Client Onboarding, from process map to payback

In the Workflow OS demo, the projected case for automating Client Onboarding at a 30-person firm takes staff time from 100 to 40 hours a month, with payback in 4.5 months. This walkthrough follows the process through Workflow OS: the map of eleven manual steps, the score, the redesign with four approval gates, and the business case, each number one you can check with a calculator.

The firm is Northstar Professional Services, the fictional 30-person company in the Workflow OS demo, with 5 departments, 11 systems and 17 mapped processes. Every ROI figure below is a projection until a pilot measures the real one.

Eleven manual steps

Onboarding starts when the client signs the engagement and ends when they get their instructions. In between, a person does eleven things. The map records who does each step, in which system and how long it takes.

Sixty new clients a month at 100 minutes each is 100 staff hours a month, about $6,000 of loaded labour at an estimated $60 an hour. The map also shows 10 changes of person or system per case, about five days of waiting per client and rework on 14% of cases. Those three facts matter as much as the hours: they are what the client feels.

A score of 88, Priority 1

With those inputs the engine gives Client Onboarding 88 out of 100: value 94, feasibility 94, strategic fit 100, minus a 7.1-point risk penalty for confidential data and the regulated KYC step. The full arithmetic is in How Workflow OS decides which process to automate first.

The engine prints its reasons in plain sentences. Two of the constraints shape everything that follows:

• "KYC / compliance review involves regulated judgement and should remain human-approved."
• "Handles confidential client data; every component needs scoped data access (privacy penalty 3 of 8)."

Software, two agents and one human decision

The recommended pattern is Hybrid AI-Assisted Automation. Steps that follow fixed rules go to plain workflow software. Steps that work on documents and free text go to AI. The compliance decision stays with a person.

A few rules in the design carry most of the risk control:

• Template-only messages. Both agents send messages built from approved templates. Any other message, and any reply the agent cannot classify, waits for the Client Services Manager.
• The CRA authorization stays a person’s action. The request is pre-filled from the client record. Staff submit it, and the client confirms it in their own CRA account.
• The KYC database stays read-only for automation. AI prepares the file and a summary; the Compliance Officer approves every file, and the approval is recorded by them.
• Scoped data. Each AI component reads only the current client’s file, and every action goes to the audit log.

Counted that way, the design has 4 approval gates. Staff time per client falls from 100 to 40 minutes: 16 for the KYC approval, 10 for record exceptions, staffing and the welcome, 7 for reviewing requests and the CRA authorization, and 7 for checks and follow-up exceptions. About 40% of new clients still need at least one follow-up for missing documents, and the design plans for it.

Five phases, twenty tasks

The plan is generated from the design. Each phase has a goal and tasks with an owner, effort, acceptance criteria and a risk level. Engineering tasks are owned by the delivery partner’s engineers; validation, pilot operation and training by the firm’s own people.

The prototype is tested on 30 sample cases, exceptions included. The pilot runs for four weeks with two or three people, under human approval, and its results are compared with the baseline on time, errors and turnaround. Ninety days after go-live the metrics are checked against the targets, and the firm decides whether to widen the scope or move to the next process.

The business case

The engine prints this arithmetic beside its results, so a finance lead can redo it by hand.

Client Onboarding, projected

Labour saved

= (100 − 40) min × 60 clients ÷ 60 × $60/h = $3,600 a month

Implementation

= 14 vendor days × $950 × 1.10 contingency = $14,630

Operating

= $100 platform + 6 integrations × $25 + 4 AI steps × $0.40 × 60 = $346, rounded to $350 a month

Payback

= $14,630 ÷ ($3,600 − $350) = 4.5 months

Year-1 ROI

= ($43,200 − $4,200 − $14,630) ÷ ($14,630 + $4,200) = 129.4%

Three choices in this arithmetic keep it conservative:

• Rework time already sits inside the 100 baseline minutes, so it is counted once. Error reduction, avoided external costs and freed capacity count as zero until they are measured.
• The day rate of $950 covers vendor engineering time. The firm’s own staff time is tracked as effort in the plan and left out of the cash cost.
• Operating cost grows with volume: every AI step is charged per case, so a busier month costs more to run.

Where the projection can break

Four inputs behind this case are still estimates: the $60 loaded hourly cost, the 14% rework rate, the 120 hours of waiting and the exception rating. The engine flags each one and asks for the evidence that would confirm it, starting with the hourly cost from finance. The Validate phase exists for exactly this: time ten recent cases, confirm the volume and replace the estimates before anything gets built.

The same engine is willing to say “no”. In the same demo company, five of the twelve generated designs show a negative Year-1 ROI:

Two of them show why. Employee Onboarding takes 200 minutes per hire, and the demo firm hires three people a month. Expense Processing is easy to automate and small: about 22 staff hours a month. Both are better left alone for now, and the numbers show it before anyone spends money.

Projects like this one sit between AI transformation and custom software: the agents need governance, and the workflow engine, connectors and audit log are ordinary, well-tested software. For how the pattern itself was chosen, read Plain software, an assistant or an agent. To run the same numbers on your own process, write to us about automation.