AI Agents for Business: Uses, Costs and a 30-Day Plan

Short answer: AI agents for business take the repetitive work your people do between systems: data entry, invoice matching, follow-ups, inbox sorting, scheduling, reporting. Start in the back office with one workload you can measure, test it on last month’s real work, and let it earn its independence. A custom agent for one workload usually costs thousands or tens of thousands and takes weeks, not the months and hundreds of thousands a large program costs. If you already know the job, a free call tells you whether an agent can take it.

I have been taking this kind of work off people for a long time. At 21 or 22 I watched people fax documents back and forth and wrote software that sent them straight through; with a business partner it reached $2 million in sales in its second year. A few years later, in 1995, I bought a $5,000 server and rented it out for close to $5,000 a month. Three decades on, one client’s $4,000 server running open models does about one and a half people’s worth of back-office work. The tools changed; the idea did not.

I have built AI products from the inside: an AI engagement platform as CEO of Product Genius, a generative AI product used by Fortune 500 engineering teams as COO of Blameless, and New Relic AI as EVP and GM of the New Relic platform. I also run two inference servers of my own and work hands on with the models and how they are trained. So this is the version I give to CEOs: plain, and grounded in what the tools can do this quarter.

What can an AI agent do for my business?

An agent can work almost any application, website, phone line or inbox, in the cloud or on your own hardware, so the limit is rarely the technology; it is picking the right job. If agents are new to you, what is an AI agent explains them in plain English. In a business, an agent does one of three things, and the third is the one most people miss:

• Do the step. Read the invoice, enter the order, send the reminder, book the appointment.
• Prepare the case for a person. Gather the documents, check that the file is complete, pull the history, summarize a long thread, so the person who decides starts with the full picture in front of them.
• Check work already done. Compare systems that should agree (the CRM says the job is closed, billing has no charge), check an outgoing offer letter against the approved terms before it reaches payroll, flag a refund above someone’s limit, catch the outcome that does not match its own evidence.

Preparing and checking keep your people deciding while the agent does the legwork, which is why they are often the best place to start. By function, the work falls into familiar groups:

• Finance: invoice intake and matching, payment follow-up, reconciliation, the weekly numbers.
• Operations: moving orders and records between systems, document intake, scheduling and confirmations.
• Customers: the shared inbox, common questions, after-hours calls, booking, onboarding a new customer.
• Sales: qualifying inbound leads, follow-up on schedule, keeping the CRM current.
• People and IT: onboarding steps, HR and policy questions, routine access requests.

More than cost

• Absorb the busy season without hiring. Back-office work grows with transactions, not revenue. Staff for the peak and you carry idle hours in the slow months; staff for the baseline and you drown in the busy ones. An agent stretches with the volume.
• Earn revenue. Answering each inbound lead within minutes, day or night, and turning around a quote while the competitor is still typing are revenue, not savings.
• Do work that was not worth doing before. Researching each contact before a personal note, or reading each closed case for what customers keep asking, was too slow for people. It is not too slow for an agent.
• Do the work you know you should be doing. The follow-up after a job is finished, the check-in with a customer who has gone quiet, the look back at last quarter’s lost quotes: work you meant to do but were too busy for. An agent can prepare it, so your people only have to make the call.
• Raise what the business is worth. If you may sell one day, a recurring cost removed raises earnings, and buyers price earnings at a multiple.

One test tells you whether you changed the work or only sped it up: imagine switching the agent off. If the old process is still standing, with a person working the queue one item at a time and the agent drafting alongside, you added a helper. The bigger wins come from asking why each step exists, and what is left for a person once the sorting, routing and first answers are done.

Start in the back office

The best first agent is rarely the flashy one. It is the work your people repeat day after day, that follows rules most of the time and costs real hours. It pays first for three reasons: it is frequent, so the savings add up; it is measurable, so you know it worked; and its mistakes stay inside the building while you learn how agents behave on your real work.

The front door is the right start when inbound volume is high and repetitive and the handoff to a person is clear, such as after-hours calls or order-status questions. Even then, read the room: an angry customer wants a person; some embarrassed ones would rather ask a machine.

How to pick the first job

My operating rule: “Any process you can model can be mostly automated, with the exceptions routed to humans.” In customer service, support and claims work, that pattern typically takes 30 to 70 percent of the routine work off your team, and your people keep the cases that need their judgment. The spread is not the model; it is whether someone wrote the process down and designed the path for the exceptions. To find your first one, test each candidate:

• Does it repeat each week, with language at the center (emails, documents, forms, calls)?
• Can the rules be written down, and can a person check the result before it has consequences?
• Is it high toil and low joy? The task people put off is often a better target than the longest one.
• Does it keep landing on your desk? The recurring job that escalates to you is usually the one that cannot keep up.
• Is the problem capacity? An agent fixes too much work for too few hands. It does not fix a bad process or a wrong strategy; it runs them faster. If a checklist or a better form fixes it, use that.
• Where does the work wait? Most of the time a job takes is waiting for a sign-off, not doing. If the invoice sits in an inbox for two weeks, automating the other steps will not make it faster.

Then rank the survivors by what the work costs today times your confidence it can be done, and start with the boring, high-volume win that does not make the board slide. It teaches your team the method at low risk, and it funds the next one.

Is your business ready? Four questions

• Where does the data for this job live right now, and who owns it? Show it in the system, not in a slide. “Not sure” or “it is complicated” means fix that first.
• How current is it? An agent deciding on yesterday’s numbers is confidently wrong on a delay.
• How many steps of the job leave a record in a system, and how many live in someone’s inbox, a spreadsheet or a hallway conversation? An agent cannot learn from work that leaves no trail.
• Can someone describe the job without saying “it depends who is doing it”? Could it run without a person quietly making up for the gaps? If not, agree how it should work before an agent learns it.

Weak answers are not a reason to cancel; they are the first two weeks of the project. Fix only the data and the steps this one job touches, then build. If the work leaves no trail today, start keeping one now: record and transcribe the calls, keep the emails in one place. Your own history is data you cannot buy anywhere, and you cannot go back and collect what you did not keep.

A morning with an invoice agent

An illustration of how the work changes, not a client story:

• Overnight. Forty invoices arrive by email and through the supplier portal. The agent reads each one, pulls the supplier, amounts, dates and purchase order number, and matches it to the order and the goods received.
• 7:00. Thirty-four match and sit in an approval queue. Six do not: a price that changed, a missing receipt, a duplicate. Each goes to the person who owns payables with the reason and the documents attached.
• 8:30. That person approves the thirty-four in one pass, using a screen that shows each amount and supplier, and spends the morning on the six that need judgment, instead of typing forty.
• Friday. The agent sends a short note of what it handled, what it escalated and the two error patterns it saw, so the instructions improve next week.

Write the job down before the build

Most of the work in a good agent is not the software. It is writing down how the job is actually done, and a process the team calls five steps usually turns out to be fifty once you sit with the people who do it. Much of how a job really works lives in a few heads, often the owner’s, the top seller’s or the long-time office manager’s. Write it with them, on one page:

• The trigger: what starts the work.
• The inputs: what information it needs, and where that lives.
• The decisions: what it has to judge, with an example of a wrong answer, a good-enough one and an excellent one.
• The checks: how it confirms each step worked.
• The destination: where the finished work goes, and how you know it landed.
• The authority: what it may suggest, what it may decide, what it may do on its own, and the conditions that hand control back to a person.

Keep that page, and the business facts it relies on, in one place with one owner. It becomes the agent’s instructions, the test plan and the handover in one document. Then test the page: could someone new meet this standard without you standing over them? If not, the page is not finished, and a new hire would struggle with it as much as an agent. The agent just tells you sooner.

Off-the-shelf tool or custom agent?

Check what you already pay for

The software you already use is adding AI features, and some of them may already be switched on. Before you build anything, list what is running, what data it reaches and where that data goes. Ask your vendors directly which models they use; the salesperson may not know, so ask for it in writing.

AI agents for small business: when a tool you can buy is enough

Common jobs inside one application are increasingly built in: meeting notes, inbox help, CRM suggestions, help-desk replies. If your job is common and lives in one app, start there. Prove the workflow with an entry-level tool before you pay for a platform.

When to build a custom agent

Packaged tools expect your business to fit their process; a custom agent fits yours. Build when the work crosses your own systems, follows rules particular to your trade, uses your own documents, or involves data that should stay on your own hardware. The old rule was buy fast, build slow. For one narrow workload that no longer holds: open models and current tools mean a custom agent can be built in weeks, and it does your job, across your systems, the way your best person does it.

What does an AI agent cost?

• Off-the-shelf tools: a monthly subscription, priced per user, per conversation, per task or in credits. Know which, because each counts something different and grows differently as you roll out.
• A custom agent, built through AI automation services: most of what I build costs thousands or tens of thousands for one workload. Large firms quote hundreds of thousands and months for a program.
• Running cost: what the model charges for the work. Frontier models charge per use; open models on your own server carry no per-use fee once the hardware is bought.

Three rules keep the running cost honest. Price the pilot on last month’s messy real items, not the clean demo set, because odd cases, retries and checks are where usage climbs. Track the cost per completed task, not only the monthly bill, since an agent can pay on hard cases and lose money on easy ones. And before you scale, ask what the bill looks like at ten times today’s volume, and who owns it. On one enterprise’s AI bill, moving the routine work to open models cut that spend by 80 percent.

How to size the prize before you build

Before a line is built, put a number on the work: how many units a month, how many minutes each takes, and what an hour of that person’s time costs you fully loaded. That is the annual cost of the work. Estimate the share an agent can take, then compare it with the whole cost to build and run, including the time your people will spend reviewing its work, which counts as much as setup.

The client example below is the arithmetic in practice: 35 percent of four people’s workload is about one and a half full-time jobs of work, worth over $150,000 a year, against a one-time $5,000 build and a $4,000 server. When the numbers look like that, the first agent pays for the second.

One caution: time saved is not money saved until you decide what the time becomes. It turns into money when it serves more customers, replaces overtime or a planned hire, or goes to work that grows the business. Decide that before you build, or the hours quietly fill with other tasks.

Your first agent in 30 days

Week 1: pick and write down

• Pick one job with the tests above and answer the four readiness questions.
• Write the one-page job description with the people who do the work today.
• Take a baseline: volume, minutes per item, error rate, cost.
• Agree what success looks like at day 30, and agree a stop rule in advance, so stopping is a finding rather than a failure.
• Map the path to production now: what it costs to run, who must approve it, and whether your security and legal requirements can be met. If you cannot map that path, pick a different job.

Week 2: build it to draft, and build the test

• Build it with read access and drafts only, on its own account, against a copy of your data and a test inbox.
• Have the people who do the job by hand collect 20 to 30 real requests from last month, in the words customers used, messy ones included, each with the right answer.
• Write the ceiling: one page of what it may not do without a person, such as re-price a quote or issue a credit. Add two caps: attempts per item and spend per day, with the case handed to a person when either is hit.

Week 3: test and try to break it

• Run the case file, grade each case pass or fail, and check it stayed under the ceiling. Run each case more than once, because one good run is a demo, not evidence.
• Try to break it on purpose, including a document with a hidden instruction in it.
• Then run it in shadow for a week or two on live work. The agent drafts each item but sends nothing, a person does the work as usual, and you compare the two each day. Count where they agree and price each miss: a wrong label costs a minute, a wrong refund costs a customer. Go live on the kinds of item it gets right and keep the rest manual for now.

Week 4: go live, small

• Go live on one workflow with a volume cap, a person approving anything that leaves the building, and a written way to switch it off and go back to the manual process.
• Measure against the baseline and decide with the numbers: scale, change or stop. Expect a dip in the first weeks while people learn to review rather than do, and judge it at the end of the month, not on day two.
• Then let it earn its next rung: after a long clean run set in advance, move from item-by-item approval to a daily summary.

When the first one works, the next job copies the method rather than starting a new project. The connections you built get reused, which is why the second agent costs less than the first.

Keeping data and decisions safe

• Sort its actions into three levels. Reading needs no approval. Changes that can be undone run on their own and are logged. Anything that cannot be undone or that the outside world sees gets a person, and the approval shows the exact email, amount or change.
• Give it its own account, with only the access the job needs. Hide fields it should not see at the source, and remember that allowed to read is not allowed to send.
• Treat what it reads as data, not instructions, and put its limits in the software, not only in the prompt.
• Separate who prepares from who approves. The agent that drafts the purchase order does not approve it, the same rule you use for people.
• Cap its attempts and its spending, so a retry loop hands the case to a person instead of running up the bill overnight.
• Log each decision with what it read, what it decided and which model made it, so when a number looks wrong you can see why.
• Watch combined access. An agent that reads HR, finance and customer records together can reveal what no single field would. A person should not learn through the agent what they are not allowed to see directly.
• Give a person a brake that stops it mid-task, and keep its activity log somewhere it cannot edit.
• Check the training setting. Some providers learn from what you send unless you switch it off. Where the data matters most, open models can run in your own office.
• Name four owners: who directs it, who reviews its work, who controls its access and who answers for the result. In a small business one person may hold the four roles; write the names down anyway.

Running it after launch

• Pace the agent to your people. The slow step is usually the person reviewing, not the agent. Release work at the pace they can clear it, ranked by customer impact, or the queue goes stale. Review becomes the work, so plan the reviewer’s time as carefully as the agent’s.
• Watch what is left for your people. When the agent takes the easy work, a person’s day can become nothing but the hardest cases, and tired people make more mistakes. Mix the work and rotate who handles the exceptions. These effects show up weeks after launch, not on day one.
• Measure the whole job, from arrival to correctly done: cycle time, first-pass accuracy, exceptions, how often people override it, and cost per completed task. Logins and hours saved are not results.
• Save the corrections. Each edit a person makes is the most valuable data you own; feed it into the instructions and the test file.
• Fix patterns, not items. Agent mistakes repeat. When you find one, look for the same mistake across its other work and fix the instruction.
• Review monthly, and re-run the test file whenever the model, the instructions or the vendor changes. Models are updated on someone else’s schedule.

What your people need to hear

Your team will ask four questions, out loud or not. Answer them before launch: which tasks will change, which decisions stay human, how mistakes will be handled, and what happens to the time the agent saves. If you can answer only two well, answer which decisions stay human and how mistakes are handled. Say plainly that the work will change, because people do not believe “nothing changes”, then show each person what their own week looks like afterwards.

• Remove waste friction, keep wisdom friction. Searching five systems and typing the same data twice should go. The pause before sensitive information leaves the company, the second look at a contract and the conversation before a staffing decision should stay.
• Keep people on decisions that could end up in court. Hiring, firing, promotion and credit are made by a person who can explain them, and so are calls that cannot be undone or that turn on ethics or ambiguity.
• Do not automate the part of someone’s job they are best at and enjoy. Take the drudgery and they will care more about the rest.
• Put your biggest skeptic on review. Ask them to write down exactly where its work falls short. Bad work does not ship, and their notes become the tests that improve it.
• Write a one-page AI policy: which tools are approved, what data may go in, who owns the output, who manages each agent, and when the policy is reviewed. Add a three-question check for any new use: is it legal, is it ethical, and am I willing to answer for it? Without it, careful people do nothing and careless ones use personal tools.

If someone’s job is simply to repeat steps one through ten, that is not a great investment of what people can do. Years ago I wrote that no-code does not remove engineers; it moves their effort from low-level plumbing to higher-value work. Agents do the same for the back office: the time goes back to customers, judgment and the work you hired people for.

Questions to ask whoever builds it

• Did they ask about your work before proposing a solution? A builder who proposes the same design to each client is selling a template.
• Who does the building? Meet the person who will build it, not only the one who sells it.
• Who owns what is built? The instructions, the code and the configuration should be yours to run and change after they leave.
• Can the model be switched later, and where will your data live and go?
• Can it show its work? Ask how it reached a given answer. A system that cannot show what it read and why it decided is hard to trust and harder to fix.
• How will it be tested before it acts, and what does success look like at 30, 60 and 90 days?
• Does it perform the exact action you need, with the permission it needs, or does it only list your software as an integration?
• What will it cost to run each month, and who maintains it when models and systems change?
• Can you talk to a customer who stopped using them?

Three real examples

One client’s agent runs entirely on open models on a $4,000 server in their own office, so no data leaves the building. The build was a one-time $5,000 development fee. It took 35 percent of the workload off four employees and saves over $150,000 a year.

A $40 million avionics company cut its legal bills by $400,000 by having AI review and draft its contracts, with the person who already reviewed contracts checking the work rather than doing it from scratch.

And a CEO with $25M in revenue asked how to use AI without an enterprise budget. We skipped the platform conversation and automated the flows his team re-typed day after day: calls, texts, emails, browser work into systems. A one-time $50,000 investment took $250,000 of annual cost out.

How to start

Pick one workload, run it through the tests above, and price it against the hours it costs today. If you have the idea and want it built in weeks, custom AI agents explains the free call: you tell me the work and the systems it touches, and I tell you whether an agent can take it, roughly what it costs and how long it takes, and whether to build it yourself or have my team build it. If you do not yet know where AI pays across the business, the AI Assessment ranks it for you.

AI agents: the questions business owners ask

What are AI agents for business?

Software that takes a defined job off your team, such as matching invoices, answering after-hours calls or chasing paperwork. It reads the work as it arrives, decides the steps, uses your systems to carry them out, checks the result and hands the exceptions to a person.

What can an AI agent do for my business?

Take the repetitive work your people do between systems: data entry, invoice matching, payment and paperwork follow-up, inbox sorting and replies, scheduling, call answering, reporting and routine HR and IT requests. It can also prepare cases for a person and check work already done, with exceptions routed to a person.

How much do AI agents for small businesses cost?

Off-the-shelf tools are a monthly subscription, priced per user, per task or in credits. A custom agent for one workload usually costs thousands or tens of thousands to build. Running costs depend on the model; open models on your own server carry no per-use fee.

Which AI agent is best for a small business?

The one that fits the job. If the work is common and lives in one application, use the agent features in the software you already pay for. If it crosses your own systems and rules, a small custom agent built for that one workload can pay back faster than a general tool.

How do I test an AI agent before it goes live?

Collect 20 to 30 real cases from last month with the right answers, run the agent on a copy of your data, grade each case more than once, try to break it, then run it beside a person on real work before it acts on its own.

Do I need a developer to build an AI agent?

Not to start. You can learn how agents behave with a current AI assistant and a drafted task. Work that crosses several systems, handles money or sensitive data, or needs reliable exception handling is better built by someone who has done it, from a written plan.

How long does it take to build an AI agent?

Weeks for one workload, once the systems and the owner of the exceptions are named. A sensible first launch takes about 30 days from picking the job to running it live on a small volume.

Should we wait for AI agents to mature?

The tools will keep improving, but waiting does not build your team’s ability to use them. Businesses that start now build the job descriptions, the test cases, the data and the habits that let them take each better model as it arrives. Start with one back-office job you can measure, and each new model makes it cheaper and more reliable.

Will AI agents replace my employees?

They take the low-value work off them. If someone’s job is simply to repeat the same steps, that is not a great investment of what people can do; the time goes back to customers, judgment and the work you hired them for.