Agentic Artificial Intelligence in Kansas City: A Field Guide for 2026

What agentic AI is, how to build it safely, what it costs per task, and where it pays off for Kansas City companies, from the founder of KC AI Lab and Vector1 Research.

By Brian James Curry, Founder of Vector1 Research, Kansas City

Agentic artificial intelligence is the part of AI that acts. Instead of answering a question, an agent takes a step inside a business system: it tenders a load, books a crew, routes a claim, or drafts a submission for review. For Kansas City companies, that shift from talking to acting is where AI starts to change the P&L, and also where it starts to carry real operational risk.

This is the third piece in my Kansas City series. The State of AI in Kansas City in 2026 described the region’s move from chat wrappers to decision systems. Kansas City Artificial Intelligence: 20 Opportunities for 2026 ranked where those systems will create the most economic value. This one is the practical guide: what agentic artificial intelligence in Kansas City should look like, how to architect it, how to price it, and how to govern it.

I’m Brian Curry. I started the Kansas City Machine Learning Group in 2016, which became KC AI Lab and grew to more than 1,500 practitioners, and I founded Vector1 Research, an independent applied AI lab in Kansas City that designs agentic systems and the causal measurement around them. What follows is the framework I use.

What Agentic Artificial Intelligence Is

An agentic AI system is a model that pursues a goal by choosing and taking actions through tools, checking the results, and continuing until the goal is met or a rule stops it. Four parts make it work: a goal, tools it is allowed to use (an API, a database query, a form in the ERP), memory of what it has done, and policy that limits what it may do.

Gartner expects the change to be fast. It predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from none in 2024, and that a third of enterprise software applications will include agentic AI by then (Gartner).

Not every agent should have the same freedom. I use five levels of autonomy, and the right level depends on the cost of a mistake, not on how capable the model is:

1. Assistant. Answers questions and summarizes. Takes no actions.
2. Copilot. Drafts an action, such as a purchase order or an email, that a person reviews and executes.
3. Supervised agent (human-in-the-loop). Executes actions itself, but only after a person approves each one.
4. Governed agent (human-on-the-loop). Executes actions within set limits on its own. People set the limits, review samples and handle the exceptions.
5. Autonomous agent. Acts with only periodic review. Rarely appropriate for decisions with real money or safety attached.

Most of the value for Kansas City companies sits at level 4. It is where volume gets handled without a person approving every step, and where accountability still stays with people.

Why Kansas City Is a Natural Home for Agentic AI

Agentic AI pays off where there are many repeatable decisions that run through systems of record, and Kansas City’s economy is built on exactly that kind of work.

The metro is the largest rail hub in the U.S. by tonnage, with four Class I rail lines and four intermodal parks, according to KC SmartPort (KC SmartPort). Freight tendering, dock scheduling and carrier billing are high-volume, rules-bound workflows, the kind agents handle well. The region’s engineering firms, health systems, insurers, banks and animal health companies run similar workflows at scale.

Three other things favor the region. First, Kansas City companies tend to be conservative operators, and that is an advantage here: the businesses that insist on limits, audit trails and measured results are the ones whose agents survive past the pilot. Second, compute is arriving locally, with Google’s Project Mica and Meta’s data center adding hyperscale capacity to the metro. Third, the region has a practitioner community that has been building machine learning systems together since 2016, which shortens the distance between a good idea and a team that can ship it.

Kansas City agentic AI does not need to copy the Silicon Valley version, which often starts from what a model can do. It should start from what a business needs decided, at what volume, and at what cost of error.

A Reference Architecture for a Kansas City Enterprise

A production agentic system is eight layers, and the most important design rule is that the limits live outside the model. A model can be persuaded or confused; a spending cap enforced in code cannot.

1. Systems of record. The ERP, transportation management system, electronic health record or claims platform. Agents act through these systems, never around them, so every action lands where the business already audits work.
2. Data foundation. A warehouse or lakehouse such as Snowflake or Databricks, with clean master data for vendors, products, customers and prices. Documents and specifications become searchable knowledge, which is the job my open-source framework MeaningFlow does with embeddings and knowledge graphs.
3. Tool layer. Each action the agent may take is a narrowly scoped tool with the least privilege it needs. The Model Context Protocol has become the common open standard for these connections; Anthropic donated it in December 2025 to the Agentic AI Foundation, a fund under the Linux Foundation co-founded with Block and OpenAI (Anthropic).
4. Orchestrator. The model, or a set of cooperating agents, that plans the steps and calls the tools.
5. Policy engine. Hard rules enforced in code: spending limits, approved vendor lists, regulatory constraints, hours of operation. The orchestrator can request an action; only the policy engine can allow it.
6. Memory and audit. A structured record of what each agent saw, what it decided, why, and what happened. My research on Memory-Node Encapsulation addresses this layer: episodic memory that makes agent decisions reviewable and lets the system learn from past cases.
7. Measurement. A holdout or comparison group so the business can see what the agent actually changed, not just what happened after it launched. This is where causal inference earns its place, and where my open-source tools Papilon and PyCausalSim apply.
8. Human console. An exception queue, a sampling review of routine actions, adjustable limits and a stop switch that works instantly.

Teams that skip layers 5 through 8 usually get a demo. Teams that build them get a system they can leave running.

The Economics of an Agent

The cost of an agent is not the model bill. It is the model bill plus tools, plus the human review it still needs, plus the cost of the mistakes it makes, and the last term usually decides whether the project survives.

I price every agent per task with c = c_model + c_tools + ρ × c_human + e × c_error. Here ρ is the share of tasks a person still reviews, c_human is the cost of that review, e is the agent's error rate on tasks it completes alone, and c_error is the average cost of one error.

Take a Kansas City distributor’s freight tendering task. A planner spends 12 minutes per tender at a loaded cost of $45 an hour, so the human cost is $9.00 per task. The agent reads about 40,000 tokens and writes about 4,000 per task. At an assumed price of $3 per million input tokens and $15 per million output tokens, the model costs $0.18. Add $0.05 for tool calls and infrastructure. A person still reviews 15% of tenders, which adds $1.35. If 1% of the agent’s tenders go wrong at an average cost of $200 each, errors add $2.00.

The agent’s total is $3.58 per task against $9.00 for a person, a saving of $5.42. At 4,000 tenders a month, that is about $21,700 a month. The model itself is only five percent of the agent’s cost; review and errors are almost all of it.

The same formula gives a break-even point. With these inputs, the agent stays cheaper than a person until its error rate reaches about 3.7%. That number, not a benchmark score, is the one to track. All of these figures are illustrative, and the comparison assumes the person makes no costly errors of their own, which flatters the human side. Put your own volumes, rates and error costs into the formula before you put a budget behind it.

Governance and Risk

Most agentic AI projects will not fail on capability. They will fail on cost, value and control. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 for exactly those reasons: escalating costs, unclear business value and inadequate risk controls (Gartner). The same release estimates that only about 130 of the thousands of vendors marketing agentic AI offer genuine agentic capability, the rest being rebranded chatbots and automation tools. Kansas City buyers should ask any vendor to show the policy engine, the audit log and the measured result.

Two security risks matter most once an agent can act. The OWASP Top 10 for LLM Applications 2025 ranks prompt injection first: instructions hidden in an email, a PDF or a web page can redirect a model away from its task. It also lists excessive agency, giving a model more tool access or authority than the task needs, as its own risk (Invicti summary of the OWASP list). The defenses are architectural: treat every document an agent reads as untrusted data, give each tool the narrowest permission it needs, and keep spending and approval limits in the policy engine where no instruction can change them.

For the governance program itself, I start from the NIST AI Risk Management Framework, released in January 2023, and its four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile in July 2024 and notes the framework is being revised (NIST). In practice that means a named owner for each agent, a written list of what it may and may not do, a measured error rate with a threshold that pauses it, and a review cadence that does not depend on someone remembering.

The audit record ties it together. If an agent took a wrong action last Tuesday, you should be able to replay exactly what it saw, which rule allowed the action, and why it chose what it did. That is the problem my Memory-Node Encapsulation research was built around, and it is the first thing I check in any agent design.

Three Agentic AI Patterns for Kansas City

The same architecture supports very different levels of autonomy, and choosing the level is the most important design decision. These three patterns cover a large share of the opportunities in Kansas City, and each sits at a different level for a reason.

Freight tendering: a governed agent

What it does. Prices each lane against current market data, picks a carrier from the approved list, tenders the load in the transportation management system, and follows up if the carrier does not accept.

Limits. Approved carriers only, a price ceiling relative to the contract rate, safety and insurance checks, and escalation for any load above a set value.

Measurement. Hold out a share of lanes for planners as a comparison group and track cost per load, on-time pickup and the agent’s error rate against the break-even rate from the cost formula.

This is a level 4 pattern because errors are frequent but cheap and recoverable, and volume is high.

Engineering specification review: a supervised agent

What it does. Reads a draft specification, checks it against the firm’s standards and applicable codes, flags conflicts between disciplines, and drafts the comments for the engineer of record.

Limits. It never issues, stamps or sends anything. Every comment is reviewed by a licensed engineer, and the agent cites the standard or prior project behind each flag.

Measurement. Track review hours per project and the share of issues caught before construction, compared with projects reviewed the old way.

This stays at level 2 or 3 because professional liability sits with the engineer, and one missed conflict can cost more than a year of saved hours. For Kansas City’s design firms, which work on infrastructure worldwide, the payoff is still large: hours saved on routine checks go into the judgment calls only engineers can make.

Hospital staffing: a copilot moving toward supervision

What it does. Forecasts admissions and discharges by unit and shift, then proposes staffing changes, float assignments and on-call requests.

Limits. A charge nurse or staffing coordinator approves every change. The agent cannot contact staff on its own, and union and regulatory rules are enforced in the policy engine.

Measurement. Overtime hours, agency staffing spend and diversion hours, compared across units or weeks with and without the agent’s proposals.

This starts at level 2 because patient safety is involved. It can move to level 3 for low-risk actions, such as drafting the on-call request, once its forecasts have earned trust.

A 90-Day Plan for a Kansas City Company

Ninety days is enough to learn whether one agent pays for itself, if the plan is built around a decision rather than a demo.

1. Days 1 to 30: choose and baseline. Pick one high-volume decision with a clear system of record. Measure today’s volume, cost per task, error rate and cost per error. Write down the autonomy level the decision warrants and the limits the policy engine will enforce. Clean the master data the agent will depend on.
2. Days 31 to 60: build and shadow. Build the tools, policy engine, audit memory and human console. Run the agent in shadow mode: it makes every decision but takes no action, and its choices are compared with what people actually did. This is where most surprises appear, and they cost nothing yet.
3. Days 61 to 90: governed launch. Let the agent act on a slice of the volume while a comparable slice stays with people. Track cost per task, error rate against the break-even rate, and exception volume. At day 90, the comparison group tells you whether to expand, adjust the limits, or stop.

Stopping is a valid outcome. A measured stop at day 90 costs far less than an unmeasured agent that runs for a year.

Building Agentic AI in Kansas City

If your company is considering agentic AI in Kansas City, start with the cost formula and the 90-day plan above, and be skeptical of anyone who cannot show you the policy engine, the audit log and a comparison group. Those three things separate an agent that runs for years from a pilot that ends in a slide deck.

At Vector1 Research, that is the work: agentic architecture with guardrails, causal measurement of what each agent delivers, and open-source tools on GitHub that let your own team check the math. You can reach me at brian@vector1.ai.

This is the third article in my Kansas City series, after The State of AI in Kansas City in 2026 and Kansas City Artificial Intelligence: 20 Opportunities for 2026.

Frequently Asked Questions

What is agentic artificial intelligence?

Agentic artificial intelligence is AI that pursues a goal by taking actions through tools, such as updating an ERP, tendering freight or routing a claim, within limits set by people. It differs from a chatbot, which only answers questions.

Where is agentic AI most useful for Kansas City companies?

Agentic AI fits Kansas City’s high-volume, rules-bound work best: freight tendering and dock scheduling, engineering specification review, hospital staffing proposals, claims routing and procurement. The strongest designs keep people on the loop, with hard limits, audit trails and measured results.

How much does an AI agent cost per task?

Cost per task equals model cost plus tool cost plus human review plus the cost of errors. In an illustrative Kansas City freight example, an agent costs about $3.58 per tender against $9.00 for a person, and review and error costs make up almost all of the agent’s total.

Is agentic AI safe for regulated industries?

It can be, if limits are enforced in code outside the model, every action is logged and reviewable, documents are treated as untrusted input, and autonomy is matched to the cost of an error. The NIST AI Risk Management Framework and the OWASP Top 10 for LLM Applications are good starting points.

What is human-on-the-loop AI?

Human-on-the-loop is a governance model where people set boundaries, review samples and handle exceptions, while AI agents execute routine decisions within those boundaries. In human-in-the-loop designs, a person approves every action.

Who is Brian Curry?

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group. He designs agentic AI systems and causal measurement for enterprises and publishes open-source tools including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast.

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.