3 AI Apps in Production in 3 Days

Results

• Speed to production: 3 tools in production within a three-day on-site sprint, and a 4th two days later
• Claude adoption: 12 consistent Claude users before the first sprint, more than 130 daily users today
• Order inbox time: about 75% fewer hours spent working the orders inbox
• Spend review: every non-production purchase order and vendor bill gets a recommendation from the Smart Spending CFO tool before anyone approves it, and roughly 43% of reviewed spend is flagged for a closer look

"We'd met plenty of firms that could talk about AI. Pulse sat down with our team and had tools running on real transactions by the end of the first week, and they've kept building with us since. We love working with Pulse, and the work has already produced a number of phenomenal successes, with more in the pipeline."

The problem

The client is a high-growth, mid-market consumer packaged goods manufacturer. By early 2026 it had Claude and Copilot turned on for employees and had set aside weekly AI time for its team. Use took off once Claude was connected to NetSuite and SharePoint, but a small group of power users carried most of it, and everyone else was learning on their own.

Leadership was building too. A production scheduling model built in Claude kept running past what a single Claude session can hold.

Meanwhile, orders that didn't come in through EDI arrived by email in a shared orders inbox, and keeping up with it took a big share of the order team's day. On the finance side, every purchase order and vendor bill needed someone to judge whether it fit the department's budget and habits. Doing that by hand across many account groups was slow and inconsistent.

Leadership was clear about what it didn't want: a firm that starts projects and doesn't finish them, or consultants who advise and never implement. The company booked a three-day on-site sprint.

How the sprints worked

Pulse calls this forward-deployed engineering: the engineers sit with the people who do the job, get the details from them directly, and build inside the same three days.

A four-person Pulse team ran the first sprint on site. They split into groups, interviewed the people who did each job, agreed with them on what finished would look like, and whiteboarded each workflow. Then they built.

On the second day, the company opened four hours of office hours to its whole team. Anyone could come down to learn about AI or bring a question from a project they were working on. It turned into a small hackathon: people showed what they'd been building and left with answers they could take straight back to their own projects.

Three tools were in production by the end of the third day. The order intake agent, in its first form as a Claude skill, read new orders from the shared orders inbox. A daily report, built on a validation pipeline the company had started, checked the production schedule against NetSuite inventory and work orders. A shift production report summarized plant-floor data for supervisors. The fourth, a quality report, went live that Friday, two days after the team left.

A second sprint centered on the Smart Spending CFO tool. The team built its first version in those three days, and it was live within about a week. Alongside it, the team worked through several smaller projects. Pulse has continued working with the company in an ongoing engagement since.

The order intake agent

The order intake agent went live in the first sprint as a Claude skill that read the shared orders inbox on a schedule. Pulse then added a dashboard where the order team reviews what it finds, and once the skill showed the approach worked, rebuilt the order intake agent as a full application.

Here is one order. A customer emails a purchase order to the shared orders inbox. Every ten minutes the order intake agent checks the inbox through Microsoft Graph, reads the new order and checks it. Orders with no flags are approved automatically, and the agent writes them into NetSuite through its API access. Flagged orders wait in the dashboard until someone on the order team reviews and approves them. If the customer needs a reply, the agent drafts one from a knowledge base built on the company's past order emails, and the reviewer sends it from the same screen.

That time went back to the team: orders reach operations sooner, and the team has more time to help customers send cleaner orders.

The Smart Spending CFO tool

The Smart Spending CFO tool reviews every open purchase order and vendor bill for non-production spend (everything outside the bill of materials) and labels each one approved, needs amendment, or not approved. It also reads the company's Ramp transactions to show spending by department over time.

The rules come from context cards written by the company's department heads. Each card covers a group of accounts in the chart of accounts and records who owns that spending and how that department decides what a good purchase looks like. Pulse built the agent harness that applies them: sub-agents work through each transaction using the card for its account group, and the harness returns one label with its reasoning. Scheduled jobs load NetSuite and Ramp data into the Smart Spending CFO tool's database, and a file fetcher pulls the documents attached in NetSuite, such as the vendor's contract, so the Smart Spending CFO tool can check a bill against what was agreed.

It catches the kinds of issues a busy approver can miss: a vendor whose insurance has lapsed, a purchase with no signed contract behind it, or a subscription the company no longer uses. The label appears inside NetSuite's existing approval request, so the approver sees the reasoning in the email they already get. The approver still makes the call.

Since launch, the Smart Spending CFO tool has flagged roughly 43% of the spend it reviewed. A flag means someone should take a second look, not that the spend was bad.

Governance and controls

• Only orders with no flags are approved automatically. Anything the order intake agent flags waits for someone on the order team to approve it in the dashboard before it reaches NetSuite.
• The Smart Spending CFO tool recommends and approvers decide, inside the company's existing NetSuite approval workflow.
• Department heads own the context cards that set the Smart Spending CFO tool's rules.
• Pulse tested the Smart Spending CFO tool with a small pilot group before releasing it to the whole organization, so the context cards could be tuned on real transactions first.
• Staff correct the Smart Spending CFO tool in plain language: they explain in its chat what it got wrong, and it updates its memory.
• The order intake agent works through API access to NetSuite and mailbox permissions in the company's own Microsoft 365 account.

Why forward-deployed engineering fit

The company brought Pulse the same problems most mid-market companies run into when their AI plans outgrow their team.

• They needed results fast. Leadership was weighing the best approach to maximize value creation and speed.
• They didn't want another pilot. Leadership had already decided AI was worth it and didn't need a proof of concept. They wanted tools in use, from a team that would finish what it started.
• Their builders were working alone. The people already building had hit the limits of what they could do by themselves, and the company wanted AI to help everyone without each person learning on their own.
• The work and business knowledge lived inside their systems. Orders, inventory and spending ran through NetSuite and Microsoft 365, so the tools had to be built inside those systems and the company's own access controls.

Pulse's forward-deployed engineers answered each one. They worked in the company's building, next to the people who run the order desk and approve spending, and had the first tools in production that same week. The team learned the systems as they were built, from the day-two office hours to the department heads who wrote the Smart Spending CFO tool's context cards. The engineers who built the tools are still working with the team, so nothing gets handed off to someone who wasn't there when it was written.

That's what companies get from forward-deployed engineering that hiring and agencies don't give them: senior engineers who ship in the first week, work inside the systems the company already runs, and stay while the team learns to own what was built.

All figures are Pulse's own measurements and estimates.