How We Built One Custom AI Layer That Connected Every System and Put the Whole Team on AI

For a fast-growing B2B distributor operating across the US and the UK, the hardest questions were never complicated ones. Why hasn't this order shipped? Has the customer paid the last invoice? Is there an open complaint on the account? Each answer existed somewhere in the business. Finding all three meant opening HubSpot, NetSuite and Zendesk, searching a shared drive and scrolling back through Slack.

Today, anyone in the company asks that question once and gets a complete, sourced answer in under 30 seconds. Behind that change is a single custom AI layer we built on top of the systems the business already ran, without replacing any of them. Within three months of launch, 94% of employees were using it every working day.

Good systems, disconnected work

The company had made sensible technology choices as it grew. Sales ran on HubSpot, orders and invoicing on NetSuite, customer service on Zendesk, and contracts, pricing agreements and procedures in Google Drive. Each platform was well configured and well liked by the teams that used it.

The friction lived between them. Over three years, the operations team had stitched the stack together with 47 Zapier and Make automations, each built to solve a specific problem at a specific moment. As order volumes grew, that web became fragile. A renamed field could break a sync without anyone noticing. Platform costs rose quarter after quarter. Finance still spent around two days each month reconciling CRM deals against ERP invoices by hand, because the automated sync could not be fully trusted.

An earlier experiment with AI had not helped. The assistants built into individual applications could only see their own data, so the CRM assistant knew nothing about invoices and the helpdesk assistant knew nothing about orders. For a business whose most important questions crossed every system, a set of partial answers was not progress. It is a common position to be in. In a recent Zapier survey of enterprise leaders, 70% said they had not moved beyond basic integration for their AI tools.

Connecting instead of replacing

The original brief was to assess a new platform that could consolidate several tools. Two weeks of mapping every system, data flow and automation pointed somewhere else. The core systems were sound, the teams were productive in them, and replacing any one would have meant a long, risky migration. What the company lacked was not another system but a reliable way to work across the ones it had.

We proposed a single AI layer, owned entirely by the client, sitting above the existing stack, an approach we now offer as part of our AI solutions for growing businesses. Before any development began, we agreed three commitments with the leadership team: every system would stay the source of truth for its own data, the AI would never reveal information a person could not already access, and nothing that changed a record or reached a customer would happen without human approval. Those commitments shaped every design decision that followed.

One layer, built once and built properly

The foundation replaced dozens of point-to-point automations with a single integration layer, built by our AI automation team. We connected HubSpot, NetSuite, Zendesk, Google Drive, Gmail and Slack through their native APIs and exposed each capability as a governed tool using the Model Context Protocol (MCP). Every connector is monitored and tested, so a change on the vendor side is caught before it reaches users rather than discovered weeks later in a mismatched report.

Above that sits a retrieval-augmented generation (RAG) knowledge layer over the company's own knowledge, including contracts, pricing agreements, product documentation and operating procedures. It re-indexes automatically when documents change and excludes superseded versions, so every answer is grounded in current, approved sources and shows where it came from.

Cost was considered from the start. Routine lookups and summaries run on fast, low-cost models, while harder reasoning, such as explaining a reconciliation exception, is routed to more capable frontier models. That balance kept running costs steady as usage grew from a pilot group to the whole company.

Designed around how each team actually works

Rather than deploy one general-purpose assistant, we built custom AI agents around the workflows that consumed the most time. Account managers now receive a short brief before every customer call, drawing together deal history, open orders, outstanding invoices and recent support tickets. Operations staff ask for the status of any order and get an answer assembled from NetSuite, Zendesk and the relevant Slack threads, with supplier follow-ups drafted for review.

In finance, reconciliation is now run by an agent that matches CRM and ERP records and surfaces only the genuine exceptions, each with an explanation. Support receives triage suggestions and draft replies informed by live order data. Leadership asks questions in plain language and gets answers that span the whole business, without waiting for a report to be built.

Every one of these interactions respects the user's existing permissions in each source system. Actions that change data or contact customers are routed for approval, and every request and response is written to an audit log.

Earning trust before scaling

The programme ran for 12 weeks and was delivered by a dedicated team of AI engineers, the same model clients use when they hire developers from us to extend their own teams. The first six were spent on integrations, the knowledge layer and the permission model. We then launched with 30 people from sales and finance, the two teams with the heaviest cross-system workload, and spent three weeks refining answers and approval flows around how they actually used the tool.

The pilot did more for adoption than any training plan. Once account managers saw a complete customer picture appear in seconds, other teams began asking for access. The company-wide release that followed needed only short, role-specific walkthroughs.

As each agent went live, the automations it replaced were retired and their logic rebuilt in tested, version-controlled code the client now owns. By the end of the programme, all 47 had been switched off.

What changed for the business

The figures tell only part of the story. Sales, operations and finance now work from the same view of every customer, because every answer draws on the same systems. Finance trusts the month-end numbers because exceptions are explained rather than hidden. And the company's AI capability no longer depends on the roadmap or pricing decisions of any single software vendor.

What comes next

With integration and governance in place, extending the layer has become a matter of weeks. The client is now connecting its warehouse management system and building a forecasting agent for purchasing, both on the same architecture and approval model.

Distribution is one of several sectors where we see this pattern. Businesses across the industries we serve, from logistics and retail to healthcare and financial services, run on similar stacks and face the same gaps between them. For teams weighing a similar decision, our guide on when to move from no-code tools to custom AI sets out the framework we used on this project, and you can explore more of our work in our case studies.

See what one AI layer could do for your business

If your teams are switching between disconnected systems, maintaining fragile automations, or getting partial answers from app-specific copilots, we can show you what a connected AI layer would look like on your own stack. In a 30-minute session, we map your systems, identify the workflows with the biggest time savings, and outline a realistic rollout plan.

Book a demo or estimate your project cost in a few minutes. To learn more about how we work with clients, read why companies partner with iSkylar Technologies.