An interview with Michael Privat
Most enterprise AI conversations get stuck in one of two dead ends: agents are either failing spectacularly or about to automate everything. Michael Privat, Chief Data and Engineering Officer at Availity — where he leads a global team of 500+ engineers — argues both narratives miss the same critical variable: workflow reality. In this interview, Privat cuts through the hype with a practitioner’s precision, explaining exactly what makes a workflow agent-ready, why most pilots never survive contact with production, and how organizations can govern AI output that now accumulates faster than humans can review it.
Top 3 Takeaways
- Agents amplify what already exists — good or bad. A well-designed workflow becomes dramatically more efficient with agents; a chaotic one becomes chaotically faster. Discipline in process design is a prerequisite, not a follow-on.
- Workflow selection is a strategic decision, not a technical one. The best agent candidates are high-repetition, low-ambiguity processes with measurable outcomes and clear escalation paths. Avoid anything that relies on tribal knowledge, interpretation, or undocumented judgment calls.
- A successful pilot proves operational reliability — not that the demo worked once. Real pilots measure outcomes, maintain observability, and put humans in the escalation loop from day one. If you can’t run it in production under governance, you don’t have a pilot — you have a proof of concept.
The Michael Privat Interview
Enterprise AI leaders are hearing two competing stories: agents are either failing or about to automate everything. What is missing from both narratives?
Both narratives are missing the same thing: workflow reality.
The people saying agents are failing are usually pointing at workflows that were never good candidates in the first place. Too much ambiguity. Too much undocumented tribal knowledge. Too many edge cases hiding inside someone’s inbox or inside the head of the one employee who has been doing the process for fifteen years.
The people saying agents will automate everything are making the opposite mistake. They assume intelligence alone is enough. It is not.
An agent is only as effective as the environment it operates in. If the workflow is already disciplined, observable, and well understood, agents can create enormous leverage. If the workflow is chaos held together by tribal knowledge and Slack messages, agents amplify the chaos too.
What gets lost in the hype cycle is that enterprise value does not come from demos. It comes from operational reliability.
You argue that workflow selection matters more than the agent technology itself. What makes a workflow agent-ready?
A workflow becomes agent-ready when the decision space is narrow enough that success and failure are objectively measurable.
The biggest misconception right now is that people think agents are about intelligence. In practice, agents are mostly about process discipline.
The best agent workflows usually have a few characteristics:
- High repetition
- Clear inputs and outputs
- Stable rules
- Observable state changes
- Low ambiguity
- Large amounts of navigation or coordination work
- Human escalation paths when confidence drops
A perfect example from healthcare IT is claims investigation work. Someone is clicking through multiple systems trying to answer a simple operational question: “Where is the payment receipt?” That is not deep strategic thinking. It is structured detective work across fragmented systems. Agents are extremely good at that.
Where organizations get into trouble is trying to automate workflows that still depend heavily on interpretation, politics, negotiation, or undocumented judgment calls.
If two experienced employees would solve the same situation completely differently, the workflow probably is not ready yet.
What are the warning signs that a team is trying to apply an agent to a workflow that still depends too much on human judgment, interpretation, or undocumented practice?
The first warning sign is when nobody can explain the workflow cleanly without saying, “Well, it depends.”
The second is when your highest performers cannot articulate why they make certain decisions. They just “know.” That is intuition accumulated over years. AI can assist around that. It cannot reliably replace it yet.
Another huge warning sign is when exceptions dominate the process instead of the happy path. If your workflow requires constant human interpretation of edge cases, your problem is probably process design before it is AI deployment.
I also look for hidden operational dependencies. The employee who everybody quietly messages when something weird happens. The spreadsheet nobody officially owns. The undocumented workaround that exists because “the system has always been broken.”
Those are signals that the workflow is not actually formalized. It is being socially held together.
Agents struggle in environments where the real process lives outside the documented process.
Many organizations have agent pilots running, but few have scaled them into production. What separates a useful pilot from a pilot that only demonstrates technical possibility?
A real pilot proves operational reliability. A fake pilot proves the demo worked once.
Most pilots succeed under controlled conditions with clean datasets, engaged teams, and artificial guardrails. Production environments are different. Production introduces messy data, competing priorities, changing systems, bad inputs, latency, compliance constraints, and edge cases nobody modeled.
The organizations succeeding right now are not treating agents like magic. They are treating them like operational systems that require governance, observability, escalation paths, and measurement.
The useful pilots usually share three characteristics:
- They target a very specific workflow
- They measure operational outcomes, not impressions
- They include humans in the escalation loop from day one
Enterprise leaders do not need another chatbot demo. They need proof that something can reliably reduce cycle time, lower operational cost, improve accuracy, or eliminate repetitive labor without introducing unacceptable risk.
How should enterprise teams document a workflow before deciding whether an agent belongs in it?
Start by documenting what actually happens, not what the process document claims happens.
Those are often two completely different things.
I tell teams to map workflows at the decision level:
- What triggers the workflow?
- What systems are touched?
- What decisions get made?
- Which decisions are deterministic versus judgment-based?
- Where do humans intervene?
- Where do exceptions occur?
- What defines success?
- What defines escalation?
Then measure the exception rate honestly.
Organizations constantly underestimate how much of their workflow is really exception handling disguised as standard process.
I also want teams documenting where tribal knowledge exists. If critical operational logic lives only inside people’s heads, that is not an AI problem yet. That is a process maturity problem.
One of the most valuable things AI is doing right now is exposing how poorly documented many enterprise workflows actually are.
In healthcare IT, workflows often include compliance, clinical risk, fragmented systems, and human exception handling. What has that environment taught you about where agents fit and where they do not?
Healthcare teaches you humility very quickly.
You learn that operational complexity is usually much deeper than the workflow diagram suggests.
A healthcare workflow may look simple on paper, but underneath it are compliance constraints, payer-specific rules, fragmented integrations, contractual edge cases, human overrides, and clinical implications that change the risk profile completely.
That environment taught me two things.
First, agents are incredibly powerful for repetitive operational coordination work. Navigation across systems. Retrieval. Reconciliation. Status tracing. Structured follow-up. Workflow acceleration. That is real value.
Second, healthcare punishes false confidence brutally.
You cannot deploy agents into workflows where hallucinated certainty creates downstream clinical, compliance, or financial risk without strong control mechanisms.
The organizations that succeed will not be the ones with the flashiest AI demos. They will be the ones with the strongest operational discipline.
What role should process owners, frontline workers, compliance leaders, and engineering teams each play in deciding whether an agent should be deployed?
This cannot be an engineering-only decision.
Engineering understands system capability. They usually do not understand the full operational nuance of the workflow itself.
The frontline workers are often the most important voices in the room because they know where the real friction lives. They know which parts of the workflow are repetitive and which parts quietly fall apart every week.
Process owners define success metrics and operational accountability.
Compliance leaders define the acceptable risk boundaries and escalation rules.
Engineering builds the operational scaffolding around reliability, observability, auditability, and execution.
The organizations that fail are usually the ones where one group tries to dominate the entire decision.
This requires cross-functional design because workflows are cross-functional reality.
You note the risk that AI-generated output can accumulate faster than human review can catch it. Where does that risk show up first inside an organization? How should the risk be mitigated?
It shows up first in operational debt.
AI is dramatically increasing output velocity. Code. Documentation. Tickets. Summaries. Decisions. Work artifacts. Everything.
The danger is not that the output exists. The danger is that organizations lose the ability to meaningfully review and govern it at the same pace it is being generated.
We already see this in software engineering. AI can now generate code faster than many organizations can properly understand, validate, or maintain it.
The mitigation is not “slow down AI.” That is unrealistic.
The mitigation is:
- Stronger validation layers
- Better observability
- Smaller deployment scopes
- Automated testing
- Confidence scoring
- Human escalation paths
- Clear ownership boundaries
Most importantly, organizations need to stop assuming generated output is trustworthy simply because it looks polished.
Fluent output is not the same thing as correct output.
How should organizations think about control points, auditability, and escalation paths when agents are executing repetitive processes at scale?
The same way good engineering organizations think about production systems.
Assume failure is inevitable somewhere.
One of the most important lessons I learned over decades in healthcare IT is that “statistically unlikely” becomes operational certainty at scale.
Agents need:
- Observable execution logs
- Decision traceability
- Reversible actions
- Human escalation triggers
- Confidence thresholds
- Permission boundaries
- Operational kill switches
You should always be able to answer:
- What did the agent do?
- Why did it do it?
- What data did it use?
- Who approved the workflow?
- Where did escalation occur?
- How do we stop it quickly if behavior degrades?
The mistake organizations make is treating AI systems differently from operational systems. They are operational systems. The standards should be higher, not lower.
If an enterprise leader wanted a practical framework for deciding where to deploy agents over the next six months, what would the first version of that framework look like?
I would start with a very simple scoring model.
For every workflow, score:
- Repetition frequency
- Operational cost
- Exception rate
- Compliance sensitivity
- Decision ambiguity
- Human coordination overhead
- System fragmentation
- Measurability of outcomes
- Liability risk when things go wrong
The best early candidates usually look like this:
- High repetition
- High coordination overhead
- Low ambiguity
- Clear success metrics
- Moderate operational pain
- Existing human bottlenecks
Then deploy narrowly.
Do not start with “transform the enterprise.” Start with one painful workflow that everybody understands.
Measure:
- Time reduction
- Accuracy
- Escalation frequency
- Failure patterns
- Human intervention rates
- Operational satisfaction
Right now, the winners are not the companies trying to automate everything overnight.
They are the companies learning where agents genuinely belong and building operational muscle around that reality.
About Michael Privat,
Chief Data and Engineering Officer, Availity
Michael Privat is a Chief Data and Engineering Officer leading a global team of 500+ engineers. With 25 years in tech, he helps organizations unlock speed, clarity, and accountability. He turns stalled engineering teams into high-performing systems built on ownership, discipline, and modern AI-driven workflows through his “accountable autonomy” model.
For more serious insights on AI, click here.
Did you find this interview with Stuart McClure useful? If so, please like, share or comment. Thank you!
The cover image is AI-generated from the author’s prompt and Aravind’s source photos.
Leave a Reply