Earlier this year, on the floor of an industry conference, I spent the better part of an hour with a shared services leader at a global logistics operator. His mandate was one most executives in this space would recognize: extend a proven offshore capability into new markets. And one of the options on his table was striking. Instead of standing up a new center at all, he could deploy a bot model. Automation in place of a building. Software agents in place of analysts.

Understand what was actually being offered. That was not a technology purchase. That was a delivery-model decision, one of the most consequential choices in any bestshoring strategy, arriving dressed as a software license. And the deeper we got into it, the clearer the real picture became. He was asking all the right questions, which is exactly why the pattern stood out: his hard problems were not technical at all. Internal stakeholders had not bought in. Budget authority was not settled. The core systems underneath the very processes he wanted to accelerate were still unstable. The fastest option on his table was the one he had the least organizational ability to control.

That conversation is the state of our industry in miniature. Bestshoring has always been the discipline of deciding where work should live, who should do it, and how it should be organized. AI has now made the speed of that work the cheapest variable in the equation. Quoting, shipment file processing, exception handling, customer communication: all of it can move faster than the structures built to manage it. Control has become the scarce asset.

So whether you are a COO standing up your first global capability center or a GBS leader running a mature captive alongside three provider relationships, the question that separates you from your peers is no longer how fast the technology can go. It is how fast your operation can go while keeping control of the car.

I watch this question land on two very different desks. The first belongs to the leader building something new: a first captive center, a GCC, a third-party BPO relationship, designed with AI in the room from day one. The second belongs to the leader who already runs the operation: an established shared services organization or outsourcing relationship that AI is about to be layered onto. The brakes matter on both journeys. They just get installed differently.

The stakes are the same, though, and they are quantified. In mid-2025, Gartner predicted that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing “escalating costs, unclear business value or inadequate risk controls.” Notice what is missing from that list: model capability. These projects do not fail because the technology cannot think. They fail because the organizations deploying them cannot govern. Flip that finding over and it becomes the opportunity: the differentiator is not the technology anyone can buy. It is the design discipline you control. And this is now squarely a shared services problem. SSON Research and Analytics reported in February 2026 that the industry has moved past asking whether autonomous agents matter, with about a third of organizations already housing generative and agentic AI initiatives directly inside GBS. The live questions are pace and control. Which is to say: architecture.

If you are designing a bestshoring solution today, AI has already changed your inputs. The volume threshold that used to justify a 200-person center may now justify 120 people and an automation layer. Functions that defaulted to a provider may stay captive because technology closed the capability gap, or the reverse. These are good problems. They become expensive problems in three specific ways I keep seeing, and each one has a governance answer.

The bot in place of the center. When a vendor offers agents instead of a facility, they are implicitly making your delivery-model choice for you. That is not a knock on the vendor; selling capability is their job. Sequencing is yours. The Bestshoring Architecture™ exists to prevent exactly that inversion: technology and AI form an enablement layer that wraps every tier of the structure, location strategy, delivery model, and operating model. An enabler of each. A substitute for none. Held in that sequence, a bot layer can be a legitimate delivery choice, evaluated against a center or a provider on readiness and risk. Allowed to jump the sequence, it is an unpriced bet that skips the questions governance exists to ask: what may the automation own, what may it touch, and who answers when it is wrong.

The contract that pays for heads. Most third-party BPO agreements still price on full-time equivalents, and an FTE-priced contract quietly pays your provider to protect headcount. Every process the provider automates shrinks its own invoice. So automation arrives as theater: pilots, dashboards, roadmap slides, while the roster stays intact. The corrective is to treat the contract itself as your primary governance instrument, and greenfield builders hold a structural advantage here. At Shared Services and Outsourcing Week this spring, the most interesting commercial conversation in the room was the shift toward outcome-based structures, where the provider shares in the savings automation creates instead of billing for the bodies it replaces. I will be honest about the state of play: I have not yet met an organization that has fully implemented one. That is precisely the opportunity. If you are signing a new relationship this year, you get to write what the industry is still debating: automation gain-share, transparency and audit rights over the provider’s use of AI on your processes, reversibility and data-exit terms, all priced at signature instead of negotiated in a crisis. The providers worth signing will welcome that conversation.

The operating model without brakes. Greenfield builds run on momentum: location scouting, entity setup, recruiting waves, transition plans. Governance gets a placeholder box on the org chart, to be filled in once things settle. Things do not settle. Then the first AI-assisted process goes live and nobody can say who has the authority to stop it. The brakes belong in the target operating model before the first hire: decision rights over what gets automated and in what order, named halt authority, exception ownership with teeth, data ownership settled while it is still cheap to settle. Retrofitting these after go-live costs multiples, and it costs something harder to recover: the business’s confidence in the center you just built.

Most leaders are not starting clean. I spent the better part of three decades inside operations like these, so this comes with respect. You inherited an operating structure: location decisions made a decade ago under different trade conditions, delivery models assembled acquisition by acquisition, exception paths that grew like scar tissue around old incidents. AI does not care about any of that history. It accelerates whatever structure it lands on, and in an existing bestshoring operation, four patterns decide whether that acceleration compounds or collapses.

Automating the unready process. A forwarder automates rate quoting on top of tariff data that three teams maintain three different ways. The AI does exactly what it was asked to do: it produces quotes at machine speed, including the wrong ones, now delivered to customers with confidence and a timestamp. The brake here is readiness gating. Automate what is standardized and measured, not what is easiest or most annoying. Which processes are actually ready is a governance decision, and it belongs to you, not to your tool vendor’s roadmap.

The exception that crosses a boundary. An agent misreads a commercial invoice in shipment file processing, and the error propagates into a customs filing before a human ever touches the file. Meanwhile your provider’s SLA rewards throughput and turnaround time, so the exception that matters most is the one the contract measures least. The brake is exception ownership redesigned to cross the boundary: who catches it, who owns it, who can halt the flow, on whose authority and in what timeframe, written into the operating model on the captive side and into the agreement on the provider side.

The protective stack. Here is a pattern I see constantly: a model that needs tuning weekly sits behind a change process built for quarterly ERP releases, six weeks of sign-offs for a threshold adjustment. That is governance built for one job, protection, doing the job so thoroughly it strangles the capability it was meant to protect. When I led a discussion group on governance in the age of AI at SSOW in Orlando this spring, this was the distinction the room kept circling, and it is the question I have been putting to leaders for over a year: is your governance framework an engine or a brake? The rooms that get furthest reach a reframe first: brakes were never the opposite of speed. Brakes are what make speed usable. The organizations that answer well design governance as a steering system. Tiered decision rights, where operators adjust within defined bounds and committees govern the bounds themselves. A governance cadence matched to the automation cadence. Steering, not stacking.

The ungoverned dividend. This one is the opportunity. In my conversations across the industry, even complex, judgment-heavy processes are seeing 20 to 25 percent augmentation from AI. In a 400-person operation, that is a hundred people worth of released capacity. Ungoverned, the dividend evaporates: quiet backfill, unmanaged attrition that takes your best exception handlers first, the twenty-year people who can smell a bad file before any system flags it, savings claimed on a slide and never found in a ledger. Governed, it becomes the scope expansion engine, released capacity redeployed deliberately into the analytics, optimization, and decision-support work that shared services organizations have spent a decade trying to claim. This is not theoretical. SSON’s Impact Awards documented Michelin cutting invoicing errors by 22 percent while nearly tripling digital adoption through governed deployment. The dividend is real. Whether you capture it is a design question.

The asymmetry deserves plain statement, because it changes what governance means in your bestshoring model.

In a captive or GCC, you own the brakes end to end. You can see the model, the data, the exception queue, the operator. Your risk is layering: enterprise governance habits imported wholesale until the center cannot adjust a threshold without a committee. Your governance work is calibration. Steering over stacking.

In a third-party relationship, your brakes are contractual. You cannot walk the floor of your provider’s model governance, and your risk is misaligned incentives and opacity: automation you cannot see, making decisions you still answer for. And you do answer for them. In 2024, a Canadian tribunal held Air Canada liable after its website chatbot invented a policy that did not exist; the company argued the bot was effectively a separate entity. The tribunal called that “a remarkable submission” and rejected it outright. The damages were small. The principle was not, and it transfers cleanly into our world.

Regulators are formalizing that logic. The EU AI Act places obligations on the organizations that deploy AI systems, not only the companies that build them. Even its own rollout carries the pace lesson: EU lawmakers deferred the high-risk requirements sixteen months, to December 2027, because the oversight infrastructure was not ready at the original speed, while transparency obligations still take effect on August 2, 2026. The world’s most ambitious AI regulator matched its speed to its steering. NIST’s AI Risk Management Framework says the same in a voluntary key: of its four functions, the one that applies to everything is Govern. Who approves. Who owns. Who can intervene. For the organization that designs its brakes on purpose, none of this is burden. It is tailwind.

The organizations that will own the next five years of this industry are not the ones with the most horsepower. They are the ones that decided how fast they could go while keeping control of the car, and then built the steering to sustain that speed.

That is architecture work, and there is a method to it. Diagnose first: the Six Dimensions of Bestshoring Readiness™ treat governance and control as a named dimension of readiness rather than an afterthought, alongside Digital and AI Readiness, so you can see exactly where the gaps sit before they compound. Architect second: place the brakes deliberately, in the operating model tier for a captive, in the commercial instrument for a provider relationship, with governance designed as optionality levers that create room to move rather than reasons to wait. Scale third: govern the dividend, redeploy it into scope, and let your governance pace be what it was always meant to be. Not a ceiling imposed by committee. A choice you made on purpose.

The fast car is already in your garage. Whether you are building your bestshoring solution or scaling the one you already run, the engine was never the question.