Autonomous agents run quietly around the clock and fresh upgrades land almost weekly. Leaders must rethink how their AI architecture is built to keep pace and protect their budget.
Sydney Scott
Editorial Strategist, AI
Workday
IT leaders have spent decades planning around biannual rollouts, but AI never clocks out.
"These are not applications that come out every six months, and then you get another version six months later,” explained Workday CEO Aneel Bhusri at Rising last month. “AI applications iterate on a weekly basis."
Autonomous agents give organizations an always-on workforce that delivers great productivity gains. But, it also means software runs around the clock, quietly piling up usage fees by the minute.
A static budget risks stalling an AI initiative before it proves its value or leaving teams stuck with yesterday’s tech. Unlocking the full value of AI begins with rethinking how agent architecture is built and governed.
"These are not applications that come out every six months, and then you get another version six months later. AI applications iterate on a weekly basis." —Aneel Bhusri, CEO, Workday
Report
If AI applications are constantly iterating, the smartest engineering move is making sure every piece can be swapped out.
Core company software works like a sturdy office building, and agents are like short-term tenants. An AI tool packs up and leaves when a better one comes along, but the building stays rock-solid.
It starts with a strong data foundation. Business logic, proprietary data and security rules belong in the system of record, not inside any one agent. An agent plugs in from the outside, does its job, and steps aside when a better option comes along. The agent can be unplugged without anyone rewriting the rules underneath it.
The same swappable design applies beneath the surface. Instead of anchoring an agent to a single model provider, modular architecture lets you point the agent to whichever commercial or open-source model currently delivers the best performance.
An AI tool packs up and leaves when a better one comes along, but the building stays rock-solid.
Before handing the keys to a new AI tenant, make sure it earns its keep. Teams can and should demand a trial run. Side-by-side tests against a company’s current setup can confirm accuracy and speed before anything changes—the way any critical system is tested before going live.
Then, make sure the checking continues. The market won't wait for the next annual review, so design decisions should be revisited every few months. That's a standing practice, not a one-time cleanup.
Lastly, every agent should get the least access it needs and nothing more. An agent should see and carry only the data its task requires. That makes switching tools painless and mistakes contained.
If an AI agent controls its own off switch, that off switch does not exist.
At Workday Rising, industry analyst Ray Wang stressed that governing controls must sit outside the system the agent operates in to guarantee clean operational boundaries, independent monitoring, and zero chance of an agent overriding its own limits.
He paired this guidance with a critical requirement for vendor RFPs: "Can you support reversibility of an action?"
Pausing an agent is only half the equation. When an agent acts on incomplete data or makes an erroneous call, teams need the ability to roll back the action, trace the audit trail step by step, and restore clean records.
If an AI agent controls its own off switch, that off switch does not exist.
Continuous systems run best with active, in-flight monitoring. Autonomous agents quietly execute tasks in the background, which makes instant visibility the best way to protect operating margins.
This was put to the test during a live demo at Rising, where an automated alert flagged an organization's agent spend. The presenter drilled down to find the source, then set up an alert that would pause the agent if it went over its allowed usage. The entire adjustment played out like a routine operational check rather than a fire drill.
That proactive control represents the enterprise standard for lawful AI. Standard vendor invoices report consumption long after spend is incurred—and long after there’s any chance to stop it. Real-time control lets you hit the brakes before it’s too late.
Pharmaceutical giant Sanofi saw how a modular system and clear budget tracking can pay off. Eight months into a two-year project, better AI tools popped up that were never in their initial cost forecast. Instead of scrapping the project or locking themselves into outdated tools, Sanofi worked with tech partners to shift on the fly, while staying within their budget.
Sanofi could take advantage of new AI upgrades mid-cycle because of how they set up their budget and architecture: with clear cost tracking and flexible tools from day one.
Leaders can learn two lessons from Sanofi. First, track what every digital worker costs and produces in real time. Second, build a plug-and-play architecture so swapping out an outdated agent takes minutes, not months. Doing both keeps work moving fast while costs stay right where finance wants them.
Real-time budget tracking and automated controls give leaders confidence in their investment, while modular design keeps systems flexible enough to adopt tomorrow’s best tools.
Real-time budget tracking and automated controls give leaders confidence in their investment, while modular design keeps systems flexible enough to adopt tomorrow’s best tools.
Teams ready to scale can think about four questions:
1. Can we see credit consumption per agent today?
2. Do we have automated alerts, and can an agent be paused automatically at its limit?
3. Does our kill switch sit outside the system it controls?
4. If a better model or agent appeared next quarter, could we swap it in without touching our core systems?
If the answer to any of these is no, that's where to start.
Report
A gap between effort and impact is the real ROI problem. Discover the solution.