A dealing desk copilot's total cost combines user or model licences with permissioned data access, workflow implementation, answer evaluation and ongoing support. Budget for the evidence and controls behind each answer, then compare a narrow read-only assistant with broader action-enabled systems as separate investments.
Choose the question before choosing the model
A copilot that assembles a shift handover from approved incident notes has a smaller scope than one investigating execution quality across trading, bridge and liquidity-provider systems. Define the initial questions and the authorised users. Decide which sources the assistant may consult, how fresh they must be and what evidence the response must include. An unrestricted assistant is difficult to price because every new question can introduce another data contract and evaluation requirement.
Start with a read-only workflow that has an identifiable desk owner. Examples include collecting the facts behind an exposure change, finding the current escalation procedure or drafting a handover for human approval. Record what it must refuse: changing limits, placing trades, disclosing another entity's accounts or providing an answer from stale evidence. If actions are added later, budget authorisation, confirmation and recovery separately.
A worked three-year cost model
All figures below are hypothetical GBP planning assumptions, excluding VAT and inflation. They are not supplier prices or a KJ Capital quote. Assume ten users, two read-only data integrations, an approved document library and business-hours support. Delivery effort uses an assumed £800 per person-day. Replace every input with an actual quotation or measured workload before approval.
| Category | Assumption | Year 1 | Years 2–3 | Three-year total |
|---|---|---|---|---|
| User licences | 10 users × hypothetical £30/month | £3,600 | £7,200 | £10,800 |
| Implementation | 25 days × £800: workflow, interface and evidence display | £20,000 | £0 | £20,000 |
| Integration | 18 days × £800: two sources, identities and permissions | £14,400 | £0 | £14,400 |
| Evaluation and security acceptance | 15 days × £800: answer sets, leakage and failure tests | £12,000 | £0 | £12,000 |
| Model usage | Hypothetical measured workload allowance of £200/month | £2,400 | £4,800 | £7,200 |
| Hosting and retrieval | Assumed £300/month | £3,600 | £7,200 | £10,800 |
| Maintenance and evaluation review | 2 days/month × £800 | £19,200 | £38,400 | £57,600 |
| Total before contingency | Sum; user licence and model fees assumed separately chargeable | £75,200 | £57,600 | £132,800 |
Avoid double-counting where a licence includes model consumption. This example excludes new market-data rights, transcription, round-the-clock support and the desk's time spent reviewing answers.
Model consumption needs a workload calculation
Estimate monthly usage from active users, questions per session, sessions per month and the typical evidence supplied to the model. Include retrieved documents, tool results and generated output, then apply the chosen provider's current charging units. Retries, long investigations and evaluation runs also consume resources. The table's £200 monthly allowance is an explicit assumption rather than a prediction; replace it with a controlled pilot measurement.
A lower-priced model can reduce one line while increasing review effort if answers become less reliable. Compare models on a fixed set of genuine desk questions with known sources and expected behaviour. Consider latency and data-processing terms alongside accuracy. Model switching should preserve identity filtering, source provenance and the evaluation suite. The ability to change a model does not remove the need to retest its behaviour.
Permissioned integration is more than retrieval
A dealer's ability to open a document does not automatically mean the assistant may combine it with every account and trade record. Map roles, legal entities and permitted purposes into each retrieval and tool call. Authentication must reach the underlying access decision rather than merely protect the chat screen. Budget tests in which users deliberately request evidence outside their authority, including through a misleading document or an apparently helpful instruction embedded in source text.
Execution and exposure facts should come from approved structured services. The model should request a calculation, display its timestamp and explain the result with links; it should not rebuild authoritative risk measures from prose. Existing calculation services may reduce delivery effort, while inconsistent identifiers or undocumented permissions can increase it. Ask an integrator to demonstrate a real permitted question and a refused question against each source before accepting the connector.
Compare suppliers on the same acceptance scope
- Answer evidence: a response links to source records and approved procedures, states the time of the data and distinguishes a fact from a suggested interpretation.
- Abstention: missing evidence, a stale service or conflicting sources produce a qualified response and a clear escalation path rather than an invented conclusion.
- Access control: permissions are applied before retrieval and tool execution, with test evidence across the broker's roles and legal entities.
- Operational ownership: incident triage, prompt and model changes, document retirement, usage budgets, retention and exit procedures have named owners.
Buy an assistant, build a workflow, or combine them?
A licensed workplace assistant can be appropriate for finding approved procedures or drafting non-sensitive internal material where the product's permissions and data terms fit. Confirm the exact edition and included controls; a per-user price alone does not establish suitability for dealing data. Test whether citations identify the correct version of a procedure and whether a removed user's access is revoked throughout the retrieval path.
A tailored application can justify its added cost when a repeatable investigation needs structured trading evidence, deterministic tools and a broker-specific approval boundary. It also creates maintenance obligations. A combined design can retain an existing interface or model provider while owning the integration and evaluation layer. Compare the cost of maintaining that boundary with adapting a packaged product, and include data export and supplier transition in both options.
Measure value in accepted work
Count completed investigations or handovers that meet the desk's quality standard, not the number of chat messages. Measure baseline preparation time and review time, then repeat with the copilot. A hypothetical handover that falls from 40 minutes to 25 minutes saves 15 minutes before allowing for corrections. Multiply only by the actual frequency and loaded staff cost. Capacity released is not automatically a cash reduction or a trading-profit improvement.
Track correction rates, unsupported claims, failed retrievals, permission incidents and dealer adoption during the pilot. Agree a decision gate: proceed, narrow the workflow or stop. Preserve funds for evaluation after launch because sources, platform behaviour and model versions change. A modest first workflow with a clear owner provides a more defensible investment case than pricing a future assistant around unmeasured efficiency promises.
Related questions
Can a general chat subscription replace the build?
It may cover generic drafting or approved document search. A desk workflow that needs live trading evidence, entity permissions and auditable calculations requires those capabilities to be integrated and tested explicitly.
Does a read-only copilot still need security testing?
Yes. It can disclose protected data or produce misleading explanations without changing a record. Test access filtering, source instructions, retention and evidence freshness.
How should we budget additional tools?
Price each tool's data contract, permissions, failure path and acceptance tests. Tools that trigger operational changes require a separate authorisation and recovery design.
Book the KJ Capital AI Diagnostic to select a desk workflow, establish its evidence requirements and compare a costed build with available products. The readiness assessment helps identify the data and ownership gaps to resolve first.