Case study
Adding an AI Copilot to a Legacy Vertical SaaS Product
How a click-heavy, 15-year-old field-service SaaS gets an in-product copilot: one intent sentence, a grounded draft, human approval, and a product that finally feels current.
No client engagement behind this piece: this is how we would transform this category of product, with benchmark-sourced targets.
Who this is for
A typical 15-year-old vertical SaaS product for field-service companies
Industry
B2B software (vertical SaaS), SMB customer base
Legacy stack
Engagement
Concept
Contents
Your product wins deals on depth: fifteen years of workflows, edge cases, and domain logic a startup cannot copy overnight. But in every demo, prospects now compare your screens to AI-native tools, and your own customers quietly paste your data into public chatbots to get answers your product should have given them. The renewal conversation has a new question in it: “what is your AI story?”
This is the transformation we would run for that product. It is presented as a concept, built on the same method we use in client engagements, so you can see exactly what the before, the after, and the path between them look like.
The busywork, quantified
Follow one work order through the legacy flow. A dispatcher takes a two-minute phone call, then spends the next ten minutes re-typing what they just heard: five screens, thirty-plus fields, customer looked up on one tab, asset history on another, the technician’s calendar in a third. Multiply by every order, every day, every customer of your product.
The product is not broken. It is faithful to how software was designed fifteen years ago: the person translates intent into fields, one click at a time. The database, the contracts, the SLA logic underneath are all still right. The waste lives entirely in the translation layer.
Before and after: the work-order flow
Drag the handle. The before is the recreated legacy module; the after is the copilot flow designed in its place.
Copilot drafted · review and confirm
- Customer / Site
- Hi-Rise · 400 King St W
- Asset / Fault
- Unit 3 · F-217 Door jam
- Priority / SLA
- High · Gold 4h
- Assigned
- M. Rossi · Tue AM
The sentence the dispatcher would have said on the phone becomes the record. The old form does not disappear; it becomes the review surface, where a person confirms what the copilot drafted instead of typing it from scratch.
What we built
The concept above is not a mockup exercise. It is the output of the same engagement steps we run on real products:
- We map the busiest workflow first: work-order creation, because it touches every customer every day and its time cost is easy to measure.
- We open a safe seam into the legacy system: characterization tests around the work-order module, then a thin API the copilot can call without touching the WebForms code.
- We build the copilot as a grounded layer: it drafts from the product’s own customers, assets, contracts, and technician calendars, never from thin air, and it cites what it used.
- We design the approval gate before the automation: every draft renders as visible, editable fields with a confirm button, an undo, and an audit trail.
- We wire evals from day one: accuracy is scored on real historical work orders, and every human correction becomes a test case.
How it works
The legacy database stays the source of truth, which is what makes this safe for a product with paying customers. The copilot reads through the same permissions the signed-in user already has, drafts against live data, and writes only through the reviewed confirm step. If the copilot is ever wrong, the cost is one correction in a review screen, not a bad record in your customers’ books.
For the product team, the copilot is also a wedge: the seam opened for it, with its characterization tests, is the first paid step of a larger modernization. Each following slice gets the same treatment, so the codebase gets healthier while the product gets smarter.
Why this pays back
For a vertical SaaS company the return is double. Your customers get hours back every week, which shows up in retention and in the demo. And your product gets an answer to the AI question that is not a bolted-on chatbot: a copilot that lives inside the workflow your competitors would need years of domain depth to replicate.
The outcomes
Screens to log one work order
5 screens, 30+ fields One sentence, one review
5 to 11
Design targetShare of routine admin work generative AI can absorb
60-70%2
Industry benchmarkTime to the first shippable slice
4-6 wk3
Design target1 Design target for the copilot intent flow, measured against the recreated legacy flow shown in the before/after above.
2 McKinsey, The economic potential of generative AI (June 2023): generative AI can automate activities that absorb 60 to 70 percent of employees' time.
3 Standard first-slice scope: one workflow, one metric, evals and an approval gate included.
Frequently asked questions
Can a copilot really be added to a 15-year-old codebase without a rewrite?
Yes. The copilot lands as a thin layer over the product's existing APIs and database, so the legacy system stays the source of truth. Where the old code is too brittle to expose safely, we write characterization tests around that slice first, then open a clean seam for the copilot to call. No big-bang rewrite, no second system to keep in sync.
What happens when the copilot gets a field wrong?
Nothing posts without review. The draft appears with every field visible and editable, the person confirms or corrects, and each correction is logged. Those corrections feed the evals, so accuracy is measured on your real work orders, not on a demo script.
How is this different from bolting a chatbot onto the product?
A chatbot answers questions next to the busywork. This removes the busywork itself: the work-order flow is re-engineered around intent, so the sentence a dispatcher would say on the phone becomes the record, with the old form kept as the review surface rather than the data-entry surface.
What would trying this cost?
A scoped proof of value on the single busiest workflow, priced as a fixed first slice. It ships in weeks with its own metric, so the decision to go further is made on a measured result, not a promise.
Go deeper on the method
AI Copilots vs. Autonomous Agents: Which Your Workflow Actually Needs
One drafts beside your team, the other runs the process end to end. A simple way to tell which a workflow needs, and how to sequence from copilot to agent without losing trust.
Read articleModernize First, Then Add Intelligence: Upgrading Legacy Software Without a Rewrite
You cannot safely bolt AI onto a codebase nobody dares to touch. Modernize the foundation first, with tests that lock in current behavior, then make it intelligent.
Read articleHow to Turn Your Existing Software Into an AI-Powered Product (2026 Guide)
Going AI-powered rarely means a rebuild. A practical 2026 playbook for turning the software you already run into an intelligent product, in weeks, with people in control.
Read articleMore transformations
AI Claim Triage From Photos: First Notice of Loss to a Reviewed Assessment
How a property claims operation turns a folder of unsorted phone photos and a 62-page policy into a structured damage assessment: every line citing the photo it came from, the coverage clause quoted, uncertain lines flagged, and an adjuster approving before anything moves.
32.4 days
Average property claim, filing to finished repairs
AI Quote Desk for a Wholesale Distributor: RFQ Inbox to Priced Quote in Minutes
How a parts distributor replaces the morning RFQ pile and the three-system price hunt with a copilot that extracts every line item, prices it from the company's own item master and contracts, checks stock, and drafts the reply, with a rep approving every quote before it leaves.
60-70%
Share of routine admin work generative AI can absorb
Get the next transformation in your inbox.
When we publish something worth your time, you will be first to know. No spam, unsubscribe anytime.