AI agents for business: a sales agent, measured

What are AI agents for business?

AI agents for business are software that takes a task, decides which tools to use and works through them: looking up a customer, pricing an order, drafting the reply. Anthropic defines agents as systems where language models “dynamically direct their own processes and tool usage.” In most businesses that use them well, a person still makes the final call.

This note shows one in daily use. Since February 2026 we have built a sales agent for Alliance Chemical, a chemical distributor in Taylor, Texas. From August 3 to September 30, 2026, it made 38,328 tool calls across 20 connected systems and sent 1,811 emails. Disclosure: Andre Taki, a PHASEKIND founder, is Alliance Chemical’s Chief Commercial Officer and the agent’s main operator.

What can an AI agent do for a business?

At Alliance, it runs most of a sales desk. Figure 1 sorts those two months into eight jobs: email, customers and pipeline, quotes, orders and invoices, purchasing, shipping, documents and corrections.

The actions are the part people notice. The agent opened 565 deals and updated 433, entered 161 orders in ShipStation, and created 129 QuickBooks invoices and emailed 95. It built 61 certificates of analysis and 57 certificates of conformance. When something went wrong it fixed it: 24 refunds, 17 voided invoices and 18 corrected shipping addresses.

Most of the work, though, was finding things. Against 4,027 actions it made 18,384 lookups: 2,595 customer searches in QuickBooks, 1,242 searches of past email and orders, 1,701 product lookups, 857 invoice checks and 308 lookups of what Alliance last paid a supplier. That ratio is the case for an agent. Each of those is a search a person would otherwise run by hand in another system; the agent runs them and brings back the answer.

How does an AI agent handle a quote?

The agent does the lookups and the person keeps the decision. A quote request arrives by email. The agent reads it, finds the customer in QuickBooks Online, searches past email and orders, looks up the product in Shopify and rates the freight in ShipperHQ. It then builds a QuickBooks estimate for a Net 30 account or a Shopify draft order for a card customer, and drafts the reply. A person reads it and decides whether it goes.

In two months that path ran 338 times and produced 136 draft orders and 47 estimates. The machine steps are quick. Median times from the log: 1.0 second to find the customer, 3.8 seconds for a freight rate, 6.6 seconds for a full quote run and 2.1 seconds to send. Sent mail rose from 754 in August to 1,057 in September.

Figure 1C shows one real quote, redacted. On September 14, 2026, a repeat customer asked at 11:05 a.m. for four drums of two products. The agent found their November order, matched its prices, rated freight to the same delivery site, created the estimate and drafted the reply. The only human input was one word, “send,” and the reply went out unedited at 11:11. The customer sent a purchase order two days later.

Each outside system plugs in through the Model Context Protocol, “an open-source standard for connecting AI applications to external systems.” Outlook, QuickBooks, Shopify, ShipStation and ShipperHQ each sit behind their own server, so the agent sees one set of tools instead of a row of logins.

AI agents vs agentic AI: what’s the difference?

In practice the two terms name the same thing; Gartner’s June 2025 forecast uses them interchangeably. The difference that matters is between an agent and a workflow. Anthropic puts it this way: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths,” while agents keep “control over how they accomplish tasks.”

A chatbot answers questions in a conversation; an agent takes actions in other systems. Gartner’s advice is to match the tool to the job: “start by using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.” Alliance runs both. An overnight run stages draft replies in fixed steps and has no tool that can send or write. During the day, the agent picks its own lookups for each email.

Should an AI agent send emails or move money on its own?

Not this one, as of October 2026. It has run in supervised mode since February 25, 2026. Three higher levels are configured, from auto-sending follow-ups up to most email, each with a minimum draft score and a daily cap, and all three are switched off. Moving up a level requires a 95% approval rate over at least 20 drafts.

Some email goes to a person at any level. First contacts, complaints and anything that mentions legal action are always escalated; the log shows 301 first contacts and 235 complaints routed that way since February. Seventeen coded rules check drafts, quotes and sends. Among them, the sending address is fixed and the recipient has to match the email thread. Two teammates use the same connections through a gateway that blocks their agents from sending email, refunding a card or voiding an invoice.

The sends show it. They peak between 9 and 10 a.m. Central, fall from 437 on Mondays to 232 on Fridays, and total 63 across all weekends. That is a working week, which is what supervision looks like in the data: email goes out when someone is there to approve it. Anthropic’s guidance gives the reason: “The autonomous nature of agents means higher costs, and the potential for compounding errors.”

Are AI agents worth it for a small business?

They can be where the same lookups repeat all day. As a general purchase, be skeptical. Gartner predicts “Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.” It also warns about “agent washing,” estimating “only about 130 of the thousands of agentic AI vendors are real.”

Expect the build to be mostly connections and rules. Alliance’s repository had its first commit on February 14, 2026, and 321 commits by September 24. The changes that most shaped its behavior were not model changes: a policy engine for sends in May, limits on teammates’ agents in May and June, and the overnight staging run in August. Anthropic recommends “finding the simplest solution possible, and only increasing complexity when needed”, so start with one desk where the lookups repeat.

The running cost is a subscription. The agent runs inside Claude Code on a Claude Max plan, which Anthropic lists “From $100” per month, so there is no per-token bill. Priced at Anthropic’s API list rates instead, September’s usage would have come to $5,504.52, a figure that includes work on building the agent itself.

We can show the volume the agent handled, its cost and the controls around it. We don’t yet have a clean measure of revenue the agent added, so we don’t claim one.