Agentic AI development for jobs that take many steps
We are an agentic AI development company. We build AI that works through a job in your ERP step by step and picks some of the steps itself. An ERP is the software that runs your orders and accounts, such as NetSuite or Odoo. We set which steps the AI may choose, which tools it may use, whose login it signs in with and where it must stop. In the builds we design, a person usually approves before anything posts.
Illustration with sample data.
What agentic AI is, and what you get
Generative AI writes an email or a summary. Agentic AI acts. It picks the next step and calls a tool, such as your ERP's API, the door the ERP opens for other software. Then it reads the result and decides again, until the job is done or a limit stops it.
A build usually comes with the written plan behind it. The plan splits the job into fixed steps and AI steps, and lists the tools the AI may call, its login and its stop points. Tests on a copy of your system and the monthly running cost usually come before go-live. What each build includes is agreed per project. Anthropic, the maker of Claude, draws the same line between two kinds of system: workflows follow “predefined code paths,” while agents “dynamically direct their own processes and tool usage.”
Our team built a system of several AI agents that runs RFQs and RFPs end to end, on both sides. An RFQ is a request for a quote, and an RFP a request for a proposal. Our team has also taken ERPs live on SAP, Oracle NetSuite, Odoo, Microsoft Dynamics, ERPNext, Xero, QuickBooks, Sage and custom ERPs, and built AI agents in Salesforce and HubSpot.
Which page is for you
We have three pages on building AI. Pick the one that sounds like your job.
Your job has many steps that change from case to case. You need to decide which steps the AI may choose and where it must stop.
You have one job that repeats, such as typing emailed orders into your ERP. You want a program that does it and hands each draft to a person.
You want AI software built and have not picked its shape yet. That page covers the whole build and what it costs to run.
Start with AI consulting services to pick the first job, or see every service we offer.
How we split agentic AI solutions between fixed steps and AI steps
The usual split puts fixed code wherever the rule is known, and the model only where the input changes. The example is an emailed purchase order that becomes a draft sales order.
Scroll the table sideways to see every column.
| Step | Runs as | Why |
|---|---|---|
| Read the email and its PDF | The model | Layouts differ by customer, so fixed code breaks on the first new form. |
| Match each line to your item list | Fixed code | A line with no exact match goes to a person, so the code makes no guess. |
| Apply price lists and customer terms | Fixed code | Prices are rules you already keep in the ERP. |
| Work out an unusual request, such as a split delivery Agent step | The model, inside a step cap (a set limit on its steps) | In this example it may look up stock and suggest a plan to a person, but it has no tool that saves the order. |
| Save the record | Fixed code, as a draft | In this example a person approves the draft before anything posts. |
Anthropic's advice is to start with the simplest system that works, which “might mean not building agentic systems at all.” Agentic systems, it says, “often trade latency and cost for better task performance.”
The architecture of an AI agent
Four parts. An LLM, short for large language model, is the kind of AI model behind Claude and ChatGPT. Anthropic describes the base of an agent as “an LLM enhanced with augmentations such as retrieval, tools, and memory.”
The model
The LLM reads the input, such as an emailed order, and picks the next step.
Tools
Functions the agent may call, such as “create draft sales order” in your ERP. NetSuite, Odoo 20 and SAP Business One each document a route to them through MCP. MCP, short for Model Context Protocol, is a standard way for an AI tool to use a system's functions.
Context
What the agent looks up before it acts: your item list, price lists and customer terms, read at the moment it needs them.
Limits and stop points
Odoo 20 agents ask the user to confirm when they reach their tool-call limit. The MCP standard says a person should be able to deny any tool call.
Five patterns behind agentic AI development, with ERP examples
Anthropic names five workflow patterns in its guide to building agents. It files all five under workflows, not agents, because your code sets the path. The ERP example for each pattern is ours, not Anthropic's.
Scroll the table sideways to see every column.
| Pattern | What Anthropic says it does | An ERP example |
|---|---|---|
| Prompt chaining | Splits a task into steps, “where each LLM call processes the output of the previous one.” | Read an emailed order, then check the lines it found, then write a short note for the approver. |
| Routing | It “classifies an input and directs it to a specialized followup task.” | Sort a shared inbox: requests for quotes to one agent, orders to another, bills to finance. |
| Parallelization | Models “work simultaneously on a task and have their outputs aggregated programmatically.” | Read five supplier replies at the same time and put every offer in one table. |
| Orchestrator-workers Many agents | In it, “a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results.” | A long RFP: one model splits it by topic, workers draft each part, and the first model puts the draft together. |
| Evaluator-optimizer | In it, “one LLM call generates a response while another provides evaluation and feedback in a loop.” | One call drafts an answer for a proposal. A second checks it against wording you have approved and sends it back with notes. |
Anthropic's guide says to “add multi-step agentic systems only when simpler solutions fall short.” We read it on 15 November 2026.
A multi-agent system our team built: RFQs and RFPs on both sides
One system, several AI agents, and two sides of the same trade. We do not name the client or give numbers from the project.
The sales side
A customer sends a request for a quote, an RFQ, or a request for a proposal, an RFP. The work is to read each line, match it to the item list, price it and draft the quote or the proposal for a person to approve.
The buying side
The business needs to buy. RFQs go out to suppliers, their replies come back, often as PDFs, and the offers are lined up so a buyer can pick one.
Work like this uses more than one of the patterns above. Sorting what arrives is routing. Reading several supplier replies at once is parallelization. How the agents split the work, and where a person approves, is agreed per project. A table of the steps is on our AI agent development page, and the full page is RFQ and RFP automation.
How to choose an agentic AI development company
The same assistant gave us two different shortlists of agentic AI companies nine minutes apart. These four questions work whichever list you get.
3 firms in common. On 14 November 2026 we asked the ChatGPT app “Best agentic AI development company for enterprise workflows” twice, nine minutes apart. The first answer listed 7 firms and the second listed 11. Only 3 were on both lists.
Our saved answers, 14 Nov 2026About 130 real vendors. Gartner estimated that only about 130 of the thousands of agentic AI vendors are real. It warns of “agent washing,” where older products are renamed as agents.
Gartner, 25 Jun 2025, read 15 Nov 2026Which tools can it call?
Ask for the list, and which tools write. Odoo's MCP server shows five lookup tools by default, such as Search and Read group, and hides the rest until someone switches them on.
Whose login does it use?
Ideally a role of its own. Oracle's NetSuite AI connector will not run under the Administrator role.
Where does it stop?
At a step cap, and at an approval before anything posts. OWASP lists failing to “independently verify and approve high-impact actions” as a cause of excessive agency.
What does it cost to run?
AI usage and hosting as monthly lines, next to the build price. Each step that runs as fixed code makes no call to the model, so the split above changes the monthly bill.
Agentic AI solutions the vendors already ship
Before a build, check what your ERP or CRM vendor sells. These are our readings of each vendor's own pages, with the date we read them.
Scroll the table sideways to see every column.
| System | The vendor's agents | What we read | Read |
|---|---|---|---|
| NetSuite | NetSuite agents, and the AI Connector Service for outside AI tools | Oracle says its agents “don't replace NetSuite records, reports, approval workflows, or your review of results.” The connector shares its request limits with your other integrations, and one AI action may send several requests. | 15 Nov 2026 |
| SAP | Joule Agents, Joule Assistants and Joule Studio | SAP says an agent gateway “governs agent communications, tool use, and human oversight.” SAP Business One is not on SAP's Joule lists. | 16 Nov 2026 |
| Microsoft Dynamics 365 Business Central | Sales Order Agent | It reads customer requests sent by email and prepares quotes and orders. Named users review and approve every outgoing message before it is sent. | 16 Nov 2026 |
| Odoo 20 | AI agents inside Odoo, plus an MCP server for outside agents | Odoo says “only approved AI clients can request access to the database.” Each connection shows up as a key, and deleting the key cuts that client off. | 15 Nov 2026 |
| Salesforce | Agentforce | Priced by use, with Flex Credits or Conversations, or by user license. | 16 Nov 2026 |
| HubSpot | Breeze agents, such as the customer agent | The customer agent answers from your synced content and hands the chat to a person when its confidence is low. Agents use HubSpot Credits. By default, usage features pause when the month's included credits run out and no more have been bought. | 16 Nov 2026 |
| Sage | MCP servers for Sage Intacct and Sage X3 | Sage lists read-only access for the Intacct MCP server. Access to the X3 MCP server is requested through Sage's AI Gateway page. | 16 Nov 2026 |
A custom build fits where these stop: a document that arrives by email in a system with no agent for it, rules of your own on every line, or work that crosses two systems. More by system: NetSuite, SAP, Salesforce, HubSpot and Sage.
What agentic AI development costs
Published ranges, with the date each was read. They describe the market, not our prices.
A production MVP workflow, meaning a first working version of one automation built end to end, costs $5,000 to $25,000. A full enterprise AI system costs $50,000 to $200,000 or more.
Layer3 Labs, AI consulting fees guide, updated 27 Oct 2026, read 15 Nov 2026Anthropic says the autonomy of agents brings “higher costs, and the potential for compounding errors,” and recommends “extensive testing in sandboxed environments.”
Anthropic, 19 Dec 2024, read 15 Nov 2026Layer3 Labs sells AI consulting itself, so read its ranges as one firm's view. Six firms' price lists, side by side, are in what an AI agent costs to build.
Tell us the job you want an agent to finish
Name the system you run and the document that starts the work. The first call is free.
Questions about agentic AI development
What is agentic AI, in plain words?
Agentic AI is AI that works through a job in steps, where a chatbot gives one answer. It reads what it needs, uses a tool such as your ERP's API, checks the result and picks the next step. It keeps going until the job is done or a limit stops it. Anthropic, the maker of Claude, describes agents as systems where models “dynamically direct their own processes and tool usage.”
What does an agentic AI development company do?
It designs and builds AI that plans steps and calls tools to finish a job inside your software. With us, the work usually starts with the split between fixed code and the model's choices. Then come the tools the agent may call and the login it uses. Stop points and approvals follow, then tests on a copy of your system and the monthly running cost. What each build includes is agreed per project.
What is the difference between agentic AI and an AI agent?
Agentic AI is the wider term, and an AI agent is one kind of it, where the model picks its own steps. Anthropic also counts fixed workflows as agentic AI. On our site, AI agent development covers one job in one system, and this page covers jobs where the steps change. Four vendors' definitions are compared in agentic AI vs AI agents.
What is agentic AI versus generative AI?
Generative AI makes content when asked, such as a draft email or a summary. Agentic AI uses the same kind of model to choose actions: it calls a tool, reads the result and takes the next step toward a goal. A model that drafts a sales order from an emailed order, by looking up items and prices, is working as an agent. The same model writing the reply email is generative.
Is ChatGPT an agent or an LLM?
It can work as an agent. Anthropic describes agents as “typically just LLMs using tools based on environmental feedback in a loop.” An LLM is a large language model. OpenAI's launch post of 17 July 2025 describes an “agent mode” in which ChatGPT “requests permission before taking actions of consequence.” OpenAI now marks that post as outdated, so check its current pages for today's features.
What is the architecture of an AI agent?
Four parts. A language model picks the next step. Tools are the functions it may call, often offered through MCP, the Model Context Protocol. Context is what it looks up first, such as your item list and price lists. Limits close the loop: a step cap, a login without administrator rights, drafts in place of posted records, and a person who can say no.
What does an agentic AI developer do?
Most of the work sits around the model. The developer writes the tools the agent calls and describes each one. They set the rules and stop points and build tests from real documents. Then they run the agent on a sandbox, a safe copy of your system, before it touches live data. Anthropic says that on one of its own agents, its team “spent more time optimizing our tools than the overall prompt.”
How can I hire agentic AI developers?
You can hire them onto your team or hire a firm that has built agents in your system. Either way, ask four things: which tools the agent can call, whose login it uses, where it stops and what it costs to run each month. For a sense of in-house cost, the US Bureau of Labor Statistics puts the median software developer wage at $135,980 in May 2025.
How can I tell if an agentic AI company is the real thing?
Ask it to show the tools its agent calls and the points where the agent stops. Gartner calls the problem “agent washing,” where products such as chatbots and AI assistants are renamed as agents without gaining the abilities. In a release of 25 June 2025, Gartner estimated that only about 130 of the thousands of agentic AI vendors are real.
When should a business not build an agentic system?
When a fixed workflow or the vendor's own AI already does the job. Anthropic's guide says the simplest solution “might mean not building agentic systems at all.” Gartner's advice runs the same way: use agents where a decision is needed, plain automation for routine work and an assistant for simple lookups. For NetSuite, we check Oracle's AI Connector Service first, which Oracle says is not a paid feature.
How much does agentic AI development cost?
Layer3 Labs, a US AI agency, prices a first working version of one automation at $5,000 to $25,000. It puts a full enterprise AI system at $50,000 to $200,000 or more. Its guide was updated 27 October 2026 and read on 15 November 2026. Layer3 sells this work itself, so read it as one firm's view. Ask for the running cost as monthly lines too: AI usage and hosting. More price lists are in what an AI agent costs to build.
How do you stop an agent from doing something it should not?
With limits inside the systems it touches. The usual setup gives the agent a role that cannot reach administrator functions, and Oracle's NetSuite AI connector refuses to run under the Administrator role at all. The agent can be set to write drafts that a person approves, and to stop at a step cap. OWASP's example of least privilege is an agent that can read one table and cannot change any records.
Do you offer agentic AI consulting before a build?
Yes. Our agentic AI consulting starts with the split on this page: which steps of the job run as fixed code and which need the model. It ends in a written brief, and the brief may say to use your vendor's own AI in place of a new agent. Our AI consulting services page sets out what the brief covers.
Have you built a multi-agent system?
Yes. Our team built a system of several AI agents that runs RFQs and RFPs end to end, on both sides. A customer's request becomes a draft quote, and requests for quotes go out to suppliers. An RFQ is a request for a quote, and an RFP is a request for a proposal. We do not name the client or give numbers from that project. The section on this page about that system names two of Anthropic's patterns that this kind of work uses: routing and parallelization.
Which systems do you build agentic AI in?
ERPs and CRMs. Our team has taken ERPs live on SAP, Oracle NetSuite, Odoo, Microsoft Dynamics, ERPNext, Xero, QuickBooks, Sage and custom ERPs. We have built AI agents for RFQs, bookkeeping and orders in ERPs, and agents in Salesforce and HubSpot, two CRMs. A CRM is where you keep customers and deals.
Sources
Every source behind a figure or a vendor fact on this page, with the day we read it.
- Anthropic, Building effective AI agents, 19 December 2024Read 15 Nov 2026
- OpenAI, Introducing ChatGPT agent, 17 July 2025 (marked outdated by OpenAI)Read 15 Nov 2026
- Gartner, press release of 25 June 2025 on agentic AI projectsRead 15 Nov 2026
- Our own answers from the ChatGPT app to “Best agentic AI development company for enterprise workflows,” asked twice nine minutes apartRead 14 Nov 2026
- Model Context Protocol specification, ToolsRead 15 Nov 2026
- Odoo 20 documentation, AI MCP serverRead 15 Nov 2026
- Odoo, Odoo 20 release notesRead 15 Nov 2026
- Oracle NetSuite help, NetSuite AI Connector Service FAQRead 15 Nov 2026
- Oracle NetSuite help, Required Features and PermissionsRead 15 Nov 2026
- SAP Business One 10.0 FP 2608, what's newRead 15 Nov 2026
- Oracle NetSuite help, Agents OverviewRead 15 Nov 2026
- SAP, Joule business AI solutionsRead 16 Nov 2026
- Microsoft Learn, Sales Order Agent overview, Business CentralRead 16 Nov 2026
- Salesforce, Agentforce pricingRead 16 Nov 2026
- HubSpot Knowledge Base, Understand the customer agentRead 16 Nov 2026
- HubSpot Knowledge Base, Test and control agent credit usageRead 16 Nov 2026
- HubSpot Knowledge Base, Manage HubSpot CreditsRead 16 Nov 2026
- Sage Developer, Sage Intacct AI GatewayRead 16 Nov 2026
- Sage Developer, Sage X3 MCP Server (1.0) overviewRead 16 Nov 2026
- OWASP Gen AI Security Project, LLM06:2025 Excessive AgencyRead 15 Nov 2026
- Layer3 Labs, AI Consulting Fees 2026 guide, updated 27 October 2026Read 15 Nov 2026
- US Bureau of Labor Statistics, Occupational Outlook Handbook: Software DevelopersRead 15 Nov 2026