How long it takes to build an AI agent for a manufacturer or distributor

A first AI agent takes about four weeks as a pilot on one job, then 8 to 12 more weeks to reach production with your systems connected. Without a pilot, add one to two weeks of discovery at the start. This guide shows the work week by week, what connecting your ERP takes and what the agent costs to run.

Derik Lawlis Founder, ThriveAI · leads every ThriveAI project

How to build an AI agent

These steps assume you start with one job in one system. Each step produces something you can check before the next one begins.

1. Pick one job with a known cost. Choose work your team repeats every week, such as turning emailed requests for quote into draft quotes or keying purchase orders into your ERP. Write down who does it and how many hours it takes each week. Then estimate what one mistake costs you.
2. Collect real examples. Pull 30 to 50 cases from the last quarter, each with the answer your team gave. These cases become the test set that tells you whether the agent is right.
3. Decide what the agent may read and change. Start with read access only. The agent drafts, and a person approves every action that changes a record or reaches a customer.
4. Choose a model and where it runs. Anthropic recommends Claude Opus 5.5 as the starting point for most work, and you can move steps to a cheaper model once your test set shows it holds up. If the job touches personal information, check where the provider processes it before you send any, as the privacy section explains.
5. Build the first version against your examples. Connect it to one system, run it on the test set and fix the most common failures. Show the output every week to the people who do the job.
6. Run it beside your team. For two to four weeks, the agent drafts while your team works as usual. Compare time and accuracy against the manual process, then decide whether the agent may write into your system.
7. Keep testing after launch. Rerun the test set whenever you change a prompt or a model, and add every case that a reviewer corrects.

The sections below expand these steps into a pilot and a production build, with the time each one takes.

What counts as an AI agent

Vendors use the word agent loosely, so settle the terms before you pay for a build.

For a first build, choose a workflow agent: one job, one team, access to one system and a person at the checkpoint. A multi-agent system adds coordination work that a single job rarely needs.

Three ways to scope the build

A build takes one of three shapes. Each has its own timeline and deliverables, so pick the one that matches how well you already know the job.

1. Pilot · 4 weeks

A pilot proves the pattern on one job before you commit to more. It is a working system on your real data with a deliberately narrow scope. It fits when the job takes at least five hours of someone's week and you want evidence before you approve a production build. If you already know the scope, skip the pilot.

• Derik interviews two to five of the people who do the job, for 30 to 60 minutes each, and maps the job with its weekly hours and failure points.
• A working agent runs on your real data, and a person approves every output before anything is written.
• A first test set of 30 to 50 labelled cases measures the agent.
• The agent connects to one system, such as QuickBooks, SAP Business One, HubSpot or Microsoft 365, with read-only or sandbox access.
• You receive a written go or no-go recommendation with the scope of a production build.

The pilot leaves out integration with several systems, single sign-on, production dashboards and long-term operating contracts. That keeps it easy to stop.

2. Production · 8 to 12 weeks

Production turns the pilot into a system that runs every day. It connects to your ERP or CRM, keeps an audit trail, runs its test set on every change and stops for a person at each consequential action.

• The agent connects to one to three of your systems, with write access where the job needs it.
• Production sign-in uses OAuth or service accounts, with secrets kept in a managed store.
• Tracing records every model call and tool call.
• The test set grows to 100 or more versioned cases and runs on every prompt or model change.
• A person approves every consequential action, such as a write to the ERP or a message to a customer.
• The audit trail records every model call and every human decision, and runbooks cover what to do when something breaks.

The depth of the integration sets the timeline. A clean QuickBooks Online account paired with HubSpot fits in about 8 weeks. A multi-company SAP Business One server on your premises, with custom user-defined fields, takes closer to 12. Plan for 12 weeks when the data needs cleaning, when part of the system runs on premises or when the approval step needs sign-off from senior management.

3. Operating retainer · ongoing

A retainer keeps the production system healthy after launch. Model providers publish retirement dates for their models, as Anthropic does for each Claude model, and connectors break when an upstream API changes.

• Monitoring covers uptime, response time, cost and error rates, with a response to any alert within 24 hours.
• The test set runs every week, and you get a report of any regression.
• Prompts and models are reviewed every month and updated when a provider changes or retires a model.
• One new job is scoped with your team each quarter.
• A quarterly review covers results and running cost, and it picks the next job to automate.

The retainer grows with the number of jobs the agent runs. One job with light monitoring is the smallest version, and several jobs with a new build each quarter is the largest.

Week by week: a 4-week pilot

This is the weekly plan for a pilot. Holidays and staff availability can shift it by a few days.

• Week 1. Map the job. Three to five 45-minute interviews with the people who do the job today capture its inputs, outputs, failure points, weekly hours and the cost of a mistake. The week ends with an agreed scope and a one-page map of the job with its baseline cost.
• Week 2. Build the first working agent. The agent is built against 30 to 50 real cases from the last quarter. It starts on Claude Opus 5.5, and steps move to Claude Sonnet 5 or Claude Haiku 4.5 later if the test set shows they hold up. The agent only reads, and its output goes to a review screen.
• Week 3. Set up the test set and the checkpoint. The cases are labelled and the test set runs. The prompt and tools are tuned against the top failure modes. The reviewer screen is designed, including what the reviewer can edit and what gets logged.
• Week 4. Decide. The agent runs on a fresh batch of cases kept out of the test set and is compared with your team on time and accuracy. The written go or no-go covers the production scope, the systems to connect, the checkpoint design and who runs the system day to day.

Week by week: an 8 to 12 week production build

This plan assumes the pilot ran cleanly and ended with a go. Without a pilot, add one to two weeks of discovery at the start.

• Weeks 1 and 2. Plan the integration and get access. Map the systems, including the ERP, CRM, Microsoft 365 and file storage. Gather credentials and OAuth scopes, and list the exact records to read and write. In QuickBooks those are customers, items, invoices and payments, and in SAP Business One they are business partners, items, sales orders and journal entries. Confirm sandbox and production access and decide where the checkpoint sits. Your IT lead signs off on the written plan.
• Weeks 3 and 4. Build the read path with logging. Structured logging starts on the first day, so every model call, tool call and response carries a timestamp. The test set grows from the pilot's cases to 100 or more. Shadow runs begin, where the agent works alongside your team and takes no action.
• Weeks 5 and 6. Add the write path and the checkpoint. Writes go behind an explicit approval step. Approval can arrive in Outlook or on a custom review screen. The audit trail captures the model output, tool calls, reviewer, decision and timestamps.
• Weeks 7 and 8. Add dashboards and runbooks. Your team gets a dashboard of throughput, test-set accuracy, cost per run, error rate and queue depth. Runbooks cover an upstream API failure, unexpected model output, a reviewer who disagrees with the agent and a regression after a prompt change. At the end, your team can run the system without the build team in the room.
• Weeks 9 to 12, if needed. Extra scope goes here, such as French-language output with French test cases or a second integration. A privacy impact assessment under Quebec's private sector privacy act also fits in these weeks.

What connecting to your systems takes

Most of the build time goes into connecting the agent to your systems. The four most common targets follow, with the time each one usually takes.

QuickBooks Online

The QuickBooks Online Accounting API authorizes apps with OAuth 2.0. The Intuit App Partner Program guide splits calls into two categories. Core calls cover most data-in operations, such as creating or updating invoices, bills, customers and vendors, and they are unmetered. CorePlus calls cover most data-out operations, such as reading accounts and reports, and they are metered.

The free Builder tier includes 500,000 CorePlus calls a month and blocks calls above that. The paid Silver, Gold and Platinum tiers charge a monthly program fee and include more CorePlus calls. To publish an app, you complete a self-attested assessment questionnaire, and a listing on the Intuit App Marketplace adds further reviews, including a security review.

Most of the friction sits in token handling for each connected company file and in the tax lines on invoices. Plan 3 to 5 days for a clean company file.

SAP Business One

SAP describes the SAP Business One Service Layer as a REST API built on HTTP and OData that gives web access to Business One services and objects. It is the path for an agent that runs in the cloud.

Most of the time goes into mapping your company's user-defined fields and approval procedures. A multi-company setup multiplies that work, and a server on your premises needs a secure network path, such as a VPN. Plan 2 to 4 weeks for a moderately customized system, and longer for a heavily customized multi-company one.

HubSpot

HubSpot publishes which of its APIs each subscription tier includes, and rate limits differ by tier. Some APIs, such as CRM forecasts, need Sales Hub Professional or higher, so check the tier table against the records your agent needs before you scope the work.

Most of the work is mapping custom properties, especially in an account where sales and marketing have added fields for years, and staying inside bulk-operation rate limits. Plan 3 to 5 days for a clean account.

Microsoft 365

Microsoft Graph exposes REST APIs for Outlook, SharePoint, OneDrive, Teams, calendars and Dynamics 365 Business Central. Most Graph APIs are standard APIs, available within documented usage thresholds at no cost beyond your user licenses. A smaller set of high-capacity and advanced APIs can be metered, and those require an Azure subscription.

Power Automate Premium is charged per user per month, paid yearly. You need it only when Power Automate runs the workflow, and an agent that calls Graph directly does not.

Most of the time goes into tenant admin consent and the choice between delegated and application permissions, after you register the app in Microsoft Entra ID. An agent that runs unattended usually needs application permissions, and a larger tenant takes longer to approve them.

How model usage is charged

Model usage is usually the smallest line in an agent's budget. Integration and testing take more. Providers charge per million tokens in US dollars, with one rate for input and another for output, and Anthropic's pricing page puts one token at roughly 0.75 of an English word. That page and OpenAI's pricing page list the rates, and Claude API pricing covers every Claude model. The table shows what each model suits, from Anthropic and OpenAI, checked on September 27, 2026.

Two features lower the bill in production:

• Prompt caching stores a repeated prompt prefix, such as your instructions and tool definitions. On Claude models, a cache read costs a fraction of the base input price.
• The Message Batches API processes requests asynchronously at a discount. Most batches finish within an hour, and a batch expires if it does not finish within 24 hours. It fits test-set runs and overnight processing.

Here is the arithmetic for one example. Suppose an agent reads 500 supplier invoices a month, and each invoice uses about 10,000 input tokens and 1,000 output tokens on Claude Sonnet 5. That comes to 5 million input tokens and 500,000 output tokens a month before caching. To price that volume, multiply it by Claude Sonnet 5's rates per million tokens on Claude API pricing. The token counts per invoice are an assumption for illustration, so measure your own on the pilot's test set.

Testing, tracing and the approval checkpoint

Testing tells you whether the agent is right. A test set is a versioned list of real cases, each with the input, the expected output and a scoring rule, and it runs on every prompt or model change. Start with 50 to 100 cases and grow the set as reviewers correct the agent.

Tracing tells you what the agent did. It logs every model call, tool call, cost and response time, so you can follow one request from start to finish. LangSmith and Langfuse both trace agents, and Langfuse is open source and self-hostable. Braintrust turns patterns in production traces into test cases.

The approval checkpoint needs four properties:

1. Right-sized scope. People review consequential decisions, such as posting an invoice or changing a price. Reviewing every output would erase the time the agent saves.
2. Confidence routing. The agent reports how sure it is. Confident outputs get a quick approval, and uncertain ones get a full review with the reasoning shown.
3. A full audit trail. The record shows who reviewed, when, what changed and why. It also helps you answer a request under section 12.1 of Quebec’s private sector privacy act when a decision was based exclusively on automated processing.
4. A feedback loop. Every reviewer correction goes into the test set as a new labelled case.

Privacy rules and where the data goes

• PIPEDA applies to private-sector organizations across Canada that collect, use or disclose personal information in the course of a commercial activity. Organizations under a substantially similar provincial law, such as Quebec's, are generally exempt for activity within that province.
• The Artificial Intelligence and Data Act was not enacted. Bill C-27, which would have enacted it, was still in committee in the House of Commons when that session of Parliament ended on January 6, 2025.
• Quebec's Act respecting the protection of personal information in the private sector, as amended in 2021 by the law known as Law 25, has sections that bear directly on an agent. Section 12.1 requires you to tell a person when a decision about them is based exclusively on automated processing, and to give the principal factors and parameters on request. Section 17 requires a privacy impact assessment before you communicate personal information outside Québec.
• Penalties under the Quebec act are large. Section 90.12 sets administrative penalties of up to $10,000,000 or 2 percent of worldwide turnover, and section 91 sets fines of up to $25,000,000 or 4 percent, whichever is greater in each case.

Where the model runs. Check where a provider processes data before an agent handles personal information.

• Anthropic’s own API offers two inference locations, global and US, and no Canadian option.
• From the Canada (Central) region, Amazon Bedrock serves Claude through cross-region inference with US or Global routing. The Claude Sonnet 5 and Claude Opus 5.5 model cards list in-region processing in Canada as not supported. AWS says data at rest stays in the Canada (Central) region, while the transient inference processing happens in the destination region.
• SAP and Cohere announced on February 10, 2026 a sovereign AI offering that begins in Canada, with Cohere's North platform inside SAP's sovereign cloud, and state that data remains in Canada.

For a job that involves personal information about people in Quebec, plan the section 17 privacy impact assessment around where the model processes the data. Many shop-floor jobs, such as reading a drawing for a quote, involve little or no personal information, which makes this step smaller.

Build these into the project plan:

1. Privacy impact assessment. Complete it before an agent touches personal information in production. The Commission d’accès à l’information oversees the Quebec act.
2. Explainable decisions. Design the audit trail and reviewer screen so you can give the principal factors behind a decision on request.
3. Breach procedure. Add the steps for a privacy breach to the production runbook.

Funding programs that share the cost

Several Canadian programs share the cost of AI work with smaller companies. Each one has its own rules, so check eligibility before you fix the scope.

Mitacs Accelerate

Mitacs Accelerate places a graduate student with your company for four or six months. You contribute $7,500 per internship and Mitacs matches it, for a $15,000 research award, and the intern receives at least $10,000. For a postdoctoral fellow, each side contributes $10,000. Mitacs reviews a proposal in about six to eight weeks, so apply well before the start date. It fits a research question inside the build, such as a new testing method or a model tuned to your parts data.

NRC IRAP

NRC IRAP gives financial support to incorporated, for-profit companies operating in Canada with up to 500 full-time equivalent employees that develop innovative, technology-driven products or services. It does not fund day-to-day operating costs or purely commercial activities. NRC IRAP also runs AI Assist, which the federal government launched in October 2024 with $100 million for Canadian SMEs that build or incorporate generative AI and deep learning solutions. To start, a senior executive calls NRC IRAP's contact centre, which can connect you with an industrial technology advisor.

Scale AI

Scale AI reimburses up to 40 percent of eligible expenses on approved projects that adopt or commercialize AI in value chains, as a grant that you do not repay. A project needs more than one participant, and at least one must be a small or medium enterprise with fewer than 500 full-time employees. Projects typically last 12 to 18 months.

NGen AI4M Challenge

NGen's 2025 AI4M Challenge reimbursed 40 percent of eligible costs on projects between $1.5 million and $8 million. Each project paired a manufacturing company with an AI company, and at least one of them had to be a small or medium enterprise. Its final application deadline was July 17, 2025, so check NGen's funding page for the next call.

Combining programs

Each program sets its own rules on combining it with other government funding. Ask about stacking limits when you apply, and design the project to fit the program from the start.

What stops agent projects before they ship

Four patterns stall agent projects, and each one has a fix you can apply before the build starts.

1. No defined job. The project starts as a wish to use AI and never names a task. Write down the job and what it costs you each week before any build begins.
2. No test set. Nobody can tell whether a prompt change helped, so decisions rest on impressions and get reversed. Label 50 real cases in the second week and version the set from then on.
3. No checkpoint on consequential actions. The agent posts a wrong invoice or sends a wrong message, and your team stops trusting it. Design the approval step before anyone writes code that changes a record.
4. Integration creep. The build starts with one job in one system and grows because each integration seems to need the next one. Limit production to one to three systems and hold the rest for later quarters.

Sources

• Anthropic: Models overview.
• Anthropic: Pricing.
• Anthropic: Batch processing.
• Anthropic: Data residency.
• Anthropic: Model deprecations.
• OpenAI: API pricing.
• AWS: Accelerate generative AI innovation in Canada with Amazon Bedrock cross-Region inference.
• AWS: Claude Sonnet 5 model card for Amazon Bedrock.
• AWS: Claude Opus 5.5 model card for Amazon Bedrock.
• SAP Canada: SAP and Cohere expand partnership to launch sovereign AI solutions.
• Intuit Developer: Set up OAuth 2.0.
• Intuit: Intuit App Partner Program guide (PDF).
• SAP Learning: Service Layer API architecture and capabilities.
• HubSpot Developers: APIs by tier.
• Microsoft Learn: Microsoft Graph overview.
• Microsoft Learn: Overview of metered APIs and services in Microsoft Graph.
• Microsoft: Power Automate pricing.
• LangChain: LangSmith.
• Langfuse.
• Braintrust.
• Office of the Privacy Commissioner of Canada: PIPEDA requirements in brief.
• Parliament of Canada: Bill C-27 (44th Parliament, 1st session).
• LégisQuébec: Act respecting the protection of personal information in the private sector.
• Commission d’accès à l’information du Québec.
• Mitacs: Accelerate.
• NRC: Financial support for technology innovation (IRAP).
• Government of Canada: Programs to help SMEs adopt and adapt AI solutions (AI Assist).
• Scale AI: The AI projects we invest in and how to obtain funding.
• NGen: AI4M Challenge.