge?

One CoE or Two? How to Govern AI and Automation Together

The end state is one converged Center of Excellence governing AI and intelligent automation together, because a single business process rarely respects tool boundaries and shouldn't be governed as if it does. But convergence has a timing problem. Where an intelligent automation program is mature and an AI practice is still finding its value levers, forcing them into one model too early can dilute the proven program and starve the experimental one. The working answer most enterprises land on: one boat, two engines, one navigation system. Converge strategy, pipeline, and value tracking. Keep the delivery modes distinct until AI maturity catches up.

The question is not academic. There is now well over a billion dollars of AI agents in circulation, and intelligent automation crossed the tipping point where ad hoc, ungoverned adoption starts creating more risk than value. That is the moment every technical evolution reaches: enough demonstrable value that the absence of governance turns an advantage into a liability.

Ashling co-founders Don Sweeney and Marshall Sied argued this out as a live debate, one taking convergence and one taking separation, across the six elements of an AI and automation governance model. The format mirrors the argument enterprises are having inside their own walls. Here is the substance of both sides, element by element, and where the honest answer lands.

Why This Debate Exists Now

Rewind to 2017 or 2018. Intelligent automation was a grassroots movement out of shared services or corporate systems, with no governing bodies. The wild west. That is normal for an evolving technology, and it works right up until there is enough value at stake that confusion, unclear funding, and unmanaged risk start eating the gains.

AI arrived at the same crossroads faster. The pattern is identical, which is why the governance question matters now rather than later.

Gartner's Business Orchestration and Automation Technologies category, the "BOAT" framework, is the market's attempt to structure this. Ashling has long used a similar internal framework it calls multimodal automation. The premise underneath both: capabilities cannot stay federated if you actually want value from AI. You need upstream and downstream deterministic workflows that are auditable, explainable, and run the same way every time, and you need to take advantage of probabilistic, LLM-based agentic solutions. BOAT is less an architecture than a capability map, and a good place to start.

The Six Pillars of a Governance Model

The debate runs through six components of any AI and automation Center of Excellence. For each, there is a case for convergence and a case for keeping things separate, and the balance shifts depending on how mature each capability is in your organization.

Element 1: Strategic Direction and Governance

The case for convergence. A tool-based CoE sends you out with a hammer looking for a nail. When an organization runs a separate AI CoE, automation CoE, and low-code CoE, those are effectively the same CoE competing for the same opportunities and solving them in different ways rather than the best way. Executive sponsorship has drifted toward AI, and funding discipline has drifted with it: automation gets held tightly to ROI because it is mature, while AI funding too often happens because the technology is exciting. That is precisely why so many AI initiatives fail. They are not held to the standards every automation initiative already meets. Combine them under one funding model and one standard, and the success rate climbs.

The case for separation. Few enterprises are actually ready for a consolidated strategy today. Many Global 2000 and smaller organizations have already earned tangible value from their intelligent automation programs. If the AI failure rate is as high as reported, converging now risks diluting a proven thing. Strategy is about the why and how you capture value. If you cannot do that in a converged state over the next one to two years, why not keep them separate for now and industrialize automation further?

Where it lands. Not "if," but "when." The instinct to protect automation's momentum is sound, and the middle ground is real: sweat the AI assets you have bought, but when a problem objectively will not benefit from AI, swing back to proven deterministic automation and let it run. Every enterprise should look for those points of intersection.

Element 2: Architecture and Security

The case for convergence. As programs mature, a real solution is one orchestrated end-to-end flow: RPA steps, AI steps, human-in-the-loop, long-running workflows, all coordinated together. You have one internal audit team, not several. Auditability for RPA is easier than for AI, but that is not a reason to hold one component of a process to a lower standard than the rest. Raise the bar across the board and hold every component, human or machine, to the same level of accountability.

The case for separation. Security postures differ too much right now. Deterministic automation means credentialized systems, SOC-ready audit trails, and provable inputs and outputs. AI means data pipelines, non-deterministic outputs, and proving you stayed inside your guardrails rather than proving an exact execution path. Converging can also slow you down. If a workflow can be done deterministically, you do not want to drag it through AI governance toll gates it does not need, because that erodes time to market.

Where it lands. The speed argument is valid only when the toll gates are genuinely unnecessary. The risk is subtler: business users see an end-to-end process, not four or five components. They expect deterministic delivery speed even when an agentic step carries extra governance and data-lineage requirements they are not thinking about. Managing that expectation gap is its own change-management task, and it is the reason to raise the standard rather than lower it.

Element 3: Talent and Velocity

The case for convergence. Capabilities are converging the way tools already did. OCR, RPA, and low-code used to be dedicated products, and now single vendors provide all of them. Build those skill sets inside one CoE rather than spinning up a separate, novel "startup" CoE. If you do not have the skills now, you will need them soon anyway, so keep every tool in one tool belt and solve each problem the best way for the organization rather than routing it to whichever CoE has spare capacity.

The case for separation. Treat the workforce the way you treat agents: shrink and specialize, because things move too fast for generalists. Even a firm that does this for a living hears constant feedback that there is too much training to keep pace with, so expecting a Global 2000 organization to keep pace across both disciplines at once is optimistic. Talent drives velocity, and velocity right now depends on experts doing expert things. On one side, data engineers, ML engineers, and AI-specific skill profiles. On the other, lean process work, rules-based development, QA, and release management. Upskilling happens over time as the program matures, not on day one.

Where it lands. The strongest point cuts toward convergence, and it is a human one. Multiple CoEs make people territorial. Turf gets defended, every problem looks like your CoE's nail, and new capabilities trigger fights over which CoE owns them. A combined CoE removes that infighting and builds change readiness, because growth flows through one team rather than several competing ones. Given the industry-wide skills gap, cross-skilling inside one CoE may be less a preference than a necessity.

Element 4: Pipeline and Priority

The case for convergence. One intake process, not several. Look at the corporate objectives and run every idea through unbiased, objective criteria based on ROI or business impact, whether that is cash flow, working capital, employee experience, or customer experience. That intake should include someone who reviews the process itself, because you do not automate a bad process, and a solution architect across all tools who brings the full tool belt to each problem. One intake, one prioritization, one pool of capacity, so no CoE sits idle while another drowns in backlog.

The case for separation. This is the element where the separation argument is weakest, and even its defender conceded that pipeline is where convergence makes the most sense. The honest counter: if your intake is truly objective and you have not updated your funding, ROI, or payback metrics for AI, the deterministic projects will always come through faster, because their value is easier to prove. That can quietly starve the AI innovation the business is most excited about. The answer is bimodal: keep pulling incremental value through the commoditized, proven automation lane, and give AI a separate mode with room for learning curves and longer-horizon bets, because AI value levers are different. Not just cost and time savings, but revenue creation and churn prevention.

Where it lands. A consolidated pipeline wins on transparency, keeping everyone on the same page, as long as the value criteria are updated to recognize AI's different levers. This is also where a Phase 2 of automation shows up. With AI in the CoE, teams revisit automations already in production, especially brittle ones generating heavy exception logs, and add LLM or AI components to make them more flexible. That is a second bite at the apple: revisit the 40, 50, or 60 automations you have shipped, mine the exception logs, and look at the pipeline that failed on technical feasibility rather than business value. Much of it is buildable now.

Element 5: Continuous Improvement

The case for convergence. Continuous improvement means two things at once: improving how you define and measure success, and improving the automations themselves. AI as a new tool in the belt lets you go back and improve exception handling and user experience on existing builds. And there is always the next frontier to scan for, like orchestration, which becomes one more capability the CoE absorbs. Part of the CoE watches what is running today; part of it watches the horizon six to twelve months out.

The case for separation. Converging too fast blurs accountability for continuous enhancement, especially with separate task forces. The right mindset is CI/CD, taking signals from both deterministic confidence levels and probabilistic LLM outputs as ways to improve operations, and most enterprises do not even do that. They drop a solution into hyper-care and post-production support and consider it finished. AI makes that worse, because an agentic system ships faster but gets more scrutinized testing, and once live it is a product that needs continual enhancement. That is why ML ops and AI ops are entering organizations: the support structure differs from an ERP package or a rules-based workflow. The feedback loops, speed, and value are all different. Deterministic CI is about throughput and stability; AI CI is about drift, model accuracy, and prompt and policy tuning.

Where it lands. A split decision. Production monitoring belongs in every CoE, but what it looks like for a purely AI solution versus a purely automation solution is genuinely different. The two sets of CI activities are discrete even if a spreadsheet says to combine them.

Element 6: Value Tracking and Insight

The case for convergence. This is the element that decides whether a program survives. Too many AI automations are failing to move past proof of concept because they lack demonstrable value. They exist because AI is exciting or because a CEO came back from a conference wanting AI. AI is a powerful capability added to automation to turbocharge it, but the value framework has to stay a value framework. That framework can go beyond hours saved to include customer experience, employee experience, and Net Promoter Score, but everything, AI and automation alike, should be held accountable to ROI and a continuous improvement metric that keeps it sustainable.

The case for separation. Clients ask three questions, and the third has spiked in volume: what is your definition of agentic, where has it been deployed in production, and how did you calculate value? That third one is in the eye of the beholder. Deterministic value is easier to track and convert into capacity returned to the business. When you introduce productivity agents that help thousands of people by a minute or two each, plus a lot of intangible value, consolidating units of measurement to report to a board becomes genuinely hard. Expect more decentralization of value tracking into line-of-business task forces before convergence, because a shared services group tracks different metrics than commercial lending. As long as the line of business sees value and funds the work, questioning their value equation is counterproductive.

Where it lands. Centralized or decentralized, the non-negotiables are the same: measurable business value, defensible business value, and ongoing measurement of it. That holds whether you run one CoE or five.

One piece not to skip: storytelling and branding. When an AI or automation project succeeds, memorialize it. Build the success story tied to business value, because someone will question that value in six months, a year, two years. A documented, well-told success builds momentum, pulls new participants and ideas into the CoE, and gives you something externally validated to point to.

The Answer: One Boat, Two Engines, One Navigation System

There is no universally right answer. It depends on maturity, timing, and organizational posture. But the North Star is clear enough to name: one boat, two engines, one navigation system.

Converge the navigation, the strategy, the pipeline, and the value framework. Keep the two engines, deterministic delivery and probabilistic delivery, distinct while AI maturity catches up to automation's. Be thoughtful and objective about which element sits where on that spectrum, and the organization wins either way.