AI agents for business operations: back-office work done inside your ERP, CRM and inbox, with approvals
AI agents for business operations are programs in which a language model reads incoming requests, looks up records in your systems and completes back-office tasks such as order entry, invoice matching and vendor follow-up, inside rules you set. Goldenroz builds them with per-user permissions, an approval step before anything costly, and a log of every action. Projects start at $25,000 at a fixed price per phase. A prototype takes about one to two weeks and production roughly four to eight weeks.
From $25,000, fixed price per phase
Updated 2026-11-14. By Shai Goldenberg
What usually goes wrong
- Orders arrive as emails and PDFsA person reads each purchase order, finds the customer, checks stock and price, and types it into the ERP. Volume grows and the team does not.
- Invoices that do not matchSupplier invoices are compared by hand against purchase orders and receipts. The exceptions are a small share, but someone has to open every invoice to find them.
- Follow-ups nobody ownsLate deliveries, missing documents and unpaid invoices each need a polite message and a note in the system. They slip when the week gets busy.
- One question, five systemsAnswering where an order stands means checking the ERP, the carrier site, the CRM and two inboxes.
What is included
- Intake from the channels you useShared mailboxes, PDFs, forms and messaging channels are read and turned into structured fields.
- System access through MCP or APIsConnections to ERP, CRM, SQL databases and Microsoft 365, read-only by default. Queries are approved in advance and the model never writes its own SQL.
- Approval queueAny action that spends money, changes a customer record or sends an external message waits for a named person until you decide to widen the limits.
- Durable runsJobs survive crashes and timeouts through leases, retries and a scheduled sweep that picks up stalled work, so a task is done once and not twice.
- Cost and audit loggingToken cost per run and a record of every tool call, input and result that your finance or audit team can read.
Enterprise AI agents: access, audit and where the data stays
Enterprise AI agents are judged by security and audit teams before operations ever uses them. They ask what the agent can read, where the data goes and what is logged.
Our answer is built into a product Shai, our CTO, owns: an on-premises MCP gateway of about 27,000 lines with 13 connector types. It lets Claude or ChatGPT answer from ERP, CRM, SQL and Microsoft 365 systems read-only, with per-user permissions, PII masking, an audit log, approved queries only and defenses against prompt injection. It installs at your site with Docker Compose. The details are on the AI integration page.
Agentic AI development in operations starts with one narrow task
Agentic AI development goes wrong when the first project is an assistant for everything. We pick one process with clear inputs, a system of record and a measurable error rate. The prototype runs on your real past examples in a test environment, so you see the failures before paying for production.
The design splits the task between model and code. The model reads the document and proposes the fields. Plain code checks the customer exists, the price matches the contract and the quantity is in stock. Anything that fails a check goes to a person, and the agent starts with every action approved.
AI agents for small business: when off-the-shelf is the better buy
Much of what is written about AI for small business describes features inside tools you already pay for: the assistant in your accounting package, your help desk or your email. For a company with standard processes, those are the right purchase.
Custom AI agents for small business make sense only when the process is specific to you and the volume is high enough that $25,000 is repaid in labor or errors avoided. If you are unsure which side you are on, AI consulting is the place to start.
How the work runs
Scoping call
Before the projectWe map the process, the systems involved and the cost of a wrong action. You get a fixed price for the first phase.
Prototype
About 1 to 2 weeksA working agent on one narrow task, run against real examples in a test environment. You see where it fails before you pay for production.
Production build
Roughly 4 to 8 weeksIntegrations, permissions, fallbacks, approval steps, logging and the evaluation set. The range depends on how many systems the agent connects to.
Launch with limits
At launchThe agent starts on a share of the volume, or with every action approved by a person. Limits widen as the logs show it holds.
What it costs
Every phase has a fixed price agreed before work starts. Model usage is paid to the model provider from your own account.
- Focused agent or agentic workflowFixed price per phaseFrom $25,000
- Larger AI platformsSeveral agents, many integrations or a full product$100,000 and up
- Scoping callEnds with a written scope and a price$0
- Model usageBilled to your own accountPaid to the provider
Who it is not for
- Companies that have not yet chosen a process to automate. Start with AI consulting.
- Agents that move money or delete data with no human approval.
- Budgets under $25,000.
Who does the work
Shai leads the technology side: business software, SaaS and mobile apps, payments and escrow systems, Postgres security, AI agents built with Claude Code and MCP, and IT and infrastructure projects with senior infrastructure engineers.
Questions buyers ask
How much do AI agents for business cost to build?
A focused agent or agentic workflow starts at $25,000 at a fixed price per phase. Larger programs with several agents or many integrations run $100,000 and up. Model usage is billed by the provider to your own account. See pricing.
Which business operations are good candidates for an agent?
Tasks with a high volume of similar requests, written inputs and a system of record to check against. Tasks that need judgment about people or contracts are poor candidates.
Can the agent write to our ERP?
Yes, through the API or interface your ERP provides, and behind an approval step at first. Access starts read-only.
How long before an operations agent is in production?
A working prototype on one narrow task typically takes one to two weeks. Production takes roughly four to eight weeks, depending on the number of systems.
Does the agent replace our operations staff?
It takes the repetitive reading and typing. Your staff handle the exceptions and approvals. We make no promise about headcount.
Guides on this topic
- Seven Claude Code practices from a production codebase: CLAUDE.md rules, guard hooks, worktrees, permission envelopes and CI review. Apply them to your team.
- Claude Code hooks run your code at fixed points and can block actions. Subagents isolate work in their own context. See a working PreToolUse guard hook.
- What an AI agent costs to build and to run: the five cost drivers, project shapes by effort, and a worked running-cost example. See where your project lands.
Talk to the person who will do the work
One hour, no sales team. You talk to Shai Goldenberg, the CTO, about your systems or your product, and you leave with a clear next step and a price range.