The verdict in three sentences
A bank with 1,400 procedures, policies and guidance notes loses hundreds of hours a day to document searches, and every wrong answer in the front office is a conduct risk. A private AI assistant, hosted in a controlled environment and restricted to the approved document base, costs between 80,000 and 200,000 GBP in 2026 for a 3 to 4 month project. It answers with the exact source, handles 25,000 queries a month and cuts search time by 4, provided it meets FCA and PRA expectations on operational resilience and governance.
The problem: 1,400 documents and busy front-line teams
In a London commercial bank, procedures cover onboarding and KYC, payments, retail and SME lending, anti-money laundering, complaints handling under Consumer Duty, and vulnerable customers. Documents are updated several times a year and stored on an intranet where keyword search returns dozens of versions.
| Indicator (2026 order of magnitude) | Standard intranet | Private AI assistant |
|---|---|---|
| Average time to find a procedure | 10 to 15 min | 2 to 4 min |
| Calls to back office for simple questions | 3,000 to 5,000 per month | 800 to 1,500 |
| Risk of applying an outdated version | High | Low (only the current version is indexed) |
| Onboarding time for a new adviser | 6 to 8 weeks | 3 to 5 weeks |
| Traceability of questions asked | None | Full, auditable log |
| Availability | Head office hours | 24/7 |
The productivity gain is real, but the decisive argument for the risk function lies elsewhere: the assistant always cites the source document, its version and date, which makes every answer verifiable.
Architecture, security and budget
The assistant relies on retrieval from the bank's document base (RAG). The language model answers only from retrieved passages and says 'I cannot find this information' rather than improvising. Hosting is either on dedicated infrastructure in a UK data centre or in a private cloud tenancy with a contract guaranteeing no data is used for model training.
| Item | 2026 market budget (GBP, ex VAT) | Timeline |
|---|---|---|
| Scoping, classifying and cleaning 1,400 documents | 10,000 to 25,000 | 3 to 4 weeks |
| Vector database and search engine | 15,000 to 35,000 | 3 weeks |
| Intranet interface, single sign-on | 15,000 to 40,000 | 3 to 4 weeks |
| Role-based access (branch, credit, compliance) | 10,000 to 25,000 | 2 weeks |
| Logging, dashboard, quality testing | 15,000 to 40,000 | 2 to 3 weeks |
| Security: penetration testing, model risk review | 15,000 to 35,000 | 2 weeks |
| Typical total project | 80,000 to 200,000 | 3 to 4 months |
| Run costs (hosting, API or GPU, support) | 3,000 to 10,000 per month | Depends on volume |
For reference, Kolonell prices this type of project in its complex application band, from 7,000,000 to 15,000,000+ FCFA (10,670 to 22,870+ EUR) depending on scope, excluding run costs. Application support then falls under monthly maintenance at 35,000, 75,000 or 120,000 FCFA depending on service level, plus hosting and model usage. London budgets are higher mainly because of regulatory testing, internal governance and UK day rates.
Compliance: FCA, PRA and UK GDPR
FCA and PRA expectations on operational resilience, outsourcing and third-party risk require documenting the tool, its suppliers, access rights and logs. The Senior Managers and Certification Regime means a named executive owns the outcome. UK GDPR applies as soon as customer data flows through the assistant. In practice, indexing procedures rather than customer files greatly simplifies the data protection impact assessment.
Mini case study
James, head of operations at a 900-employee bank in the City, counts 25,000 procedure searches a month. Illustrative scenario: project at 140,000 GBP, run costs at 6,000 GBP a month, i.e. 212,000 GBP in year one.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Time saved: 9 minutes per search x 25,000 = 3,750 hours a month. Valuing only half of that time at a 45 GBP loaded hourly cost gives about 84,000 GBP a month, roughly 1 million GBP a year. The project pays back in under 3 months, before counting fewer procedural breaches found by internal audit.
FAQ
How much does a private AI assistant cost for a bank in London in 2026?
Budget 80,000 to 200,000 GBP ex VAT for 1,000 to 2,000 documents, plus 3,000 to 10,000 GBP a month in run costs. Cost varies mainly with hosting choice and depth of security and model risk testing.
Does bank data leave the UK?
Not necessarily: the assistant can run on dedicated UK infrastructure with an open-weight model. A private cloud tenancy is cheaper upfront but must be approved under outsourcing rules.
What if the AI gives a wrong answer?
Every answer cites the source document and version, which the adviser can open in one click. Accuracy above 90% is measured on a set of 300 test questions before go-live.
How is the assistant kept up to date?
The policy owner publishes a new version in the tool, and the old one is removed from the index within minutes. No new development phase is needed.
How long does deployment take?
Three to four months, with a pilot in 2 or 3 teams from month two.
Let's scope your project. Tell us your document volume, hosting constraints and user count: we will price a scope, an indicative budget and a pilot then rollout schedule. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
