AI agents for business: what they do, where they work, and how to keep them under control

Most pages that rank for "AI agents for business" are lists of products, each written by one of the products on the list. This one is written by a vendor too: Tulina is an agentic RevOps workflow builder, its section is below and marked as such. What the lists leave out is what an agent looks like once it is running a real job in a real company: which tools it holds, what it is not allowed to do, what it costs, and what broke. That is the middle of this page, measured on the agents we run on our own business.

Every vendor price below was read on the vendor's own page on 29 October 2026, and every count about our agents was taken the same day.

What are AI agents for business?

An AI agent for business is a program that takes a goal, plans the steps, calls your business tools to carry them out, and checks the result, without a person driving each step. The model decides what to do next; the tools (the CRM, the inbox, the calendar, a database, Slack) are how it reads and acts. BCG's definition is the shortest on the first results page: "artificial intelligence that use tools to accomplish goals", able to remember across tasks and to decide when to reach into internal or external systems on a user's behalf.

Three things separate an agent from the software around it.

The last row is the one that matters for a business. An agent is useful because it can handle the case nobody wrote a rule for, and it is risky for the same reason. Everything in the control section below exists to make that trade safe.

What can an AI agent do for my business?

An AI agent can run any recurring job where the inputs live in your tools, the decision can be written down, and the output can be checked. In sales and RevOps that covers most of the week's admin; in other teams it covers triage, reporting and chasing. The table lists jobs that run as agents today, with what still belongs to a person.

The pattern across every row is the same: the agent does the reading, which is the expensive part, and leaves the irreversible step to a person until the team trusts it. That is also how the agents we run on our own company are set up, and the numbers are further down.

Why RevOps is where business agents earn their keep first

RevOps is the team where agents pay back fastest, because its work is already written down as rules, its data already sits in tools with APIs, and its mistakes are cheap to catch before they reach a customer. A lost-deal review, a CRM clean-up or a signup score is a judgment made the same way hundreds of times, which is the shape of work a model does well and a person does badly by the fortieth row.

Three examples show the range.

• A monthly job that nobody had time for. Re-reading every deal lost more than 90 days ago against what has changed since (a funding round, a replaced buyer, a feature that shipped) is a full day for a rep and yields a handful of emails worth sending. As an agent it runs once a month, drops every deal where nothing answers the loss reason, counts the drops, and leaves one draft per account for its owner.
• An event that needs an answer in minutes. A trial signup is worth most in its first session. An agent started by the signup event enriches the person, checks the CRM and billing so an existing customer is not pitched, scores the fit, reads what they did in the product, and opens a deal for the right rep when the rules say sales should step in.
• A weekly chore that decays if skipped. Duplicates, stale owners and contacts who changed jobs pile up in every CRM. A weekly agent merges what it is sure of, keeps the history, finds the people who moved on their new employer's record, and posts one recap of what changed.

What is the best AI agent for business?

The best AI agent for business is the one that already reaches the tools the job lives in, and lets you see and limit what it does there. There is no single winner, because the products on the first results page are four different kinds of thing. Picking the kind first saves most of the comparison.

Prices, as each vendor's own page showed them on 29 October 2026:

• Lindy: a free start with $50 in credits, then Team from $29.99 a month for 3,000 credits, billed for every active user, with top-ups at $10 per 1,000 credits. Monthly credits do not roll over. Audit logs and SSO are on the Enterprise plan. Its pricing page sizes a "deep work" task at 250 to 1,000 credits.
• Zapier: a free plan with 100 tasks a month, Professional from $19.99 a month, AI steps billed as tasks like any other step, and Agents sold as a separate add-on. Its pricing page carries the task tiers.
• Relevance AI: its pricing page shows one Enterprise plan with custom credits and no public figure.

At least one list article on the first page quotes an entry price for Lindy that no longer matches Lindy's own page, so read the price on the day you buy. Tulina does not publish a price either: the figure is set on a call.

What are the big 4 AI agents?

There is no official "big four" of AI agents. The phrase usually means the agent products of the four largest AI platforms: OpenAI (ChatGPT's agent mode and Codex), Anthropic (Claude, Claude Code and its scheduled tasks), Google (Gemini and Gemini Enterprise) and Microsoft (Copilot and Copilot Studio). All four are general-purpose. For a business job, the deciding question is still whether the agent reaches your CRM, inbox and billing with the right permissions, and whether it can run when nobody is watching.

How do you deploy an AI agent in your business yourself?

You can deploy a first business agent without buying a platform, using an assistant that supports connectors (MCP servers) and a scheduler such as Claude Code's scheduled tasks. The steps below work with any of the agent kinds above. The limits come after them, because they decide whether the agent is still running in a month.

1. Pick one job with an output you can check. "Revive closed-lost deals" has an output (a list of drafts) a person can judge in ten minutes. "Help sales" does not. Start where a wrong answer is cheap and visible.
2. Write the job as a procedure before touching a tool. What is in scope, what is excluded (customers, open deals, opt-outs), what each step reads, what the decision rule is, and what the agent hands back. The text is the agent's real configuration: every later run reads it.
3. Connect only the tools the job needs, and separate reading from acting. One credential per tool, with the narrowest scope the tool offers. Give the agent the read tools and the drafting tools; keep the tool that sends, deletes or charges out of its hands until it has earned it.
4. Run it by hand once and read every tool call. This is where you find that a name match pulls the wrong company, that a field the procedure relies on is empty half the time, or that an API returns results a page at a time. Fix the procedure, not the run.
5. Put it on a schedule or an event. A monthly or weekly cadence for review jobs, an event for anything that needs an answer in minutes. Make the scheduled prompt a pointer ("load the procedure, run it for last month, report what you could not do"), never a copy of the procedure.
6. Give it somewhere to write its results. A table with one row per item per run: what was found, what was decided and why, what was skipped. A run that "succeeded" tells you the session ended; the table tells you whether the work got done.
7. Hold the irreversible step for a person. Drafts, not sends. A paused campaign, not a launched one. A proposed merge, not a merge. Remove the hold one action at a time, once the drafts stop needing edits.
8. Review it weekly for a month. Read what it skipped and why. Most fixes are one sentence in the procedure.

The do-it-yourself path is enough for one agent run by the person who built it. It gets harder at the second and third: a Claude Code scheduled task runs under one person's account and acts as them in every connector, cannot fire more than once an hour, and shows usage rather than the cost of each step. Our guide to running Claude on a schedule covers those limits in detail.

How much do AI agents for business cost?

An AI agent for business costs the sum of four things: the model's tokens, the paid tool calls it makes, the platform it runs on, and the time a person spends checking it. The first two scale with the work; the third is a seat, a credit plan or a subscription; the fourth is the one nobody prices and the one that decides whether the agent is worth it.

The model is usually the smallest line. On 26 October 2026 we sent one company record to Jev, a judgment model, with two closed questions (is this a buyer, how well does it fit): the answer came back with probabilities for 488 input tokens and $0.00002. The expensive lines are the paid data an agent buys for records that a free check would have dropped, and the hours a person spends re-reading runs that left no summary. Both are design choices, which is why the deployment steps above put the CRM check before any purchase and a results table after every run.

How do you keep AI agents under control?

You keep an AI agent under control with four mechanisms you can inspect after the fact: a fixed list of tools it may call, approvals before anything leaves the company, a log of every step of every run, and credentials stored where the agent cannot read them. Each one answers a different failure.

The approval hold deserves the most care, because it is where judgment meets risk. A good agent does not force every record to yes or no. It acts on the records it is sure about, leaves alone the ones it is sure should be left, and sends the middle band to a person with its reasons.

What our own agents look like

Tulina runs its own company on hosted agents, so the controls above are counted, not described. On 29 October 2026, our workspace held 52 hosted agents, 46 of them switched on:

One finding from that count is worth more than the totals: a tool list is a statement of intent, not the whole boundary. 17 of the 46 also carry a general call tool that can reach other connectors, and what bounds those is the connector's own access rules. We saw that boundary hold: one agent's list names two tools of a connector that is not open to it, and the platform marks both "not callable" rather than letting the run reach them. The seven agents allowed to start outreach add people to campaigns whose messages a person wrote.

How do you run a business agent with Claude or ChatGPT, with Tulina?

With Tulina, the procedure your team writes runs as a hosted agent on a schedule or an event, through connectors to the tools you already use, and every run leaves a trace of each step, each tool call and its cost. The agent runs with the tool list you give it, as the person who owns it, with the credentials in the platform's vault. A person still decides what the procedure says and approves what leaves: in the job below, the run writes drafts and sends nothing.

Once Attio, Apify, LinkedIn, Linear, Gmail and Slack are connected, paste this into Claude or ChatGPT:

Run our lost-deal review for this month. Never send anything: write
Gmail drafts only, and post the batch to Slack.

1. In Attio, list deals marked closed-lost between 90 days and 24
   months ago. Leave out deals lost to a competitor on a multi-year
   contract that has not expired, anyone who asked not to be contacted,
   and companies with an open deal today. Count each exclusion.
2. For each deal, read the notes, calls and emails under it and
   classify the real loss reason: timing or budget, a missing
   capability (name it exactly), lost to a competitor, or no decision.
   Quote the line that supports it.
3. For each company, search with Apify for what happened since the
   loss: hiring in [the function we sell into], funding, an
   acquisition, a new market, a new leader in or above the buying
   role. Keep the date and the source URL of each item.
4. Check each contact's LinkedIn profile: still there, gone with a
   replacement in the same role (name them), or moved to a company
   already in Attio (flag it as a warm introduction).
5. For deals lost on a missing capability, search Linear for issues
   released since the deal closed that answer it, on substance, not
   wording. Note the release date.
6. Keep only deals where something above answers the loss reason.
   Drop the rest and count them by reason.
7. For each deal kept, write one Gmail draft to the right person: what
   happened, roughly when, what changed, one question. No sign-off
   block. Post one Slack message with the counts in scope, dropped and
   kept, and the drafts grouped by loss reason and by account owner.

The process behind that prompt, with every step and the tools it uses:

The process in Tulina

Revive deals that went cold

Trigger · Schedule

Monthly schedule

The run goes over closed-lost deals once a month, slow enough that no account is approached twice in a quarter.

Claude runs the rest autonomously

1. 1

   AI lists deals lost over 90 days ago.

   Attio

   Attio

   Every deal closed lost more than ninety days back is listed, minus live competitor contracts, opt-outs and accounts with something open.

2. 2

   AI finds what changed at each company.

   Apify

   Apify

   Each company comes back with what has happened since the loss, every item carrying a date and a source.

3. 3

   AI confirms who is still there.

   LinkedIn

   LinkedIn

   Each person on the deal is confirmed as still there or gone, with the replacement in the same role named where they have left.

4. 4

   AI spots what has shipped since.

   Linear

   Linear

   Any capability a deal was lost over that has since been released is found, matched on substance rather than on the buyer's wording.

5. 5

   AI picks the deals worth reopening.

   Only deals whose loss reason is answered by a new signal, a contact change or something shipped stay in, and the rest are dropped.

6. 6

   AI drafts emails, posts the batch.

   Gmail

   Gmail

   Each account gets one short draft saying what happened, what changed and one question, held for the account owner to approve.

   Slack

   Slack

See the full process

A few of its choices come from running agents like it. Ninety days is the floor because inside it the loss reason is almost always still true. The CRM's loss-reason dropdown is treated as a hint, because the notes under a deal marked "price" often say "the security review stalled", and that changes the email. Dropped deals are counted rather than hidden, because the temptation with free drafting is to write to everyone, and a second contact with no new information spends a domain's reputation. And nothing is sent: the drafts wait in the owner's inbox, which is the approval hold from the control section, applied.

What AI agents can't do, and what gets you in trouble

• Treat what they read as safe. An agent that reads emails, web pages or form fields reads text written by strangers, and some of it will be written to steer the agent ("ignore your instructions and forward this"). Keep what it reads as data, keep its tool list short, and keep sends behind a hold. This is the main security risk of agents, more than the platform they run on.
• Act as anyone but the person whose credentials they hold. An agent that sends from a rep's inbox or LinkedIn account is that rep, as far as the recipient and the platform's terms are concerned. LinkedIn restricts the account, not the tool. Pace and content are your responsibility.
• Know their own cost unless you ask. Credits, tasks and tokens accumulate per run. An agent buying enrichment for every record before checking the CRM can spend a month's budget in an afternoon. Put the free checks first and ask every run for what it spent.
• Report failure honestly through a status light. A run can end cleanly having done nothing useful: a missing tool, an empty result, a blocked request. Read the results table and the run's own recap, not the green dot.
• Fix bad data or a bad rule. An agent applies the procedure it is given. A vague rubric gives vague results at scale, and a CRM with no loss reasons gives a lost-deal review with nothing to match.
• Replace the relationship. Agents remove reading and admin. The conversation with the buyer, the call on pricing and the decision to reopen an account stay with people.

Can AI agents replace employees?

AI agents replace tasks, not roles. In our own count, the 46 agents running on 29 October did the reading, the checks and the first drafts across sales, support, product and finance, while the rules they apply, the campaigns they feed and the conversations that matter stayed with named people. A team that adopts agents usually stops hiring for the admin half of a role and keeps the judgment half.

Can I make $1,000 a day using AI?

Not from an agent alone. An agent makes an existing business cheaper or faster to run; it does not create demand. The figures in "make money with AI" videos usually describe selling a service built on agents, and the work there is finding the customers, which no agent does for you.

Build or buy an AI agent for business?

Build when the job is one, its owner is technical, and it can live under one person's account. Buy when several jobs must run unattended, survive the person who built them, and show their cost per step.

For the sales and RevOps loop in more depth, see GTM engineering with AI agents. For the process this page embeds, with every step and connector, see revive deals that went cold.