IP Location.net

Business, Software, Information Technology

An API, Not a Bot. How a Print Shop Let AI Read the Business Without Letting It Touch Anything

I run a commercial print shop with my brother. We ship 15,000 to 20,000 pocket folders a day. About a year ago, we started asking a simple question. We had Claude and ChatGPT on our desks for emails and spreadsheets. Why could they not tell us what was due today?

The answer was that our shop data was locked inside the management software, and the only way to get it out was to open the screens and read them. AI had gotten so easy that anyone could schedule a task or write a script, but none of the systems we demoed had an API or connector that worked well for our needs. So we built our own, and when we did, we made two decisions about AI access that shaped how we use it in our daily operations.

Decision One: Read Everything, Change Nothing

Our assistants can access information across the shop, including jobs, proofs, shipping, and outstanding invoices. They connect via a read-only connector that uses the Model Context Protocol (MCP), enabling compatible AI assistants to retrieve information from external systems. I can ask what is on hold, what is waiting on proof approval, what department a job is in and when it is due, or how accurate last month's estimates were. I can also request a morning briefing, and the answer comes from live operational data. Authorized administrators can retrieve invoice reports in a similar way.

What the assistants cannot do through this connector is change a job, a price, or an invoice. The connector does not have a write path. This restriction is built into the integration rather than relying on an instruction telling the AI not to make changes. For us, that distinction was important because it allowed employees to retrieve information without introducing unnecessary risks to production records.

Decision Two: If It Acts, It Goes Through the API

There is a difference between an assistant that reads and an agent that performs an action. We wanted to explore both. However, we did not want agents clicking around our screens like browser bots, because an incorrect action could affect production schedules, shipping, or customer orders.

Instead, we developed a REST API with scoped access keys. One key can be permitted to create quote drafts without modifying other records. Another can be authorized to generate shipping labels. This approach allows us to define which actions an automated process can perform and limit access to the functions it needs.

Over time, we introduced several automated workflows into our daily operations:

  • Web store importer: Orders from our online storefront come in as quotes tagged "Online storefront." Duplicate orders are identified, and paid orders can move into production according to predefined rules.
  • Shipping agent: The system reads the UPS and FedEx queues, generates shipping labels, associates them with the appropriate jobs, and updates tracking information.
  • Proof automation: A prepress script uploads proofs and sends them to customers. Customers can approve a proof or request changes through a link.
  • Estimating agent: An agent reads an email or specification sheet and prepares a quote. Pricing is calculated using the shop's existing press sheet rather than generated independently by the AI. Whether the quote is sent automatically or requires approval depends on the rules we establish.
  • Production monitoring: An automated process monitors jobs as they move through departments and can help assign work as capacity becomes available, subject to predefined operational rules.
  • Morning briefing: A single request provides an overview of upcoming deadlines, jobs on hold, and orders waiting for approval.

These workflows were introduced to reduce repetitive administrative tasks and improve access to information. They still require monitoring, especially when an automated action could affect customers, production schedules, or material usage.

Why the Separation Matters

One reason we started exploring AI was to reduce the time spent checking job statuses, preparing reports, and moving information between systems. However, commercial printing involves physical production. Once a job goes to press, mistakes can result in wasted paper, ink, time, and materials.

That is why we distinguish between retrieving information and performing actions. Read-only tasks can generally be automated with fewer operational risks, provided that access to sensitive information is properly controlled. Actions involving money, materials, or customer orders require additional safeguards.

For example, an assistant can retrieve a list of jobs waiting for approval without affecting production. Creating a shipping label or releasing a job to production is different because those actions can have financial or operational consequences. We use permission controls, predefined rules, and human approval where appropriate to manage those risks.

The level of automation will vary between businesses. Some shops may be comfortable automating routine shipping tasks, while others may prefer employees to review every action. The important consideration is maintaining control over what automated systems are permitted to do.

How Print Shop Software Can Support Read-Only AI Access

Some print shop management systems support integrations that allow AI assistants to retrieve operational data without modifying records. In our case, we use InfusePrint, the management system developed for our shop, as an example of this approach. Its read-only MCP connector allows compatible AI assistants to retrieve information such as job status, press hours, and outstanding invoices without directly editing those records.

When changes are required, they must be made through a separate REST API with permissions scoped to specific tasks. This separation allows us to use AI for reporting and operational visibility while keeping changes to jobs, pricing, and invoices subject to additional controls.

Other print management platforms, including PrintSmith Vision, ePS Pace, and printIQ, have different integration capabilities. API availability, supported functions, and compatibility with AI tools can vary between systems. Depending on the software and the information a shop needs to access, a custom integration may be necessary.

Our experience highlighted the importance of evaluating more than whether a management platform offers an API. We also needed to understand what information could be retrieved, which actions could be performed, how permissions were managed, and whether access could be restricted without affecting normal operations.

What It Took to Build and Implement the System

We did not begin this project with the intention of developing commercial software. We were trying to solve operational problems in our own print shop. After producing more than 13 million folders over a little more than two years, we had a clearer understanding of where our existing management processes were creating unnecessary work.

Some of the systems we had used previously involved additional setup or storage fees, limited storefront integration, and workflows that did not fully match our production process. We wanted a system that could follow a job from the initial quote through production, shipping, and the final paid invoice while making that information accessible to our assistants.

My brother developed the first version in approximately six months. We began using it in daily operations and continued making adjustments based on how employees interacted with the system, which information they needed most frequently, and where automation could reduce repetitive tasks.

The system was eventually made available to other print shops as InfusePrint. Developing our own print MIS software gave us a better understanding of how operational data, API permissions, and AI integrations can work together. Although our implementation reflects the needs of our particular business, the experience taught us several lessons about integrating AI into operational software.

One of the most important was that AI does not necessarily need permission to change records to provide value. Providing an assistant with reliable access to up-to-date information can already reduce manual reporting and help employees make decisions. When automated actions are needed, separating them from read-only access provides a clearer way to manage permissions and operational risks.

For our shop, the goal was never to have AI independently run every part of the business. It was to make existing information easier to access and automate selected tasks without giving up control over production decisions.

Conclusion

One of the most important lessons we learned was that AI does not necessarily need permission to change records to provide value. Providing an assistant with reliable access to up-to-date information can reduce manual reporting and help employees make decisions. When automated actions are needed, separating them from read-only access provides a clearer way to manage permissions and operational risks.

For our shop, the goal was never to have AI independently run every part of the business. It was to make existing information easier to access and automate selected tasks without giving up control over production decisions.

Featured Image generated by Google Gemini.

Share this Post

Comments

Comments are available to signed-in users and are moderated to keep the discussion useful and respectful. Spam, automated submissions, and low-value promotional comments are removed. Outbound links may be approved when they are relevant and genuinely helpful to readers, but they are displayed as plain text rather than clickable hyperlinks.

No comments have been published yet.

Please sign in to submit a comment.