Kurt FischmanFounder, Marshal
Kurt is the CEO of Marshal, the Managed Agent Operations company.
Twenty-five jobs appear on this page, and each one shares three traits: it happens often, it ends in text or a record rather than a judgment call, and a person can check the result in about a minute. An AI agent is software that runs those steps against your real systems and hands back something finished. The list is the easy part. Picking three is the work.
Most lists of this kind fail the same way. They name a department, attach the word automation, and leave you guessing what the software actually does on a Tuesday afternoon. So every job here is written as an output plus the person who checks that output. An agent, in plain terms, runs a short sequence of steps against the tools you already pay for, your inbox, your CRM, your invoicing system, and reports back when it's done. An assistant waits to be asked a question. An agent runs on a trigger and produces a thing.
The 25 jobs, and who checks each one
Each row names one job, the finished thing the agent hands back, and the person who confirms it. Read the third column first: it's the one that decides whether the handoff is real.
| Job | What the agent produces | What a person checks |
|---|---|---|
| Support: first reply on repeat questions | A drafted or sent answer using your existing policy language | A weekly sample of answers for accuracy and tone |
| Support: order and job status lookups | A status reply pulled from your system of record | Any reply where the record looked incomplete |
| Support: return and refund intake | A completed request with photos, order number, and reason | The refund decision itself, before money moves |
| Support: appointment booking and rescheduling | A confirmed calendar slot and a confirmation message | Double bookings and same-day changes |
| Support: after-hours message capture | A structured summary of each message with urgency noted | The urgency call on anything flagged high |
| Support: review and complaint routing | A categorized item assigned to an owner with a draft response | Every response to a public complaint before posting |
| Sales: inbound lead qualification | A scored lead with budget, timeline, and scope captured | Leads the agent scored low, once a week |
| Sales: first response to web forms | An acknowledgment and a proposed next step within minutes | The first week of outbound copy, then spot checks |
| Sales: follow-up on stalled deals | A sequence of short, specific follow-up messages | The stop rule: which accounts get left alone |
| Sales: meeting scheduling and reminders | Held slots, invitations, and reminder messages | Reminders sent to accounts in a sensitive stage |
| Sales: CRM cleanup and deduplication | A merge list with proposed field corrections | The merge list, before it's applied |
| Sales: quote drafting from a price list | A populated quote against standard rates | Every quote that departs from the standard rate |
| Finance: invoice data extraction and coding | Extracted line items coded to your chart of accounts | Coding on anything above a set dollar threshold |
| Finance: payment status lookups | An answer on what was paid, when, and against which invoice | Mismatches between the ledger and the reply |
| Finance: overdue invoice reminders | A dated reminder against a confirmed open balance | The list of accounts excluded from reminders |
| Finance: expense receipt matching | Receipts matched to card transactions with gaps listed | The unmatched pile and any duplicate claim |
| Finance: weekly cash and receivables summary | A short report of position, movement, and what changed | The numbers against the source system, monthly |
| Operations: work order intake and dispatch prep | A complete work order with address, scope, and parts noted | Scheduling conflicts and access requirements |
| Operations: purchase order and delivery chasing | A current status per open order with dates confirmed | Late orders that affect a committed customer date |
| Operations: inventory reorder flagging | A reorder list with quantities and lead times | The reorder list, before purchase |
| Operations: recurring compliance document collection | Requests, reminders, and a tracker of what's outstanding | Whether each received document is actually valid |
| Hiring: applicant screening summaries | A summary per applicant against your written rubric | Every advance and reject decision |
| Hiring: interview scheduling and candidate updates | Booked interviews and status messages to candidates | Messages to finalists and anyone being declined |
| Back office: meeting notes into tasks | Notes converted to tasks with owners and dates | Owner assignments, at the end of the meeting |
| Back office: weekly performance report | A plain-English report of the numbers and what moved | Any figure you plan to act on |
The third column is the constraint. A job only leaves your desk when a named person can confirm the output faster than they could have produced it.
Nothing on this list asks software to price a job, settle a dispute, choose a hire, or approve a payment. That isn't squeamishness about the technology. It's arithmetic about where the cost of a mistake lands. A wrong status reply costs you one embarrassed follow-up email. A wrong payment approval costs you the payment, and possibly a vendor relationship you spent four years building.
The shape underneath all twenty-five is worth naming, because it's the thing you'll use later to evaluate a job nobody listed. Volume, so the agent's practice compounds and the time saved is real rather than theoretical. A text or record output, so the work product is reviewable instead of buried in someone's judgment. And a cheap check, so a person can confirm the result faster than they could have produced it. Miss the third trait and you haven't handed off a job. You've hired a very fast intern whose homework you now grade every evening.
Adoption isn't the same as a handoff
Nearly every business your size has already bought something with AI in the product description. The SBE Council's Small Business Technology Use Survey, a national online survey of 517 employers with 2 to 99 employees fielded by TechnoMetrica from February 17 to 23, 2026, found that 82% of small business employers had adopted at least one AI tool and that the typical firm was running five of them. Owners reported saving a median of five hours of their own time per week, businesses a median of 11.5 employee hours, and the median annual spend on all of it was $2,200.
Read those numbers precisely, because they describe tools, not handoffs. Five tools and a bundle of recovered hours is the profile of a business where the owner got faster at their own work. It isn't the profile of a business where a job left the owner's desk permanently. Moving from the first state to the second is what this list is for, and it's harder than buying, because it forces you to say out loud what the job produces and who is accountable for checking it. Most businesses never write that sentence down, which is why the fifth tool delivers so much less than the first.
Cost has the same gap between the brochure and the calendar. One vendor's hands-on testing page claims that a single agent running around the clock replaces what used to take 10 to 20 hours of staff time per week. Further down the same page, that vendor tells you to expect the first two to three weeks to be configuration and calibration before any of it shows up. Both statements can be true. Only one of them is on the pricing page, and only the second one has a claim on your February.
The test that picks your first three
Three questions, in this order, and a job that fails any of them goes back in the queue.
Can you count it? If you can't say roughly how many times last month someone answered that question, chased that document, or retyped that record, you can't tell whether the agent helped. Frequency is the whole return. A job that happens twice a quarter is a job you should simply do.
Does the checker already exist? Not the checker you'd hire, or the process you'd write if you had a free weekend. The person who, today, would notice a wrong answer within a day. If that person is you, and you're already the constraint on everything else in the business, then handing this job to an agent moves the bottleneck rather than removing it.
Can you undo the mistake? Ask what happens on the bad day, not the good one. A duplicated appointment reminder is an apology. A refund issued against the wrong order is money, and the customer learns something about you that you'd rather they didn't.
Take overdue invoice reminders as an illustration of a job that passes all three. Suppose a firm sends forty of them a month, each one currently written by hand by whoever remembers. The output is a short email against a known balance. The checker is the bookkeeper who already reconciles the ledger weekly, and the failure mode is a reminder sent to someone who paid on Thursday, which costs one gracious correction. That job can leave your desk. Now hold the same three questions against contract negotiation and watch it fail all three at once.
What has to stay with you
My recommendation, offered as judgment rather than as measured fact: keep exceptions, approvals, disputes, hiring decisions, pricing, and anything a regulator might ask you to explain. Not because agents are incapable of drafting an answer in those areas, but because the value of those decisions lives in the accountability behind them, and accountability isn't a thing you can delegate to software. An agent produces the draft, assembles the evidence, and flags what looks unusual. A person decides and signs.
There's a quieter reason too. The exception is where you learn what's actually wrong with your business. The disputed invoice tells you the estimate was vague. The angry customer tells you the handoff between sales and delivery drops things. Hand every exception to software and you get a smoother week and a slowly worsening blind spot.
Start with three, not twenty-five
Pick three jobs from the list that clear all three questions, and run them for a month with a named person checking the output. Measure two things only: hours that came back, and how often the checker had to intervene. If interventions stay flat while volume rises, you've found a real handoff and you can add the fourth. If the checker is correcting a quarter of the output at week four, the job wasn't wrong, the specification was, and that's fixable before you spend another dollar.
The twenty-five jobs will still be there. What you're actually short of isn't candidates. It's supervision, and you should spend it deliberately.