Nobody is asking anymore whether AI is worth adopting. That question is settled. The harder question, the one keeping IT leaders up at night, is operational: how do you keep AI from turning into a pile of fragmented tools, duplicated effort, uneven adoption, and governance gaps once it moves past the pilot stage.
That shift from access to operating model was the center of a recent episode of the Amplitude of Tech podcast, in conversation with Ed Keisling, Chief AI Officer at Progress Software, one of the first people to hold that title anywhere in enterprise software. Ed walked through how Progress scaled AI culture from a handful of early adopters to more than 1,200 people using AI tools daily, why the company’s original 20-page AI policy nearly killed adoption before it started, and why inference costs going from effectively free to unpredictable is the budget problem nobody planned for.
| Quick Answer: An AI governance framework works best when it is built before a pilot proves itself, not after. Progress Software’s Chief AI Officer Ed Keisling scaled enterprise AI adoption by pairing a Vanguard cohort (fast-moving experimenters with a larger inference budget) with a Champions cohort (regional and departmental scaling), rewriting AI policy from restrictions to permissions, and treating shadow AI usage as a signal of tooling gaps rather than only a risk to shut down. |
The AI Adoption Curve Has Moved From Access to Operating Model
A year or two ago, most executive conversations about AI started with access: who is allowed to use it, which tools are approved, what data can go in. Those questions still matter, but they are no longer the whole conversation.
At small scale, informal coordination works. Somebody messages somebody else and it gets sorted out. At the scale Ed described at Progress, with 1,200-plus people using AI-enabled development tools, copilots, and internal agents, informal coordination becomes a liability. Somebody has to actually answer the cross-functional questions: which use cases get funded, which capabilities get reused instead of rebuilt, which teams are quietly building something that already exists somewhere else in the business.
Vanguards and Champions: How Progress Software Scales AI Culture
Rather than governing every team the same way, Ed described splitting AI experimentation into two distinct lanes.
Vanguards: Room to Push the Edges
The Vanguard cohort at Progress is a small group of the most forward-leaning AI users, given the organization’s largest inference budget and the most latitude to experiment. Their job is to find the actual limits of what AI can do inside the business, in a controlled environment with direct access to leadership and IT for support. When something breaks, or a new pattern proves out, that learning feeds back into how the rest of the organization works.
Champions: Scaling What Already Works
The Champions cohort solves a different problem: getting proven practices to every region and department without losing what makes them effective locally. Progress operates in offices across Bulgaria, the Czech Republic, India, Ireland, and beyond, so a single global mandate rarely lands the same way twice. Champions translate general guidance into something bespoke enough for their region to actually stick, then follow through to confirm it did.
Shadow AI Is a Signal, Not Just a Risk
Every organization has some version of shadow AI, employees reaching for tools outside what IT has sanctioned. The instinct is to treat that purely as a compliance problem, and in some cases it is: access to certain models has to be blocked outright for legitimate security reasons.
But Ed’s team also reads shadow AI usage as a diagnostic. If a company rolls out one tool expecting it to solve a problem, and employees keep reaching for a different tool instead, that gap is information. It usually means either the sanctioned tool has a real limitation, or the organization failed to enable people well enough to get full value out of what they were given. Both are fixable, but only if shadow AI gets investigated instead of just blocked.
Governance Has to Show Up Before the Pilot, Not After
One of the clearest lessons from the episode: Progress’s original AI policy ran about twenty pages, and it backfired. Employees read it, got nervous about the consequences of getting something wrong, and the safest option became doing nothing at all.
The fix was rewriting the policy from a list of prohibitions into a list of permissions: here is what you can build, here are the controls already in place to experiment safely, here is the process if you need more. That single reframing, from restriction to enablement, changed adoption more than any technical rollout did.
Ed also pointed to a related discipline: matching the level of review to the level of risk before building starts, not after a demo has already generated executive excitement. An internal assistant summarizing public documentation is a different animal from an agent that can touch customer data or trigger a production workflow. Deciding which tier a project falls into up front avoids the expensive rework that happens when governance shows up only after a team has sprinted in the wrong direction.
The Real Budget Problem: Inference Costs and Model Routing
For most of the last two years, AI usage inside Progress was effectively free from a cost-tracking standpoint. That changed quickly once providers like Microsoft and Anthropic moved from subsidized, flat-rate access to metered, usage-based pricing. Teams that had built habits around unlimited access to the most capable models suddenly faced bills that made those habits unsustainable.
Ed’s team responded by building routing logic that matches the model to the task rather than defaulting to the most powerful option available: routine work like writing unit tests goes to a smaller, cheaper model, while genuinely complex reasoning gets reserved for a premium one. They also started assigning inference budgets by team and project instead of leaving usage open-ended, and built tooling that manages context size, since repeatedly resending full conversation history to a model quietly drives up cost with no added value.
Where CIOs and CTOs Should Start
Ed’s advice for leaders earlier in this process was direct: treat AI as a participatory discipline, not a topic to plan in meetings before anyone touches the tools. Learning happens by doing, and by failing fast enough to correct course.
From there, the practical steps mirror what worked at Progress: separate a small cohort of fast-moving experimenters from the broader workforce that needs consistency; rewrite AI policy around what people can do rather than what they cannot; treat shadow AI as a source of signal about tooling gaps; and build cost visibility and model routing into the architecture before usage scales past the point where it is easy to control.
Frequently Asked Questions
What is an AI governance framework?
An AI governance framework is the set of principles, guardrails, and decision rights that determine how an organization builds, deploys, and monitors AI. It covers what data models can access, which use cases require human approval, and how reuse and risk get managed before a project reaches production.
Do you need a Chief AI Officer to scale AI?
Not necessarily as a dedicated title. What matters is clear ownership of AI coordination, reuse, and governance across business units. Some organizations assign this to a Chief AI Officer, while others build the function into program management or a cross-functional AI council.
What is shadow AI and why does it matter?
Shadow AI refers to employees using AI tools outside of officially sanctioned platforms. Beyond the compliance risk, shadow AI usage patterns often reveal gaps between the tools a company provides and what employees actually need to do their jobs.
How should enterprises manage AI inference costs?
Enterprises should route routine tasks to smaller, less expensive models and reserve frontier models for genuinely complex work, set usage budgets by team or project rather than allowing unlimited access, and build visibility into which use cases are driving token consumption before costs scale unpredictably.
What is an AI Center of Excellence?
An AI Center of Excellence is a centralized function, often paired with a hub-and-spoke model of regional or departmental champions, that sets standards for AI use, promotes reusable components, and helps the organization scale proven practices instead of letting every team build in isolation.
Build Your AI Governance Framework With Amplix
Scaling AI without governance is how enterprises end up with sprawl instead of results. Amplix helps organizations build the operating model, policy, and cost controls that let AI scale safely, drawing on the same lessons Ed Keisling shared about culture, reuse, and inference budget planning.
Contact Amplix today to talk through your AI governance and adoption strategy.
For more on how enterprises are approaching AI at scale, see our related insight on enterprise AI orchestration and explore Amplix’s AI capabilities.
