Most AI roadmaps start with the same slide: a bold promise of “2 or 3x productivity” or “automating entire workflows” to transform entire departments. We’ve heard it so often now, we may be tempted to roll our eyes. But here's the thing: that's not wrong. It's just incomplete.
As organizations start to make good on the roadmap, they begin to discover the gap between demo and deployment: the model needs guardrails, the data needs tidying, that killer use case from the keynote turns out to be 80% edge cases, and they haven’t figured out how to ground AI without context bloat. What we need to remember is that this is simply the normal learning curve of any incredibly powerful technology.
AI as a technology is ready, but are we?
The teams seeing real transformation didn't expect magic. They understood that beyond hype, AI is any other technology with vast potential that must be grounded in reality.
If we stop treating AI as a magical truth engine and start treating it as what it actually could be, which is a very fast, very powerful, but occasionally literal-minded digital colleague, a different set of goals appears. One that’s much more realistic, and paradoxically, much more transformative.
Read on and learn how to design AI enterprise transformation based on what’s achievable right now.
Seven steps to designing enterprise AI for real-world impact
1. Stop asking “Where can we use AI?” and start asking “What can we safely delegate?”
Most AI initiatives begin focusing on the wrong question:
“Where can we use AI in our business?”
It sounds strategic, but it invites grand aspirations with over-engineering that may fail to live up to requirements.
A better starting point is a more focused, although potentially more uncomfortable question:
“If you could wave a wand and never do three tasks again, what would they be?”
When teams answer honestly, you get very specific tasks:
- Tedious research across many different data sources
- Lack of credible or stale information
- Summarizing long documents into the same three bullets
- Re-writing the same narratives over and over again when evaluating peers
- Identifying the latest relevant news and determining if and how my portfolio is impacted
- Running pro-forma analysis
- Copy-pasting data between systems
- Reconciling slightly inconsistent numbers from financials or standardizing reporting units or currencies
- Triaging inbound requests
- Checking for obvious inconsistencies
Those tasks are the building blocks of real AI delegation.
They describe workflows that:
- Have a clear goal (decide X, extract Y, check Z)
- Are highly repetitive
- Require low to moderate cognitive load for a human
- Are too dynamic for old-school RPA (robotic process automation), but structured enough that “good data in can lead to reasonable decision out” is possible
This is the real “AI sweet spot” or “the low-cognition, high-repeat work band” for most enterprises today: taking the grinding, low-variety work off people’s plates so they can focus on the parts of the job that requires judgment, strategy and debating with other humans. Automate the mundane.
2. Raise the floor of decision quality before you chase the ceiling
Most AI messaging leans on “superhuman performance”: better forecasts, better research, better insights, better anything.
In practice, especially in complex domains like banking and risk, the first big win should be much more controlled.
Make the worst version of a task significantly better and do it more often.
Think about a junior analyst on a tight deadline:
- They might miss a footnote where management quietly redefines a key metric.
- They might mix up units (“thousand” vs “million”), shifting numbers by orders of magnitude.
- They might ignore three data points that contradict a narrative because there isn’t time to reconcile them.
This is where LLMs excel. They are extremely good at:
- Scanning many more documents than a human has time for
- Checking consistency across sources (dates, units, entities, trends)
- Applying first-principles sanity checks:
- “Does this leverage figure match the balance sheet and income statement?”
- “Does this growth rate align with macro conditions?”
- “Is this claim actually supported by the citations attached to it?”
Moody’s previous write up on AI in banking makes this explicit: without accurate, consistent, well-governed data, even the best models will produce misleading insights and hallucinations, and increase compliance risk rather than reduce it.
If you invest in the data foundation and treat models as relentless reviewers instead of oracles, something powerful happens:
- The baseline quality of every memo, every investigation, every forecast goes up.
- Obvious errors get caught early and quietly.
- Senior employees spend less time fixing preventable mistakes, and more time arguing about the truly hard and strategic questions.
That’s not as headline driven as “AI outperforms experts,” but it’s far more achievable and far more valuable in year one.
3. Use AI to redistribute attention, not just “save time”
Time saved is the easy metric. “We reduced manual review time by 59%” is a great case-study line and real systems already deliver numbers like that on abuse-handling and triage workflows.
But the more interesting question is:
“What did my team do with the reclaimed time?”
In a well-designed agentic workflow, the answers should be:
- More time to review edge cases
- Investigate more atypical and potentially impactful anomalies
- Spend more time on cross-disciplinary decision making
Take a memo generation agent as an example:
- The agent does the grunt work: news archaeology across credible licensed news sources, web search on company websites or government sources, crawl through annual reports or interim quarterly updates, read through macroeconomic forecasts or review company structures.
- It applies a structured and formulaic approach to finishing the majority of the write up
- A RM can then spend time to review the write up, add their own narratives on top instead of writing everything from scratch. That time saved is additional time with clients.
You haven’t removed the human. You’ve changed their job:
- From “researcher + typist”
- To “editor + strategist + pattern-spotter”
That’s a better job. It’s also a job where experience compounds and where your best people have more surface area to apply their judgment.
Efficiency is the visible surface level and redistributed attention is the real structural shift.
4. Make peace with the fact that LLMs are not truth engines
A recurring failure mode in AI projects looks like this:
- “We’ve integrated our data lakehouse with an LLM.”
- “We’ve wired up some dashboards.”
- “Therefore, we now have an intelligent assistant that always tells us the truth.”
We don’t.
LLMs are pattern machines. They are astonishingly good at generating plausible continuations of text conditioned on the data and instructions you give them. They are not (and have never been promised to be):
- A guarantee of factual accuracy
- A substitute for missing data
- A license to skip governance, lineage, or documentation
The Moody’s-aligned view in financial services is blunt: good AI without good data is a risk, not an asset. If you feed models inconsistent, unstandardized, poorly governed datasets, they will:
- Confidently hallucinate correlations that don’t exist
- Hide data quality issues under fluent prose (the smooth lie)
- Make it harder, not easier, to trace how a decision was made
So a realistic goal for enterprise AI is:
Build feedback loops
That means:
- Small, low-risk projects to validate data and model behavior before scaling.
- Transparent schemas and entity IDs so every number can be traced back to a source.
- Multi-dimensional evaluation: accuracy, feasibility, compliance, cost. Not just “did the model answer something.”
- Human-in-the-loop by design
An agent that routes every recommendation through a human approval step in your internal tooling is not “unfinished.” It’s acknowledging reality: you are layering statistical reasoning on top of an imperfect world. Oversight must be a feature.
5. Exploit what AI is uniquely good at: crossing silos and checking assumptions
Humans are very good at deep, narrow reasoning within a domain.
We are less good at:
- Holding dozens of weak signals across different disciplines in working memory
- Remembering every caveat attached to every data source
- Doing the same first-principles check for the 200th time with equal care
AI agents, when wired correctly, are weirdly good at exactly this as we’ve stated in point 2.
A single workflow that can:
- Retrieve financials data from a core system
- Join it with implied probability of default models
- Overlay macro forecast indicators or policy changes for a region
- Cross-check a narrative against external news, filings, and reference data
- Flag: “This story doesn’t reconcile with the numbers; here are three contradictions.”
Specifically in risk and banking:
- Cross-checking a bank’s stated “sector exposure” with external datasets to see if it’s consistent.
- Stress-testing a company’s rosy projections against macro scenarios and sector utilization thresholds.
- Flagging that a borrower’s “improving” metrics are almost entirely subsidy-driven, while core operating cash flow is flat.
These are things a human analyst can do but not for every file, every time.
The realistic goal isn’t about “AI finds all the hidden truths” but about doing more of what’s achievable:
“For any important decision, we want one brutally honest, cross-silo sanity check before it reaches a committee.”
If you can automate that sanity check and make it cheap enough to run on everything, you quietly upgrade the entire organization’s signal-to-noise ratio.
6. What “success” actually looks like in year one
If we strip away the hype, what are reasonable goals for an enterprise AI program over the next 6–12 months?
Something like this:
- A portfolio of agentic workflows
- Each focused on a narrow set of verbs (triage, summarize, check, qualify, reconcile).
- Each with a clear owner and success metric.
- Measured ~30–70% time reduction on targeted tasks
- Not across the whole org but just in the specific workflows you redesigned.
- Documented via time-and-motion or ticketing data, not vibes.
- Decision quality that is “at least as good as humans,” not worse
- Conversion rates or risk outcomes stay flat or improve slightly; they don’t need to be significantly better
- Because when AI replaces a large chunk of manual work, your win is the speed and scale to value as long as your key performance curves can stay where they are.
- A visible reduction in “preventable errors”
- Fewer unit mix-ups, fewer missed footnotes, fewer inconsistent numbers across decks.
- More coherent stories that reconcile with source systems.
- A cultural shift in how work is described
- People start saying “I’ll have the agent draft it and then I’ll review” instead of “I’ll brute-force it myself.”
- Teams become more explicit about what they want to stop doing, not just what new technologies they want to adopt.
- A clearer understanding of your data reality
- You discover which datasets are actually production-ready and which are aspirational.
- You invest in sourcing, quality, standardization, transparency, and governance
None of this requires sci-fi autonomy. It requires clear-eyed understanding about what models can and can’t do today, and ambition about redesigning work around those capabilities.
7. A different slogan for enterprise AI
If we had to summarize this mindset in something that isn’t “it’s not about efficiency, it’s about transformation,” it might sound more like:
“Be precise about what you’re willing to delegate, and relentless about how you measure it.”
That means:
- Start with the tasks people hate, not the use cases that sound impressive on stage.
- Design agents to raise the floor of quality and coverage.
- Use the model’s superpower in breadth and cross-silo pattern-spotting to challenge your assumptions.
- Treat data quality, context and governance as first-class citizens
- Keep humans in the loop where stakes are high, and use their feedback as the engine of continuous improvement.
The long-term potential of AI in enterprises is enormous. But the path there is surprisingly grounded, and it must be taken one clearly defined task, one well-designed workflow, and one honest metric at a time.
About the author:
David Pan is a Director - Industry Practice Lead for Asia Pacific at Moody’s and is responsible for exploring innovative applications of Moody's data exposed through GenAI. He advises organizations across the region on distinguishing hype from reality, identifying practical use cases, and guiding effective adoption strategies.
Before joining Moody’s, David led generative-AI solution design, development handbooks, and supervised production deployments that drove measurable business impact. He has held leadership roles in Professional Services, Solution Architecture, and Business Development across financial crime & compliance, fraud & identity, and data science consulting.
David currently holds an Executive MBA at INSEAD.