Kurt FischmanFounder, Marshal
Kurt is the CEO of Marshal, the Managed AI Ops company that designs, deploys, and operates AI agents as critical infrastructure for founder-led businesses.
AI agents for business are software systems that own a workflow end to end: they read a situation, decide the next step, act across the tools a company already runs, and work toward an outcome. The real buying decision is not which agent is smartest. It is how much the agent may do alone, which tools and data it may touch, and where a human signs off.
Essential Insights
- AI agents for business decide the next step in a workflow rather than following a fixed branch, which is what separates them from automation and from chatbots.
- Autonomy is the amount of work an agent finishes before a human has to approve the next move, and it is a dial the business sets, not a fixed property of the tool.
- Tool access and data access are the supply line: an agent is only as capable as the permissions, integrations, and freshness of what it can read and write.
- Human oversight for a production agent has four working parts: approval gates, exception queues, human review, and audit trails.
- The gap between the companies experimenting with agents and the far smaller number running them in production is an oversight gap, not a model-capability gap.
- The four questions every buyer asks (how much autonomy, which tools, what data, how much human review) are one question asked four ways: how much can this thing do before a human has to sign off.
How the Four Buying Axes Trade Off
AI agents for business get sold as a shopping problem, and the whole market is arranged to keep it that way: ranked listicles, best-of grids, a leaderboard of tools. That framing hides the actual decision. The four questions a buyer asks about an agent are not independent settings you tune one at a time; they move together, and the amount of rope you hand the software is set by all four at once. The table below compares an AI agent run with real oversight against a bare autonomous agent and a plain scripted automation across autonomy, tools, data, and human review, the dimensions that decide whether an agent survives past the pilot.
Comparison of an AI agent run with oversight, a bare autonomous agent, and scripted automation across four operational dimensions that decide production readiness.
| Dimension | AI agent with oversight | Bare autonomous agent | Scripted automation |
|---|---|---|---|
| Autonomy | Acts alone below a risk line, escalates above it | Acts alone on everything, including edge cases | Runs only the exact path it was coded for |
| Tool access | Scoped permissions per tool and action | Broad access, rarely least privilege | Fixed connectors set at build time |
| Data | Reads fresh, permissioned records with limits | Reads whatever it was pointed at | Reads only pre-mapped fields |
| Human review | Approval gates and exception queues for unsure cases | No sign-off path when confidence drops | Breaks and waits for an engineer |
| Audit trail | Logs each decision, reason, and action | Logs output, rarely the reasoning | Logs runs, not judgment |
| Production fit | Survives real inputs and rare cases | Stalls in pilot on the first ambiguous case | Reliable until the process changes |
The agent with oversight is the only column that treats ambiguity and the sign-off path as first-class work, which is exactly why it reaches production.
What AI Agents Actually Do
AI agents for business read a situation, decide the next step, act across the tools a company already runs, and work toward an outcome rather than a single reply. AI agents for business are not a product you buy off a shelf and switch on. That loop is the whole distinction. A chatbot answers the question in front of it and forgets. A scripted automation runs the one branch it was built for and stops the moment reality diverges. An agent holds a goal in mind, checks the current state, picks a move, and repeats until the job is done or it hits something it cannot resolve.
Marshal builds these as AI Agent Systems that own a workflow end to end on top of a client's existing stack.
The concrete version is duller than the pitch, which is the point. The agent reads the CRM record, checks the calendar, drafts the follow-up, and then stops at the approval gate for a human to release it, or clears it itself when the confidence and the risk both clear the line. No new platform. No migration. The work that used to bounce between three people and a spreadsheet gets held by one system that never forgets the fourth step. That is what owning the workflow means in practice, and it is why the useful question is never which agent is cleverest. The observe, decide, act loop is the same whether the job is lead routing or invoice reconciliation; only the tools and the stakes change.
Autonomy Is a Dial, Not a Feature
Autonomy is the amount of work an AI agent finishes before a human has to approve the next move. It is the first axis buyers ask about and the one they most often get backwards, because they treat it as a property of the tool rather than a setting they own. A support agent that drafts replies and waits for a click is low-autonomy. One that answers, refunds, and updates the account on its own is high-autonomy. Same software, different rope. The right amount is a business decision keyed to how reversible the action is and how much it costs to be wrong.
This is where agent governance stops being a compliance word and starts being the design. High autonomy on a reversible, low-stakes action is free money. High autonomy on an irreversible, high-stakes one is how a pilot becomes a headline. The move is to set the dial per action, not per agent: let it clear the routine cases alone, route the unsure ones to an exception queue, and keep a human on the rare decisions that actually carry weight. You can buy the smartest agent on the market and still watch it sit in pilot forever, because nobody wrote down who signs off when it is unsure. Autonomy without that graduated sign-off is not confidence. It is an unowned risk waiting for its Tuesday.
Tools and Data Are the Supply Line
Tool access and data access are the supply line for any AI agent, and supply lines are where campaigns die. A reasoning engine with no permissions is a very expensive text box; the value shows up only when the agent can read the real record and write the real update. What it may touch, how fresh that data is, and where its permissions stop are not setup details to rush past. They decide the ceiling on everything the agent can do. Marshal treats this as data sync and admin relay work: the unglamorous plumbing that keeps the record current and the writes scoped.
The failure here is quiet, which is what makes it dangerous. An agent pointed at stale CRM data does not throw an error. It confidently acts on last quarter's truth and files the result before anyone notices the drift. Least privilege is the discipline that contains the damage: the agent gets exactly the tools and fields the job needs and nothing more, so a bad instruction or a hijacked prompt cannot reach systems it was never meant to see. Every extra integration is another door, and doors get walked through. The tool everyone shops for is the easy part. The permissions and the freshness are the project, and they are where the weeks actually go.
The Deployment Gap Is an Oversight Gap
Marshal reads the 62/11 deployment gap as an oversight gap, not a capability gap: the agents that stall in pilot are rarely too dumb to do the job, they are too unsupervised to be trusted with it. A May 2026 founder guide from Dan Cumberland Labs summarizes McKinsey and Deloitte research showing 62% of organizations experimenting with agents and only 11% running them in production. The shape of that gap matters more than the exact figures. Every one of the four buyer questions is really one question asked four ways: how much can this thing do before a human has to sign off.
That same guide cites a Gartner prediction that more than 40% of agentic AI projects will be canceled by the end of 2027, blamed on escalating cost, unclear value, and inadequate risk controls. Read the last item closely, because it is the tell. We build AI Agent Systems that run on a client's existing stack, wrapped in approval gates, exception queues, human review, and audit trails, because the pilots we audit almost never fail at the demo. They fail in month six, on a Tuesday, when the input is weird and nobody wrote down who decides. Oversight is not the tax you pay on autonomy. It is the thing that lets you grant autonomy at all, and calibrating it is what an AI agent risk assessment framework exists to do. The agent that reaches production is rarely the smartest one in the bake-off. It is the one whose owner could answer the boring question before the first ambiguous case arrived.
Where AI Agents Are the Wrong Call
AI agents for business are the wrong call when no single workflow is drowning anyone yet. The math works when manual coordination is already eating real hours every week: a founder rekeying leads at midnight, an ops lead reconciling the same three systems by hand, a rep losing deals to slow follow-up. If that pain is not present, an agent is a solution shopping for a problem, and it will show up as cost without return. The other poor fit is any job that needs human judgment on every case, where there is no routine majority to automate and no rare exception to escalate, just exceptions all the way down. Start where the volume is high, the steps repeat, and the cost of a mistake is survivable. That is where an agent earns its keep, and it is a narrower slice than the market wants you to believe.
Frequently Asked Questions
What is the best AI agent for business?
The best AI agent for business is the one that fits the specific workflow you want owned, not the one that ranks highest on a generic list. Selection follows the job: a support workflow, an outbound workflow, and a finance workflow each reward different tools, permissions, and oversight settings. The better question is which workflow is drowning someone today, then which agent can own that one with the least new risk.
What can an AI agent do for my business?
An AI agent can own a repeatable, multi-step workflow: qualifying and routing leads, keeping a CRM current, drafting and sending follow-ups, reconciling records across tools, or preparing recurring reports. The realistic scope is a job with a clear trigger, a defined outcome, and a bounded set of tools. Work that needs genuine human judgment on every case is a poor fit, and admitting that early is how the good deployments stay good.
Who are the big 4 AI agents?
The phrase big 4 AI agents usually points at the platforms behind most business deployments rather than four ranked products: OpenAI, Anthropic's Claude, Microsoft Copilot, and Google's Gemini line, with Salesforce Agentforce often named alongside them for CRM-native teams. For a founder-led business the platform label matters less than whether the agent can reach your existing tools and respect the permissions you set.
What are the 5 types of agents in AI?
The five types of agents in AI, per the standard classification, are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents, and learning agents. The list runs from the least to the most adaptive. Most business workflows are served by goal-based and learning agents, which hold an objective and improve against outcomes rather than reacting to a single trigger.
How much human oversight does an AI agent need?
Human oversight for an AI agent scales with the reversibility and stakes of what it does, not with a fixed rule. Low-stakes, reversible actions can run unattended with an audit trail. Irreversible or high-cost actions belong behind an approval gate, with unsure cases sent to an exception queue for human review. The point is to spend oversight where being wrong is expensive and withdraw it where being wrong is cheap.
Why do most AI agent pilots fail to reach production?
Most AI agent pilots fail to reach production because nobody defined ownership and sign-off before the first ambiguous case, not because the model was too weak. Industry research attributes cancellations to unclear value, escalating cost, and inadequate risk controls. Agents that reach production pair scoped tool access, fresh permissioned data, and graduated human review from the start, so the rare hard case has a place to go.
Build a business that runs itself.
Join hundreds of small businesses operating at machine speed with agents on the job.