A working demo of AI added to an existing onboarding and KYC review process, with the business case, the model choice and how it would run in your own environment.
A fictional UK FX and payments firm onboards business customers by email and runs periodic KYC reviews by hand. It wants to know whether AI can take work off its team without changing that process or its systems. The demo assumes little or no redesign of that process, so the AI fits around what the team already does. That isn’t necessarily best practice, but it is a realistic starting point for many firms, because redesigning the operating model, managing the change and retraining staff is a big barrier to adopting AI.
The demo answers four questions, in the same order as the tabs, and ends with takeaways and next steps. Click any of them to go straight there.
Every company, person and figure here is invented. The attachments are text versions of the documents so you can edit them; a real build reads the PDFs and photos. Volumes and costs are placeholders until they're replaced with the firm's own numbers.
To start, a person approves every email. Once the model has proved itself on a routine type of case, such as asking for an in-date passport, the firm can let it send those itself. Periodic reviews follow the same flow.
The model reads the email and attachments, the firm's rules check what it found, and you get one of three outcomes here: approve, email the customer with a draft reply ready to send, or pass to compliance.
To see each step behind the outcome, open AI workings after a run.
The same case, step by step: what the model did, what ordinary code did, and how the outcome on Team view was reached. The model only reads and makes judgement calls. Code does the hard checks and decides the next step.
These stand in for the firm's own KYC policy, which a real build would write down with compliance. In every rule the model first reads the details out of the documents in call 1. Exact checks then run in code. Judgement calls are answered by the model in the same call.
New applicationsDeciding the next step is code. Anything the customer can fix gets an email. A case goes to compliance when it needs a risk decision rather than a document: a possible politically exposed person, a high-risk country, a dissolved company, or payments far above what was declared. For unusual payments the analyst shouldn't ask the customer directly, because if the activity turns out to be suspicious that could tip them off. Compliance decides whether to ask for more, apply enhanced checks, report or decline.
The model is called twice. Each call sends fixed instructions, written once with compliance, plus the case. The model replies in a fixed format that the code can read.
Step 1 instructions: read and judgeOpen this box to load them.Step 4 instructions: draft The case, as sent
Three ways to run onboarding and reviews over the next three years: carry on as today, add AI to the current process, or move onto a KYB onboarding platform: a customer portal with automatic register, ID and sanctions checks, priced per case. Every number is an assumption you can change at the bottom of the page.
Until a route goes live the team keeps working as today, so those months carry today's staff cost on top of the route's project cost. A later go-live means fewer months of saving inside the three years.
| Today | AI on the current process | KYB platform | |
|---|---|---|---|
| What changes for the team | Nothing. | Same inbox and steps. The model checks and drafts; a person approves. | Customers upload documents through the platform’s portal instead of by email, and the team works each case in the platform. |
| Time to first result | – | Weeks | A few months |
| What you get beyond the saving | – | Choice of model, and the cost of every case in view. | Automatic register, ID and sanctions checks, and ongoing monitoring that flags changes at the company. |
A KYB platform changes how customers send their documents, which is the step this firm has chosen not to change for now. A full client lifecycle management (CLM) platform is a bigger step again, for firms that need one client record and one policy engine across many countries.
| Assumption | Value | Why this number |
|---|---|---|
| Volumes | ||
| Active business customers | An illustrative UK FX firm with around £100m of revenue. Published figures put revenue at roughly £13,000 to £20,000 per active business customer (Ebury and Moneycorp), so about 5,000 to 8,000 customers. Replace with the real number. | |
| New applications a month | Scaled to the same £100m firm from published onboarding and customer growth figures, plus applications that are abandoned or declined. A range of 250 to 700 is plausible. Replace with the real monthly figure. | |
| Share of customers reviewed each year (%) | High-risk customers yearly, medium every two to three years, low every three to five. 40% is a blend; the firm's risk mix sets the real figure. Trigger reviews add to it. | |
| Reviews | – | Active customers × share reviewed each year, then divided by 12 for the monthly figure used in the rest of the model. |
| Share of cases that go to compliance (%) | Cases the model passes to a person for a judgement, such as a company being struck off or payments well above what was declared. These are worked fully by hand, at today's time, on every route. | |
| People | ||
| Fully loaded cost per person a year (£) | Salary plus employer costs (national insurance, pension, benefits, desk) for an onboarding analyst in the UK. Replace with the firm's own figure. | |
| Hours worked a day | A standard UK office day. | |
| Days a week | Full time. | |
| Working weeks a year | 52 weeks less holidays, bank holidays, sickness and training. | |
| Cost per working hour | – | Annual cost ÷ (hours a day × days a week × working weeks). |
| Hours per person a month | – | Hours a day × days a week × working weeks ÷ 12. |
| Today, by hand | ||
| Email rounds per application | Customers rarely send everything right first time. Three rounds is the first email plus two chasers. | |
| Staff minutes per round, application | Reading the email and attachments, checking them against Companies House, writing the reply and updating the file. | |
| Email rounds per review | One request for updated information and one follow-up. | |
| Staff minutes per round, review | Comparing the file with the register and recent payments, then writing the request. | |
| Minutes per case today | – | Rounds × minutes per round, for applications and reviews. |
| AI on the current process | ||
| Email rounds per application | The model asks for everything missing in one email, so fewer chasers. The customer still has to reply. | |
| Staff minutes per round, application | A person reads the model's check and draft, edits if needed and presses send. The minutes are per email, so they add up across rounds. | |
| Email rounds per review | Reviews where nothing has changed close with no email at all, which brings the average down. | |
| Staff minutes per round, review | Reviewing a shorter draft against the file note. | |
| Minutes per case with AI | – | Rounds × minutes per round. |
| Build and integration, one-off (£) | About 30 days of work: connecting the mailbox, document reading, the rules, testing on the firm's past cases, the audit log and training the team. | |
| Running and support a year (£) | About £3,000 a month for hosting, monitoring accuracy, re-testing when model versions change, and support. | |
| Model | ||
| Model cost per case (£) | Set from the model above once it has results on the Models tab. You can type over it. | |
| Months to go live | One process first, run alongside the team, then switched on. | |
| KYB platform | ||
| Portal rounds per application | The portal asks for every document up front and pulls the register itself, so fewer chasers. | |
| Staff minutes per round, application | The platform runs the register, ID and sanctions checks. A person still reviews the case, makes the judgement calls and writes any request. | |
| Rounds per review | Ongoing monitoring flags changes at the company, so many reviews need no request. | |
| Staff minutes per round, review | Reviewing the platform’s alerts and the file. | |
| Minutes per case on the platform | – | Rounds × minutes per round. |
| Set-up and integration, one-off (£) | Configuring the portal and rules, and connecting the platform to the core systems. Placeholder. | |
| Internal project team, one-off (£) | Operations, compliance and IT time, plus moving customers onto the portal. | |
| Platform fee a year (£) | Case management, users and the portal. Placeholder; replace with a quote. | |
| Checks per case (£) | Company check, ID checks on owners and directors, and sanctions and PEP screening. Placeholder based on published per-check prices. | |
| Months to go live | Mostly configuration, with less integration and migration than a full platform. | |
Several KYB providers publish per-check prices, from about $2 to $3 for a basic company check, with ID checks on owners and screening charged on top; mid-market platforms add platform fees or monthly minimums. The figures above are placeholders to replace with a quote. Examples: The KYB pricing; Didit, business verification pricing.
The same cases through ten models. Missing a request or referral is the costly error: a case that needed a person or a document was approved or closed. A cheaper model that never misses those can do the job at a fraction of the cost, which changes the business case and what a vendor should charge. Each run is steps 1 to 3 from AI workings: the model reads and makes the judgement calls, and the same code checks and rules decide. The draft step is left out because it doesn't change the decision.
Missed a request or referral is the costly error: the case needed a document from the customer or a compliance decision, and the model approved it or closed the review, so a problem would have gone through. Other wrong decisions are the less costly kind, for example emailing a customer whose application was fine, which wastes time and annoys the customer but lets nothing through. Failed answers are runs where no usable answer came back, usually because the provider was busy. Press Re-run failed to try those again.
What the demo shows, using the figures and choices entered on the other tabs, and how a firm would take it forward.
Why the model bill stays this small, and when it doesn’t, is in the article The Agent Cost Curve.
This demo uses an invented firm and invented cases. Happy to talk through any of these steps: scalepointpartners.com.
Shown here for Microsoft 365 and Azure, which many payments firms use; the same pattern works on Google Workspace or AWS. In production this runs inside your own tenant. The team stays in Outlook, the data stays in Azure, and nothing is sent without a person pressing send. This demo page uses invented data only.
Every input, model answer, rule result and human decision is logged in your own cloud account for audit. Outside your account: the Companies House API, and the model provider where the table below says so.
Your choice here also decides which models are lit up on the tab. The steps for a single case are shown on . A real build adds three steps around them: watching the mailbox, gathering the documents and register record, and saving the draft in the shared mailbox with an audit log. All three are ordinary code, and their cost is in the build and running costs on the tab.
| Area | How it is handled |
|---|---|
| Mailbox access | |
| People stay in control | The model can only create drafts and file notes. It cannot send email, approve an account or change a customer record. |
| Instructions hidden in emails | Customer emails and attachments are treated as data. Because the model can only draft, a malicious instruction has nothing to act on. |
| Attachments | |
| Audit trail | |
| Personal data | A data protection impact assessment and an update to processing records before go-live. |
| Third parties | The model provider and any build partner go on the third-party register, including DORA ICT rules where the EU entity is in scope. |
| Model governance | Tested on your own past cases before go-live, accuracy tracked once live, re-tested when the model version changes. |
| ScalePoint access | Built in the firm's own cloud account. No client data is held on ScalePoint systems. |