Presenter console

“Good afternoon. Jon Keam and I are going to talk about AI foundations for intelligent tax administration, but we are not starting with a model name. Models change too quickly, and the mission is too important, for that to be the foundation.”

“We are going to use one ladder throughout the hour. At the bottom is data: returns, transcripts, correspondence, logs and telemetry. Predictive systems turn some of that into information—a score, a match, a classification or a diagnosis. Generative systems, grounded in approved sources, help turn information into usable knowledge. At the top is judgement: a determination a named person signs. That top step stays human.”

“We will not dig into every term today. This companion carries the detail so the session can focus on making better technology decisions now and preserving options for later.”

“The colors will follow that ladder on every slide. Red is data; orange is information; green is knowledge; purple is judgement. If something we show does not help the agency climb a rung, it is technology looking for a mission.”

“Red Hat provides the operating foundation, Four Inc. supports the federal distribution path, and Carahsoft is hosting us today. Our promise is simple: before you leave, the major claims will connect to something we can run in a shell. We will define the terms, show where AI belongs, show the platform around it, and finish with hands-on labs rather than a thank-you slide.”

“Before we go any further, we need to separate four things that get collapsed into the word AI. Predictive AI estimates. It gives you a score, a class or a probability. Generative AI creates. An LLM is simply a generative model specialized in language. Agentic systems add tools and rules. Automation executes known steps.”

“Take an out-of-memory event. Predictive AI can flag that the host is trending toward exhaustion. The LLM can read the logs and approved runbooks and explain the likely cause. An agent can query the metrics, look at the CMDB, open the incident and invoke an Ansible dry run—but only with an allowlisted tool belt and rules: read-only first, approved systems only, stop if the evidence conflicts, require approval for disruptive change, and record everything. Then automation runs the tested remediation, validates it and rolls it back if health gets worse.”

“The audit example is the same architecture. A model flags a deduction pattern. The LLM summarizes the evidence and current guidance. The agent may use the filing entity’s approved calculator and retrieve the relevant records, but it cannot change source values, use an unapproved calculator or turn a mismatch into a determination. Automation moves the approved fields, captures the calculation and routes the exception. The auditor signs the judgement.”

“And here is the hard part: starting AI without a mature automation strategy is mistake number one. AI increases the volume and velocity of information. If every recommendation still becomes a manual ticket, you have built a faster backlog. Islands of scripts are not an enterprise toolchain. Institutional automation is reflexive: detect a known condition, apply policy, execute an approved response, validate it and preserve the evidence. AI can make that loop smarter, but it cannot substitute for the loop.”

“AI is an umbrella word. Under it are different ways to do work. We will explain the useful distinctions, then move on. The deeper definitions stay in this companion so you do not have to memorize them.”

“Predictive AI learns patterns from examples and uses them to estimate something about a new case. A weather forecast estimates the chance of rain. An operations model might estimate which server is at risk of running out of memory. An audit-support model might flag a deduction pattern for review. None of those estimates proves what happened or what is allowed.”

“Generative AI produces content. For our examples, an LLM—a large language model—constructs text using learned patterns and the material we give it. It can summarize a return, explain a log or draft a response. It is generating an answer; it is not automatically looking up and applying the current IRS rule correctly.”

“Calling that an educated guess is fair as an introduction. But a forecast estimates an outcome, while a language model generates a response. Neither is IRS policy. A model can say something true, but sounding sure does not make it authoritative.”

“Retrieval helps by putting the relevant approved document in front of the model. A calculator helps by doing the arithmetic using explicit rules. Those tools give us stronger evidence, but we still need the right tax year, inputs, exceptions and reviewer.”

“Agentic means the system can choose what to do next: query the logs, inspect a recent change, retrieve a document or call a calculator, then look at the result and decide whether another permitted step is needed. Tools are its available functions. Rules are its limits. The software must enforce those limits; asking the model to behave is not access control.”

“Automation executes the procedure we have already defined. It can collect records, run a calculation, route an exception or perform an approved repair. It does not need a language model to follow a known checklist.”

“These are capabilities you can combine, not mandatory stages. The decision is what the work needs. Getting that distinction right helps us solve today’s problem without committing every future workflow to one model or one vendor’s way of working.”

“Why now? Because the work is seasonal, textual and unforgiving. The workforce is comparatively flat, but demand is not. A disproportionate share of demand arrives during the filing-season peak, and the shaded space between the demand curve and staffed capacity is where backlogs, overtime and inconsistent handling appear.”

“AI can help because much of the work is language: correspondence, guidance, notes, notices and the institutional memory carried by experienced employees. But AI also produces information at machine speed. If every classification, explanation or recommendation lands in a manual queue, we have not removed the bottleneck—we have amplified it.”

“That is why automation strategy comes before AI scale. A governed toolchain has to route, validate, execute, record and escalate that output. Otherwise the agency gets a faster backlog.”

“Mistake number one is starting an AI program without understanding the automation strategy that will carry it. A few isolated scripts are not an enterprise toolchain. You can learn in a bounded pilot while you build the foundation, but the path into institutional operation needs to be part of the plan from the beginning.”

“The final constraint determines the architecture: the data boundary and approved uses determine where sensitive information may be processed. The model comes to the data. Choosing a particular model may be a six-month decision. Choosing where models run, how they are governed, and how work moves around them is a ten-year decision.”

“The model is one component. Think of a car: the engine matters, but transportation also needs steering, brakes, controls and somewhere to go. An impressive engine does not give you the whole vehicle.”

“Training is the learning phase. Examples change the model’s internal numerical settings so it becomes better at a task. Inference simply means running that trained model on a new request. A token is a small piece of text, often a word or part of a word. Those are the only mechanics we need today.”

“A typical generative language model builds the answer piece by piece, using patterns learned during training and the context supplied now. It can learn facts. But it is not opening the current IRS policy book every time it answers unless we build a process that retrieves that book. Its learned knowledge may be old, incomplete or applied incorrectly.”

“Prompting means giving the task and instructions. Context is the material available for this answer. Retrieval finds relevant passages and puts them into that context. Fine-tuning is additional training to improve a particular behavior or task. Use the right lever for the problem: a missing current document usually calls for better retrieval, not a new training project.”

“For an auditor, retrieve the applicable guidance and use a verified calculation tool for arithmetic. For a system administrator, retrieve the relevant logs, change record and runbook. In both cases the model helps interpret evidence; its explanation still needs checking.”

“We will not unpack every box here. The point is that a useful service includes the data connections, model runtime, permissions, monitoring and automation around the model. Those are the parts we want to operate consistently even when we replace the model later.”

“This staircase is the spine of the presentation. We are very good at creating data and information and then pretending we created knowledge. The test is simple: if it does not inform an outcome we need, it is not knowledge.”

“Predictive AI climbs the first riser. It scores, matches, flags or classifies. Generative AI, grounded through retrieval, helps climb the second. It reads, synthesizes, cites and drafts. But the final riser is different. Judgement is a determination someone is accountable for and signs.”

“That does not mean knowledge cannot drive automation. A directive can say patch the devices every Tuesday, and automation can carry out that procedure. If the rules already say to stop when an upstream service is unavailable, ordinary automation can enforce that exception too. People handle the situations and competing priorities that fall outside the defined rules.”

“In this design, accountable review remains on the purple step. It is to remove the transcription, matching, routing and repetitive evidence gathering that keeps that person trapped on the red and orange steps. Every hour a machine safely absorbs at the bottom is an hour an examiner can spend on the decision only a person should make.”

“Predictive and generative describe the job we ask the technology to do. Prediction estimates a value, category or outcome. Generation produces content. They overlap: a generative language model itself uses prediction to build text. We separate them here because forecasting a failure and drafting an explanation require different evidence and tests.”

“A predictive error is not automatically obvious. A score can look reasonable and still be wrong. We need known outcomes to check missed incidents, false alarms and whether the scores mean what people think they mean.”

“A generated paragraph can also look reasonable and be wrong. It might cite the wrong year, omit an exception or state a possible cause as a proven cause. We check whether the source supports the claim, whether the answer covers the relevant facts, and whether the model stops when evidence is missing.”

“A forecast is not a diagnosis. A diagnosis draft is not an approved repair. An unusual deduction is not an improper deduction. The useful system keeps those distinctions visible.”

“Sometimes the best solution combines a score, a grounded draft, a reviewer and automation. Sometimes an exact rule and a calculator are enough. Selecting the smallest reliable combination is more valuable than using every term on this slide.”

“Where does AI belong first? Look for work that is high-volume, text-heavy and already reviewed by a person. Intake, classification, guidance retrieval and drafting fit that pattern. AI prepares; the existing reviewer remains responsible.”

“Notice what is not highlighted: exam selection. That is deliberate. Selecting a taxpayer for examination is not just another classification problem. It is a consequential determination with fairness, authority, documentation and appeal implications. It may eventually use analytical support, but it does not inherit permission from a low-risk drafting use case.”

“Two rules of thumb keep us honest. First, start where a human already checks the output. Second, automate the toil around the model: provisioning, patching, data collection, routing, scaling, evidence capture and hand-offs. A brilliant model surrounded by manual operations is still a fragile service.”

“These are the five outcomes we are actually selling—not AI for its own sake. Modernize operations. Improve efficiency and accuracy. Unlock data-driven insight. Strengthen security and compliance. Build an AI-ready foundation.”

“Each outcome names the product capability and the lab where you can test the mechanism yourself. That matters because a phrase such as intelligent operations is otherwise just buzzword bingo.”

“The easiest first outcome is often operations. An error string appears in a log. An Event-Driven Ansible rule recognizes the known condition. A tested playbook runs on approved hosts. The on-call is notified. The ServiceNow or Remedy record opens and closes. The workflow records the job and links the outcome to the incident.”

“That first outcome does not require a grand AI strategy. It requires institutional automation—and it creates the toolchain that lets later AI output become governed action rather than another queue.”

“Integration failures often appear at the seams between systems. Gain a little weight and your pants split at the seam, not through the center of the fabric. Architecture has the same problem.”

“Read this from the bottom. Hardware and compute provide the physical base. RHEL and OpenShift provide the common operating foundation. Above that sit model serving, training, pipelines, registry and retrieval. Mission applications consume those capabilities. Systems of record remain where they are; APIs and automation connect them rather than forcing a rewrite.”

“The important parts are the two vertical bands. Trust is not a compliance phase added after deployment. Automation is not a script added when operations becomes painful. Both span the stack. Identity, signing, FIPS configuration, scanning, inventory, deployment, remediation and evidence need defined owners and reusable controls wherever those controls apply.”

“A seam is acceptable when we know where it is, who owns it, what contract crosses it, how it is observed, how it fails and how it is repaired. One foundation does not mean one homogeneous estate. It means heterogeneous systems present a governed, repeatable interface instead of becoming operational snowflakes.”

“That is how information becomes actionable knowledge from bare metal to the taxpayer—and how the model comes to the mission instead of mission data being dragged to somebody else’s model.”

“The federal problem is not getting a model. Models are abundant. The problem is operating AI without giving up control of the data, location, hardware, lifecycle or future choices.”

“Our proposition is: bring the model to the mission, not mission data to somebody else’s model. The unique asset is your data and the procedures around it. Keep visibility into both. Run on-premises, in an accredited cloud or in a disconnected enclave as policy requires.”

“Open source gives us a glass box rather than a black box. That does not make every model output explainable by magic. It means the components, interfaces and lifecycle remain inspectable and replaceable.”

“For platform teams, Models-as-a-Service is the practical pattern. A central team publishes governed model endpoints. Projects receive identity, quotas, monitoring and an upgrade path instead of rebuilding an entire stack.”

“Automation is a differentiator. The durable discriminator is the connected estate: operating system, platform, security controls, model lifecycle and execution layer working as one governed system. That keeps today’s decision from limiting tomorrow’s mission.”

“On the left is the definition of the service: infrastructure as code, a signed model and a GitOps repository. On the right are three places the same governed service may run: the datacenter, an accredited cloud or an air-gapped enclave.”

“The point is not that every workload moves everywhere without testing. The point is that placement becomes an explicit policy decision rather than a manual re-architecture. Data sensitivity, capacity, latency, cost and accreditation determine the venue.”

“This changes three things. Evidence becomes reusable, so workflow two starts with workflow one’s controls. Filing season becomes a scheduling and capacity event rather than a redesign. And the exit door remains visible.”

“Modern lock-in is subtle because it limits future functionality, not merely today’s invoice. Think of a phone plan that can call Domino’s but not Pizza Hut. Sometimes a native service is the right choice; we should say that honestly. The discipline is to know which choice creates dependency and to preserve portability where the mission requires it.”

“The useful pattern is a loop, not a chatbot. I can talk to a chatbot at night and it will tell me how wonderful I am. That is pleasant. It is not what I want during an outage, an audit or an investigation.”

“This example combines prediction, explanation, retrieval and action. Other workflows can use fewer of those capabilities. Predictive AI detects a pattern in telemetry. The language model explains the evidence in ordinary language. Retrieval grounds that explanation in approved runbooks, incidents and change records. We configure the automation platform so only approved jobs, targets and permissions are available to this workflow.”

“Agentic does not mean unlimited autonomy. The agent gets a restricted tool belt: perhaps logs, metrics, CMDB, ITSM and an Ansible dry run. The rules define identity, targets, maintenance windows, thresholds, approval gates, evidence and stop conditions. The agent may choose which approved tool to use next; it may not invent a production action.”

“In this workflow, each change updates the incident-management record and runs through the governed automation console. The workflow checks the outcome and preserves the evidence. Diagnostic reads are logged according to the agreed process. After review, the incident summary may enter the knowledge base. Generated text never declares itself true.”

“That is what makes the loop institutional: the model improves interpretation, while the automation toolchain supplies controlled execution and memory.”

“Think of this as a private operational notebook. It has read your logs, runbooks, incidents and change records, and it runs beside the systems it observes. Nothing FTI-adjacent has to leave the boundary.”

“The blue portion of the chart is detection. That was already fast. The large orange portion is diagnosis—the time humans spend asking what changed, whether it happened before and which runbook applies. That is information work.”

“The example in the deck shows diagnosis falling from two hours forty-one minutes to twenty-four minutes, and total resolution time falling from three hours sixteen minutes to fifty-nine minutes. The scenario illustrates the value of connecting logs with changes and runbooks. It is not a performance promise for another environment.”

“Say the numbers accurately: twenty-four minutes is diagnosis, not MTTR. We still include detection, remediation and verification. That precision is what makes the story credible.”

“The four layers are visible here. Prediction flags the risk. Generation produces the cited diagnosis. An agent chooses among approved diagnostic tools. Automation performs and verifies the approved remediation. The opportunity is to reach a supported diagnosis faster while still checking the repair and its result.”

“Crawl, walk, run sounds obvious, but most programs skip the grounding step. They prove that a model can produce an interesting answer and immediately start planning enterprise scale.”

“In the prove phase, choose one recurring, painful workflow with a baseline. How long does it take today? Where does the queue form? What is the error rate? What evidence does the reviewer reconstruct by hand? Change one thing and measure it.”

“Before that proof can scale, ground it. Agree in writing on which data may be used. Identify the automation owner. Establish capacity, identity, evaluation, approval, logging, rollback and the security boundary. Inventory the current automation estate so the pilot joins the toolchain instead of creating a new island.”

“Only then scale: more users, more models, more workflows and more venues. The success test is that we understand the dependencies we accept today and retain a practical path to the alternatives the mission may need tomorrow.”

“Plan for problems, not predicaments. A hypothetical future GPU shortage is a predicament to monitor. A missing data agreement is a problem stopping the pilot today. Solve the problem in front of you.”

“Do not leave this hour with a vague instruction to go do AI. Step zero is a lab walk-through. We put the right people together, run the mechanisms and find where in the stack you actually need to prove value.”

“Then we hold an architecture workshop around one workflow. We map the data boundary, the existing automation toolchain and its islands, the human decision, the security controls, the baseline and the measurable outcome. That scopes a guided pilot instead of a science fair.”

“If you are already a Red Hat customer, ask for your Solution Architect. If you do not know who that is, email Jon or me, open a Customer Portal request, or ask through your management line. Ted Craig is the IRS account executive, and pilots should connect to the account team early. Four Inc. and Carahsoft belong in the acquisition path before the end, not after the technical work is complete.”

“Bring your platform team, mission owner, automation owner and cyber team together from the start. Unanswered foundational questions are free to ask now and expensive to rebuild later.”

“Everything is public at the site shown here. Bring us your hardest question. If we do not know the answer live, we will find the person who does.”

“We are ending the narrative portion with something you can run, not a thank-you slide. These labs are an opinionated path through the stack so you can prove the mechanisms to yourself.”

“Sometimes people give you all the bits and tell you to build something. Sometimes you want the pot before you plant the plant. These labs are the pot: one coherent path using the upstream tools that sit inside supported Red Hat solutions.”

“The disclaimer matters. Run this on a personal lab machine. It is not an MVP, never production, and never something to install casually on a government computer. The community bits teach the architecture; the supported Red Hat products, integrated into the agency boundary, are what the agency accredits.”

“We will run Lab 01 and stop on the line showing the model bound to 127.0.0.1. That shows where the service listens; proving that prompts stay local also requires checking its outgoing connections and network controls. Then we will run Lab 05, whose representative output shows a governed rule and playbook resolving a known condition in sixty-two seconds and recording the job.”

“Anyone can take the labs, and they cost nothing. The objective is confidence. If you cannot prove the behavior to yourself, an authorizing official will hear that uncertainty immediately.”

“This is the entire hour in one loop. Start with logs from the mainframe and distributed systems. They land in Loki as data. A predictive model uses the approved features to score which service is most likely to fail in the next seventy-two hours. That score is information, not a decision.”

“Lightspeed then retrieves the relevant incident history, change records and runbooks. It drafts a playbook and an Event-Driven Ansible rulebook with citations. That is drafting assistance. When the model also chooses the next permitted tool based on what it finds, the system is acting agentically.”

“Then the loop deliberately stops. A person reviews the pull request, edits it and decides whether to merge. In this configured workflow, a merge after the required checks triggers promotion. The approved content enters the repository, Ansible Automation Platform and the curated knowledge base.”

“When the known event fires, Event-Driven Ansible launches the approved job. The procedure follows the approved steps and programmed branches; the result still has to be checked. It validates the outcome, preserves the job history and escalates if the condition falls outside the policy envelope.”

“Now the assistant can answer what happened, what was approved, what ran and whether it worked—using the evidence the loop itself produced. It does not SSH into a host and improvise. It launches job templates a person already approved.”

“That is institutional automation: not a person noticing the same problem again, not a model inventing a fix, and not an orphaned script. Review, indexing and ownership turn the promoted changes and their outcomes into a usable knowledge record.”

“This appendix is here because security questions deserve evidence, not a drive-by reassurance. We will not read all ten rows. Pick the questions that matter to your boundary and trace each one to the control and the lab that proves it.”

“Where does the data go? The model runs inside the enclave, and Lab 01 shows the service bound locally. Is it the approved model? Verify the signature before admission. What cryptography is covered? Check the relevant module validation, product version, approved configuration and FIPS state; document components outside that coverage. Does the system remain compliant? Run the Compliance Operator and retain the result, including failures and remediation.”

“What about smart cards, TPMs and hardware cryptography? Test the supported integration with the vendor hardware. What about applications and containers the agency builds? Their lifecycle needs signing, scanning, admission policy and evidence too. Platform security does not erase application responsibility.”

“Is the model drifting? Define the measure and threshold before deployment, monitor it and route violations. Can it run disconnected? Mirror images, models and updates through the approved path. Who is accountable? A named reviewer, enterprise RBAC, approved automation content and complete job history. The model drafts; it never sends or promotes itself.”

“Avoid absolutes. Controls inherited from RHEL and OpenShift provide a strong base, but anything not inherited must be documented and handled elsewhere. Federal authorization adds artifacts—attestations, POA&Ms, control evidence and reporting. Do not invent those alone; bring the security team and the appropriate specialists in early.”

“Understand the job before choosing the label. Verify the evidence before acting on the answer. Build the automation that makes the result useful. And make today's choices in a way that leaves tomorrow's options open.”