What Drives the Cost of AI Automation for a 200-Person Financial Services Firm

The most useful conversation we have with mid-market financial services COOs and CTOs is about the cost of an AI automation project. Most of the cost figures in the public domain are miscalibrated for a 200-person firm. They are either consumer figures (the monthly price of a chatbot seat) or large-enterprise figures (a bank's multi-year AI programme). Neither helps plan a real engagement at a mid-market FCA-regulated firm.

This piece walks through what goes into the cost of designing, building, running and governing AI automation in a typical 200-person UK financial services firm: a wealth manager, an advisory firm, a broker, a specialist insurer, or a similarly sized regulated business. It sets out what moves each layer of cost, so you can plan before anyone quotes.

The four cost layers

AI automation cost has four distinct layers, and the balance between them matters. Most cost surprises come from underestimating one of them.

1. Discovery and design. Before a model is selected, the workflow has to be understood as it actually runs. The Evolve Workflow Audit is a fixed fee, quoted on the first call, and what moves it is the number of people in scope and the complexity of the operating model. The output is what the rest of the cost is planned from, which is why it comes first.

2. Build and integration. The cost of building the automation, integrating it with your existing systems, and standing up the eval harness, observability, and rollback path. For a single well-bounded workflow at a 200-person firm, say, automating first-pass suitability report drafting from structured client data and adviser notes, this is a fixed price per phase quoted from the audit. What moves it is the integration scope. A workflow that lives entirely inside one system, with a clean API and well-organised data, is cheaper. A workflow that touches the CRM, the document management system, and the email archive, and where the data needs cleaning before it is useful, is more expensive.

3. Inference and infrastructure. The running costs of the model and the supporting infrastructure, billed to the firm's own AWS Bedrock or Azure OpenAI tenancy in UK regions so every line item is visible. Model inference has fallen sharply in price since 2023 and is rarely the number that surprises anyone. Where firms get caught is on log retention and observability infrastructure, which can cost as much again, but is the kind of cost the regulator will expect you to be paying.

4. Governance and ongoing operation. The costs that are most often left out of business cases. Quarterly re-evaluation of the eval harness against new edge cases. Monitoring dashboards reviewed by the second line. Periodic refresh of prompts and curated test sets as the underlying business changes. For a single deployed workflow this is a recurring cost that belongs in the business case from the start; for a portfolio of three or four workflows, the per-workflow cost falls because much of the eval and monitoring infrastructure is shared.

What that adds up to

For a typical 200-person UK financial services firm starting with one well-scoped AI automation workflow, the year-one cost is the sum of those four layers: discovery, build, the first year of running costs, and the first year of governance. We quote the first two as fixed prices after the audit, and size the second two so they can be checked against the cloud bill. The next workflow added to the same eval and monitoring infrastructure costs noticeably less than the first one, because the foundations carry over.

For comparison: the human-time equivalent of automating a first-pass suitability drafting workflow at a typical wealth manager is six advisers each saving several hours a week of report drafting time. Multiply that by what an adviser hour costs your firm and you have the number the business case is built on, and it is a number you can check because the audit recorded it.

Where firms blow the budget

Three patterns repeatedly. None are technical; all are scoping problems.

Wrong workflow. The most expensive failure mode is an automation pointed at a workflow that should not have been automated, or that should have been a different shape of automation. The Workflow Audit prevents this; firms that skip it because “we already know what we want to build” are the ones that find themselves rebuilding twelve months later.

Underestimated data work. The build cost ranges above assume the underlying data is in roughly the state the discovery suggests. Where data needs cleaning, deduplicating, or restructuring before the automation can use it, that work has to happen and it has to be budgeted. Most firms underestimate it on the first project and right-size it on the second.

Governance retrofit. Firms that build the automation first and bolt on governance later spend far more than they would have spent designing it in. The FCA, the ICO, and the firm's second line all expect the same artefacts; building those as part of the original engagement is much cheaper than reverse-engineering them under audit deadline pressure.

How to scope the first project to land well

The pattern that lands best for mid-market firms: pick one well-bounded workflow with clear quality criteria and recoverable errors. Run the Workflow Audit. Build, eval, and pilot. Take the learnings into the next workflow with the same governance scaffolding. The compounding effect after three or four workflows is substantial, because the per-workflow build cost falls as the surrounding infrastructure carries over and the firm's confidence in scoping increases.

The wrong move is the opposite: committing to a multi-workflow programme before any single workflow has been delivered into production. The cost of that approach scales linearly with scope and the risk of the first workflow being the wrong one is highest.

For a deeper look at the patterns we see most often in financial services, see the AI automation for financial services page. For the question of when an AI automation should instead be an agentic AI deployment, and what the cost differential looks like for multi-step agentic systems, the natural next read is Agentic AI explained: a UK operator's guide.