Artificial Intelligence
An Agent Walks Into a Bar
Enterprises are racing to give AI agents more access, but access isn't the bottleneck—clean, current context is.
Julie Colwell
Principal Strategist
Workday
An AI agent walks into a bar. And tries to expense the entire tab.
It's there on behalf of an employee who just wrapped a team dinner.
- It sees a calendar invite titled "Team Dinner — Acme Deal"
- It sees a receipt for $1,240
- It has a standing instruction to "handle expenses automatically"
So it does exactly that. It approves the whole tab, tip included, no questions asked.
The catch: Two of the six “team” members were contingent workers, whose per-diem is capped lower than full-time staff. The policy that actually applies isn't the global one the agent pulled from a shared drive — it's a regional addendum finance updated 11 weeks ago, sitting in a different system the agent never checked. And anything over $500 needs a director's sign-off before it's paid, not after.
The twist: The agent didn't hallucinate a policy or botch the math. It acted on the context it had and ignored the context it lacked. That's the whole problem with agentic AI right now.
Report
AI Agents Are Here—But Don’t Call Them Boss
The agent didn’t hallucinate a policy or botch the math — it acted on the context it had and ignored the context it lacked.
Why it matters
Every enterprise is about to hand agents keys to real decisions — expenses, approvals, headcount, payroll. Those decisions are only as good as the information behind them. And right now, most enterprise information is scattered, contradictory, and stale.
By the numbers:
- 95% of generative AI pilots inside businesses never reach sustained production use, per MIT NANDA's State of AI in Business 2025 report — driven largely by systems that can't learn, adapt, or integrate with real workflows, not by weak models.
- 82% of employees say they spend significant time acting as translators between systems that don't talk to each other, according to Workday's 2026 research surveying 6,100 finance, HR, IT, and ops professionals.
- 43% say their days feel busy, but not productive. Just 27% of organizations have embedded AI into the core systems that run the business.
- Employees at companies with AI embedded in core systems are more than twice as likely to report time savings of 25% or more, per the same Workday research.
The pattern is consistent across the research: it isn't that agents lack access to information. It's that nobody has done the work of making that information clean enough to act on.
Zoom in: quantity isn't the solution
It's tempting to think agents fail because they don't know enough. Usually, it's the opposite.
Researchers have a name for this now: context rot. As the information an AI system has to sift through grows, the odds it weighs the wrong parts, or misses the one detail that mattered, go up, not down. Chroma's research on context rot found performance degrading across every major model as input length increases, independent of whether the task got harder.
A related, well-documented failure mode: the “lost in the middle” effect, where models pay disproportionate attention to information at the very start or end of what they're given — and are more likely to overlook something buried in the center. Bury a regional meal-cap addendum in the middle of a 40-page policy PDF, and there's a real chance it never gets weighed.
The emerging discipline for fixing this is called context engineering — treating what an agent sees as something to be curated, not just collected. The core idea, laid out in a widely cited industry write-up on the practice: the goal isn't feeding a model more information, it's finding the smallest set of information that actually gets the job done. That's a different design problem than the one most enterprise software has been solving. For 20 years, the instinct was to capture everything, make it searchable. Agents punish that instinct instead of rewarding it.
The real failure point
A related body of research — including a recent survey on agent reliability — makes the same point from a different angle: agents rarely fail in isolation. Their behavior is shaped entirely by the instructions, memory, tools, and retrieved information in their context. When that context is weak, agents drift off task, act on stale information with total confidence, or apply the wrong rule to the right situation.
None of that shows up in a demo. Demos are, almost by design, the one moment an agent's context is curated and clean. Production is real employees, real regional variation, real exceptions, and data that was accurate six months ago and isn't anymore. That gap is where agentic pilots go to die — and it's also why Gartner expects a sizable share of agentic AI projects to be shelved within the next two years.
This isn't unique to a team dinner. The same setup could play out in the finance department closing the books, in a benefits enrollment window, in a vendor renewal. Wherever an agent is asked to act, there are really only two paths. Give it clean, current, unambiguous context, and it knows exactly what to do — apply the right rule, flag the real exception, ask the one clarifying question that matters.
Or leave it working off a patchwork of scattered, contradictory information, and it approves the wrong thing, defers to nothing, and asks no questions.
Access ≠ context
There's a version of this problem that looks solved but isn't: an agent technically connected to everything — the HR system, the policy wiki, last quarter's spend data.
Access is no longer the constraint most organizations are fighting. Quality is.
Cognizant ran into a real version of it managing payroll in Australia, where companies must report tax and superannuation details to the tax office within 24 hours of every payroll run — a compliance window with essentially no room for context that's stale or split across systems.
It used to take Cognizant's team nine days to close a pay run, much of it spent on manual reporting and journals to reconcile what one system knew against what another needed. Consolidating payroll onto a single connected system cut that to two days because the right context finally lived in one place instead of being stitched together by hand.
Access is no longer the constraint most organizations are fighting — quality is.
Back to the bar
The agent had access to a meal policy — just not the current, region-specific one. It had access to attendee data — but no way to tell an intern from a full-time employee inside a calendar invite. It had access to an approval workflow — but no signal this expense needed to route through it before payment, not after.
Every piece of context it needed technically existed somewhere in the enterprise. None of it arrived clean, current, or attached to the decision at the moment the decision was being made.
Connecting an agent to a data source is not the same as giving it context. Context is judged, structured, and current. A policy document isn't context until someone — or something — has resolved which version is authoritative, tied it to the right role and region, and defined what happens when the facts don't fit neatly.
Connecting an agent to a data source is not the same as giving it context.
Bottom line
Give that same agent clean context — the current regional policy, correctly attached to role and headcount; a clear signal on required approvals before payment; an unambiguous read on who was in the room — and the outcome flips. It flags the two intern per-diems. It routes the total for sign-off before the card is charged. It leaves the right tip.
The agent didn't get smarter. Its context did.
The agent didn’t get smarter — its context did.
The organizations that win with agentic AI won't necessarily be the ones with the most advanced models. They'll be the ones that did the essential work of making their enterprise context worth acting on.
Report
AI Agents Are Here—But Don’t Call Them Boss
A gap between effort and impact is the real ROI problem. Discover the solution.