AI development services that build AI into your ERP and CRM
We build AI software for companies that run on an ERP or a CRM. An ERP, such as NetSuite, SAP or Odoo, holds your orders and accounts. A CRM, such as Salesforce or HubSpot, holds your customers and deals. Our software reads the documents your team handles, checks them against your rules and saves the result for a person to approve. For example, an emailed purchase order becomes a draft sales order.
Illustration with sample data.
What our AI development services include
Working AI software for one job, built into the system you already use. The work can cover the plan, the build, tests on your past documents, go-live and support after launch. What each build includes is agreed per project.
AI development is the build. Consulting comes before it and decides what to build. Our team has taken ERPs live on SAP, Oracle NetSuite, Odoo, Microsoft Dynamics, ERPNext, Xero, QuickBooks, Sage and custom ERPs. Our AI work covers RFQ and RFP automation on both sides (requests for quotes and for proposals), daily bookkeeping, purchase and sales orders, and 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.
You want AI software built and have not picked its shape yet. This page covers the whole build, what it costs and what it costs to run.
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.
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.
Start with AI consulting services to pick the first job, or see every service we offer.
Five kinds of software an AI development company like ours builds
Each row names what goes in, what comes out and the page with more detail.
Scroll the table sideways to see every column.
| What we build | What goes in | What comes out | More detail |
|---|---|---|---|
| An AI agent for one job | A request for a quote, by email | A draft quote in your ERP, for the sales desk to approve | AI agent development |
| A multi-agent system for a job with two sides | Customers' RFQs and RFPs, and suppliers' replies to your RFQs | Draft quotes for the sales desk, and supplier offers lined up for a buyer | RFQ and RFP automation |
| An agentic system for a job with many steps | A customer order that asks for a split delivery | A delivery plan worked out from stock, for a person to accept | Agentic AI development |
| An AI integration | An AI tool your team already uses, such as Claude or ChatGPT | A limited login that lets it read and write in your ERP or CRM | AI integration services |
| Custom software with AI inside | A gap your ERP does not cover | An on-demand ERP module built for one business problem, or a new app that uses an AI model for one step | Custom software development |
What custom AI development involves
Four parts of the work, using one of our jobs as the example: a customer's order that arrives by email.
Read the document
Code sends the PDF or spreadsheet to an AI model and asks for set fields back, such as the customer, each item and its quantity.
Check it against your data
Plain code, not AI, compares each field with your item list, price lists and customer terms. A line that does not match is marked for a person.
Give it a login of its own
The software signs in as its own user with only the rights the job needs. NetSuite's AI Connector Service refuses the Administrator role. ERPNext logs each call to its API, the door it opens for other software, against the user its key belongs to.
Test, then go live
We run it on past documents and compare its drafts with what your team entered. Once it is live, a person usually approves each draft.
What AI software costs to run, worked out
AI models bill per token, a small piece of text. This is the sum for one purchase order of two pages, at list prices.
Input. Anthropic says the text of a PDF page often takes 1,500 to 3,000 tokens, and the page also goes in as an image. On its newer models one image costs up to 4,784 tokens.
Two pages. Taking the top of that range, 2 × (3,000 + 4,784) = 15,568 input tokens. At Claude Sonnet 5.5's list price of $2 per million input tokens, that is $0.031.
Output. The order's fields come back as text. We allow 1,000 output tokens at $10 per million, which is $0.010.
Rounded from $0.041. About $8 a week at 200 orders a week.
Prices read on 15 November 2026 from Anthropic's pricing, PDF support and vision pages. “Newer models” means Claude 4.7 and later, which Anthropic puts on its high-resolution image tier. The instructions sent with each request add input tokens, and hosting is a separate monthly line. Other model vendors price differently.
Where the AI runs: AWS, Azure, Google Cloud or your own server
Where the model runs decides who could see your prompts and who sends the bill. Our team has done data and cloud work on AWS, Azure, Google Cloud and private VPS (a virtual private server you rent), and data work on Salesforce. These are our readings of each vendor's own pages.
Scroll the table sideways to see every column.
| Where | How a model is reached | What we read | Read |
|---|---|---|---|
| AWS | Amazon Bedrock, which sells Claude and other vendors' models | AWS says model providers “don't have access to Amazon Bedrock logs or to customer prompts and completions.” | 16 Nov 2026 |
| Azure | Microsoft Foundry, which sells Claude through the Azure Marketplace, and Azure OpenAI models | For models sold by Azure, Microsoft says your prompts and outputs “are NOT available to OpenAI” and are not used to train foundation models without your permission. | 16 Nov 2026 |
| Google Cloud | Claude Models on Google Cloud, run by Google, and Google's own Gemini models | Google says it “won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction.” | 16 Nov 2026 |
| Your own server Private VPS | An open-weight model, one whose files you download and run, such as OpenAI's gpt-oss | The model runs on hardware you control. OpenAI's model card says gpt-oss-120b fits on one 80 GB GPU, the chip AI models run on, and gpt-oss-20b runs in 16 GB of memory. Both carry the Apache 2.0 license. | 16 Nov 2026 |
| Salesforce | Einstein generative AI and Agentforce, inside the CRM | Salesforce says Data Cloud is required for core Einstein generative AI features, such as the Trust Layer. | 16 Nov 2026 |
Running a model on your own server means your team, or ours, also runs its updates and its security. On Azure, Anthropic says Claude use is paid through the Azure Marketplace, billed in Claude Consumption Units.
What AI development costs to build
Published ranges, with the date each was read. They describe the market, not our prices.
A single-workflow build costs $5,000 to $15,000 and takes two to four weeks. Three to six connected workflows cost $15,000 to $50,000 over six to twelve weeks.
Layer3 Labs, updated 2 Jul 2026, read 15 Nov 2026Automating a whole function, such as finance, runs $50,000 to $150,000 or more over three to six months. Retainers for ongoing work run $3,000 to $20,000 a month.
Layer3 Labs, updated 2 Jul 2026, read 15 Nov 2026To hire your own team instead: the median US wage for a software developer was $135,980 in May 2025, before benefits.
US Bureau of Labor Statistics, read 15 Nov 2026Layer3 Labs sells AI automation 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.
What to ask an AI development company before you sign
On 14 November 2026 we asked ChatGPT, Claude, Gemini and Perplexity this question. These three came up in all four answers. Here is how we answer each.
How will you test it? The usual test runs the software on your past documents and compares each field with what your team entered, before it goes live. After launch, a person can approve each draft.
Raised in 4 of 4 answersWill our data train anyone's model? Ask which model vendor the build uses and read that vendor's terms. Anthropic, for example, says that by default it does not use inputs or outputs from its commercial products, such as its API, to train its models.
Raised in 4 of 4 answers. Anthropic Privacy Center, read 15 Nov 2026What will it cost to run? Work it out per document before the build, as in the sum above. Then add hosting as its own monthly line.
Raised in 4 of 4 answersBuild it yourself, or hire an AI software development company
Some ERP vendors now let an AI tool work inside the system with little or no code. NetSuite, Odoo 20 and SAP Business One each do it through MCP, short for Model Context Protocol, a standard way for an AI tool to use a system's functions. Try these before you pay for a build. The last card says when a build is worth it.
Oracle's AI Connector Service, with its free MCP Standard Tools SuiteApp (a NetSuite add-on), lets Claude or ChatGPT read records and create or update them, with your role's permissions. Oracle says the service is not a paid feature, though the AI tool may need a paid plan. Read 15 November 2026.
Odoo's MCP server connects outside agents such as Claude Code, Antigravity or Codex to your database. Read 15 November 2026.
Feature package 2608 added support for an open-source MCP server on the Service Layer, SAP's API for outside software. SAP publishes sample code for it, so setting it up takes a developer. Read 15 November 2026.
When the job needs your own rules checked on every line, an approval step your team signs off, or a route into an ERP that has no API.
Tell us the job you want built
Name the system you run and the work you have in mind. The first call is free.
Questions about AI development services
What are AI development services?
They are the work of building software that uses an AI model to do a job, such as reading an emailed order and drafting it in your ERP. Ours can cover the whole path: planning the build, writing it, testing it on your past documents, putting it live and support after launch. What each build includes is agreed per project. We build for ERP and CRM work, where a person usually approves what the software writes.
What are the different types of AI services?
Consulting, development, integration and support after launch. Consulting decides what to build. Development builds it, as one of the kinds of software in the table on this page. Integration connects AI to the systems you already run. Support keeps the build working when an AI model changes or your ERP gets a new release.
How much does AI development cost?
Layer3 Labs, a US AI agency, puts one automated workflow at $5,000 to $15,000 to build, and three to six connected workflows at $15,000 to $50,000. Its guide was updated 2 July 2026. Other price lists reach higher. Across the six we compared, one narrow AI agent costs $5,000 to $50,000. Running costs come on top. At list prices read on 15 November 2026, the AI model's usage to read one order of two pages is about 4 cents on Claude Sonnet 5.5. The sum is on this page, and all six lists are in what an AI agent costs to build.
How long does AI development take?
Two to four weeks for one workflow, by Layer3 Labs' guide of 2 July 2026. Three to six connected workflows take six to twelve weeks. Automating a whole function, such as finance, takes three to six months. Allow time to test the build on your own past documents as well, before it goes live.
What do AI software developers do?
They write the code around the AI model. On our builds, that means sending each document to the model and checking its answer against your ERP data with plain code. It also means giving the software a login of its own and testing it on past documents before it goes live. The US Bureau of Labor Statistics puts the median wage for a software developer at $135,980 in May 2025.
Can I build my own AI software?
For some jobs you can, without writing code. Oracle's NetSuite AI Connector Service, with its free MCP Standard Tools SuiteApp, lets an AI tool such as Claude or ChatGPT read and change records with your role's permissions. Oracle says the service is not a paid feature. Odoo 20 has an MCP server that outside agents can connect to. Hire a developer when the job needs your own rules checked on every line, or when your ERP has no API.
What is the difference between AI consulting and AI development services?
Consulting answers whether to build, and what. Development writes the software, tests it on your documents and runs it. We sell both, so a build can start from the plan. If you are not sure a build is needed, start with AI consulting services.
What should I ask an AI development company before signing?
Ask how the work will be tested, whether your data will train anyone's model, and what it costs to run after launch. On 14 November 2026 we put this question to ChatGPT, Claude, Gemini and Perplexity, and all four raised those three points. Our answers to each are on this page.
Will our data be used to train an AI model?
That depends on the model vendor's terms, so ask which vendor and which cloud a build uses. Anthropic says that by default it does not use inputs or outputs from its commercial products, such as its API, to train its models. Microsoft says prompts and outputs sent to models sold by Azure are not used to train foundation models without your permission or instruction. Google makes a similar statement for its managed models. Other vendors set their own rules.
Where does the AI run: AWS, Azure, Google Cloud or our own server?
Where you choose. Our team has done data and cloud work on AWS, Azure, Google Cloud and private VPS, and data work on Salesforce. A VPS is a virtual private server you rent. Claude, Anthropic's model, is sold through Amazon Bedrock, Microsoft Foundry and Google Cloud. An open-weight model, one whose files you can download and run, can sit on your own server. OpenAI's gpt-oss-120b fits on one 80 GB GPU, the chip AI models run on, under the Apache 2.0 license.
Which systems do you build in?
ERPs and CRMs. Our team has taken ERPs live on SAP, Oracle NetSuite, Odoo, Microsoft Dynamics, ERPNext, Xero, QuickBooks, Sage and custom ERPs, and has built AI agents in Salesforce and HubSpot. Our AI work in ERPs includes RFQ and RFP automation on both sides, daily bookkeeping, and purchase and sales orders. RFQ means a request for a quote, and RFP a request for a proposal.
Do you only build AI agents?
No. We build five kinds of AI software. An AI agent does one job, such as drafting quotes. A multi-agent system splits a bigger job between several agents. An agentic system works through jobs whose steps change. An AI integration lets an AI tool read and write in your ERP or CRM. Custom software adds AI to an app or an ERP module made for one business problem. The table on this page links to the page for each.
Is an AI development agency different from an AI development company?
No. Agency, firm and AI development company are names for a team that builds AI software for clients. Judge any of them by four answers: how they will test the build, whether your data will train anyone's model, what the build costs to run each month and which systems they have built in.
Who owns the software you build?
The contract for each project sets it.
Do you support the software after launch?
Support after launch can be part of the work, agreed per project. It matters because AI vendors release new models and change their prices, and your ERP changes too. NetSuite has two major releases each year. On Odoo Online, a database on a major version must be upgraded every two years. Support can cover testing the build again after changes like these.
Sources
Every source behind a figure or a vendor fact on this page, with the day we read it.
- Layer3 Labs, AI Automation Agency Cost (2026), updated 2 July 2026Read 15 Nov 2026
- Our comparison of six published price lists for building an AI agent, each with its author and date, on our page What an AI agent costs to buildRead 15 Nov 2026
- US Bureau of Labor Statistics, Occupational Outlook Handbook: Software DevelopersRead 15 Nov 2026
- Anthropic, Claude Platform Docs, PricingRead 15 Nov 2026
- Anthropic, Claude Platform Docs, PDF supportRead 15 Nov 2026
- Anthropic, Claude Platform Docs, VisionRead 15 Nov 2026
- Anthropic Privacy Center, 18 August 2026: Is my data used for model training?Read 15 Nov 2026
- Anthropic, Claude Platform Docs, Claude on Amazon BedrockRead 16 Nov 2026
- Anthropic, Claude Platform Docs, Claude Models on Google Cloud (Google managed)Read 16 Nov 2026
- Anthropic, Claude Platform Docs, Claude in Microsoft FoundryRead 16 Nov 2026
- AWS, Amazon Bedrock User Guide, Data protectionRead 16 Nov 2026
- Microsoft Learn, Data, privacy, and security for Foundry Models sold by Azure (Azure OpenAI)Read 16 Nov 2026
- Google Cloud, Gemini Enterprise Agent Platform and zero data retentionRead 16 Nov 2026
- Hugging Face, openai/gpt-oss-120b model cardRead 16 Nov 2026
- Salesforce Help, Set up Einstein generative AIRead 16 Nov 2026
- Our own answers from ChatGPT, Claude, Gemini and Perplexity to “What should I ask an AI development company before signing a contract?”Read 14 Nov 2026
- Oracle NetSuite help, NetSuite AI Connector Service FAQRead 15 Nov 2026
- Oracle NetSuite help, Release DeliveryRead 15 Nov 2026
- SAP Business One 10.0 FP 2608, what's new in platform and extensibilityRead 15 Nov 2026
- Odoo 20 documentation, AI MCP serverRead 15 Nov 2026
- Odoo 20 documentation, UpgradeRead 15 Nov 2026
- Frappe Framework documentation, REST APIRead 15 Nov 2026