Inside a London discretionary manager quietly funding private LLMs.

The firm and the decision it made quietly

Ashworth Calder is the kind of firm that does not appear in league tables unless a journalist specifically requests it. Founded in 1993 as a breakaway from what was then County NatWest Investment Management, it built its client base through referrals from solicitors and accountants in the Home Counties, then expanded northward through acquisitions of regional discretionary managers in Leeds and Edinburgh during the consolidation wave that followed RDR in 2012. Its £18.4 billion AUM sits below the Rathbones and Quilter tier by total size, but its client concentration — 78 per cent in mandates above £500,000 — gives it a profit margin that its larger competitors have not historically matched. The firm runs 23 model portfolios across five risk profiles and a tax-efficient overlay. Those portfolios sit at the centre of what chief executive Helena Mercer describes, in her annual letter to advisers, as "the repeatable excellence that justifies discretionary fees."

Mercer, who joined Ashworth Calder in 2017 from a senior position at Brewin Dolphin and became chief executive in 2021, brought one conviction to the role that her predecessor had not shared: that the firm's model portfolio service was simultaneously its greatest competitive asset and its greatest operational vulnerability. The MPS generates approximately £4.3 million in annual revenue at current fee levels. It also requires a degree of manual coordination — between the investment committee, the portfolio operations team, and the twelve wealth managers who adapt the model allocations for individual client circumstances — that Mercer described in an internal strategy document reviewed by this reporter as "structurally incompatible with the margin demands of the next decade."

The decision to build a private LLM was not, in Mercer's account, a reaction to what frontier model providers were offering. She had evaluated enterprise contracts with both OpenAI and Anthropic in the first quarter of 2023. The capability case was not in doubt. The problem was elsewhere: client data residency, FCA operational resilience requirements, and a fundamental unwillingness to put Ashworth Calder's proprietary rebalancing logic — the accumulated investment committee judgment of thirty-one years — into a prompt that traversed a third-party API endpoint. "We are not a technology company," she told the executive committee in April 2024. "But our process is our product. That process does not leave this building."

The FCA framework that shaped every architectural choice

The FCA's operational resilience framework — set out in PS21/3 and in force since March 2022 — requires FCA-regulated firms to identify their important business services, set impact tolerances for disruption, and demonstrate that they can remain within those tolerances through severe but plausible scenarios. For a discretionary manager, the important business services are the portfolio rebalancing and client reporting functions that sit at the core of the MPS. Any technology that becomes part of those services is subject to the same operational resilience requirements as the services themselves. When Ashworth Calder's chief risk officer, Daniel Forsyth, began the firm's regulatory dialogue with the FCA's Supervisory Technology and Authorisations division in May 2024, this was the first question he addressed. The firm's position — that a self-hosted system on exclusively leased UK colocation infrastructure constituted an internal analytical tool rather than an outsourced technology service — was accepted by the FCA in principle, with conditions. Final discretion on every portfolio action had to remain with a named human portfolio manager. The system's outputs had to be logged with the same audit trail as any other investment research. And the infrastructure had to be hosted in a jurisdiction subject to UK data protection law — a condition that, after the Brexit-era adequacy decisions, pointed firmly toward dedicated UK colocation rather than any cross-border cloud arrangement.

The GDPR successor framework — the UK GDPR, administered by the Information Commissioner's Office — added a layer that the FCA dialogue did not. Ashworth Calder's client data includes decades of portfolio histories, risk appetite assessments, and suitability records for approximately 9,400 individuals. The firm's data protection officer, Sarah Kowalczyk, spent eight weeks in the summer of 2024 mapping which elements of that data set could be included in a training corpus under the ICO's existing guidance on AI and data protection. The conclusion was narrower than the initial estimate: client personal data, including anything that could identify an individual client's investment preferences or portfolio composition, was excluded from the training corpus entirely. What remained was the firm's own investment committee output — thirty-one years of model portfolio changes, rebalancing decisions, asset allocation rationale documents, and the internal research memoranda that the investment committee had produced on each change. That corpus, stripped of any client-identifying information, ran to approximately 180 million tokens. It is, in Forsyth's phrase, "the intellectual history of how we make money."

The pension tailwind that made the build urgent

The UK pension reform programme — the staged abolition of the lifetime allowance, the extension of auto-enrolment to smaller employers, and the government's pension pot consolidation measures — has sent a wave of newly advised clients toward discretionary managers in the £500,000 to £2 million mandate bracket. Ashworth Calder's new business pipeline shows this clearly: in the twelve months to September 2024, 34 per cent of new mandates came from clients who had previously held their pension assets in an employer-scheme default fund and were moving them into bespoke discretionary management for the first time. These clients arrive with more specific tax structuring needs, less familiarity with the adviser-manager interface, and a higher frequency of query volume than the firm's historical book. The wealth management team's average client contact hours rose by 22 per cent in the same period. The model portfolio rebalancing cycle, which had previously run on a quarterly schedule with ad-hoc adjustments, is under pressure to accelerate as client circumstances diversify. Jonathan Grieves, Ashworth Calder's head of wealth management, describes the current situation in terms that the firm's 2019 leadership team would not have recognised: "We are receiving the pension reform dividend. The question is whether our operating infrastructure is scaled to collect it."

The private LLM project is, at its operational core, an answer to Grieves's question. The system is not being built to replace portfolio managers. It is being built to handle the analytical workload that currently sits between the investment committee's model portfolio decisions and the individual wealth manager's client-specific implementation — a gap that, as mandate complexity grows, consumes roughly a third of every wealth manager's working day in manual cross-referencing of model weights, tax positions, and client suitability records. If the system absorbs that analytical layer, Grieves estimates that the wealth management team can handle 30 per cent more mandates without adding headcount. At Ashworth Calder's current average mandate fee of £8,400 annually, 30 per cent more mandates is approximately £23 million in additional annual revenue at full maturity. The investment committee approved the project's economics on this basis. The £6.8 million build cost is, on this framing, a 3.4-month payback against the revenue case. Mercer presented it to the executive committee with a single line: "This is the cheapest operational hire we will ever make."

The build: Mistral, a Manchester data centre, and one hire

The technology lead Ashworth Calder hired is not drawn from the firm's existing operations function. Thomas Aldridge joined the firm in June 2024 from his previous role as a principal machine learning engineer at a London-based quantitative hedge fund — a firm whose name he is contractually prevented from disclosing but which manages assets in the tens of billions. Aldridge's mandate at Ashworth Calder was precise: design, build, and operate a private LLM that runs exclusively on UK-sovereign infrastructure, integrates with the firm's existing portfolio management system without requiring a rebuild of that system, and produces outputs that a wealth manager with no AI background can interpret and act on within two minutes of receiving them. His title is Head of Intelligent Systems. His team, as of the time of writing, consists of himself and two engineers seconded from the firm's existing technology operations group.

The base model Aldridge chose is Mistral 7B Instruct, fine-tuned on Ashworth Calder's proprietary corpus across fourteen weeks of iterative tuning cycles. The choice of Mistral over DeepSeek or Llama reflects a specific licensing concern: Mistral's Apache 2.0 licence imposes no restrictions on commercial deployment or on the number of users accessing a fine-tuned derivative, a condition that Aldridge required given that the system will eventually be accessed by the full wealth management team. The infrastructure runs on a dedicated cluster of 24 NVIDIA H100 SXM nodes colocated at a Tier IV data centre in Salford — a facility whose primary tenants are UK financial services firms with data residency requirements, and which holds ISO 27001 and Cyber Essentials Plus certifications that satisfy the FCA's operational resilience documentation requirements. The compute lease costs £1.4 million annually. Aldridge's loaded annual compensation is £480,000. The total ongoing operating cost, including the two seconded engineers and Aldridge's own salary, runs to approximately £2.3 million per year — 12.5 basis points of the firm's annual MPS revenue, against the 30 per cent additional revenue the system is projected to enable.

The corpus that underpins the fine-tuning contains one element that Aldridge describes as the system's most defensible competitive asset: thirty-one years of Ashworth Calder's investment committee minutes, including not only the final decisions but the intermediate working documents that the committee produced during its deliberations. Most discretionary managers archive their final model portfolio decisions in structured formats accessible to compliance systems; the working documents that explain the reasoning behind those decisions typically survive only as unstructured PDFs in shared drives that no one has incentivised anyone to index. Ashworth Calder's previous head of investment operations spent three years, ending in 2021, systematically converting those working documents into a structured archive. The project had originally been undertaken for compliance retrieval purposes. Aldridge recognised, when he reviewed the firm's existing data estate in his first month, that it was inadvertently the most valuable asset in the build. "Most firms," he told Forsyth in their first working session, "are starting from deal memos. We are starting from thirty years of recorded argument."

What to watch

Ashworth Calder's deployment is in its early production phase. The signals below will determine whether this becomes a template for UK discretionary managers or remains an outlier at the top of the sub-Rathbones tier.

• FCA guidance on AI in discretionary management. The FCA's current operational resilience framework was written before private LLMs were a live question for wealth managers. The regulator's AI Discussion Paper DP23/3, published in October 2023, signalled that sector-specific AI guidance for investment management was under development. If that guidance arrives before mid-2025 and addresses self-hosted AI systems explicitly — particularly around the question of whether a fine-tuned internal model constitutes a material outsourcing arrangement — it will function as a compliance template that either accelerates or constrains the deployment curve for the fifteen to twenty discretionary managers that advisors in the space estimate are currently in evaluation. Ashworth Calder's informal dialogue with the FCA's supervisory division has given the firm a reasonable expectation of which way that guidance will land. Other firms are waiting for the paper.
• Thomas Aldridge's hiring signal. Aldridge has posted no additional roles since his two internal secondments in August 2024. The composition of a private LLM engineering team at a firm of Ashworth Calder's size — one principal engineer plus two generalists — reflects a deliberate constraint on headcount cost during the validation phase. If Aldridge posts an external role for a second ML engineer with financial document parsing experience, the firm has concluded that the system's performance justifies scaling the engineering function. That signal will appear on LinkedIn before it appears anywhere else. It has not appeared yet.
• Grieves's mandate growth figures in the Q1 2025 internal review. The business case for the LLM project rests on a specific claim: that the system will free enough analyst capacity to support 30 per cent additional mandates without headcount growth. The firm's internal review cycle runs quarterly. The Q1 2025 review will be the first to show the system's operational contribution against the mandate growth baseline established in 2024. Mercer has committed to revisiting the revenue case at that review. If the figures are positive, the executive committee has indicated it will approve an expansion of the compute budget to add capacity for the client reporting automation module that Aldridge has scoped but not yet funded.
• The pension consolidation pipeline's acceleration. The government's pension pot consolidation measures — which are in the implementation phase and require pension schemes below a certain asset threshold to transfer assets to larger vehicles by staged deadlines — will continue pushing newly freed pension capital toward discretionary managers through 2025 and 2026. If that pipeline accelerates beyond current Treasury projections, the operational pressure that made the LLM case compelling will intensify. Ashworth Calder is better positioned to absorb that pressure than it was twelve months ago. Its competitors at the same AUM tier, most of whom have not made an equivalent infrastructure investment, are not.
• The ICO's stance on AI training data in financial services. Kowalczyk's exclusion of client personal data from the training corpus is a conservative interpretation of UK GDPR that eliminates most regulatory risk but also constrains the system's usefulness on client-specific queries. The ICO published a series of AI and data protection guidance documents in 2023 and early 2024. If the ICO issues updated guidance that clarifies the conditions under which pseudonymised client data can be used in proprietary model training within a regulated financial services context, Aldridge has indicated he would revisit the corpus design. A more permissive ICO interpretation would make the system significantly more valuable on the client suitability and reporting tasks that currently require the most manual wealth manager intervention.

Frequently asked

Why would an FCA-regulated discretionary manager build a private LLM rather than use an enterprise contract with OpenAI or Anthropic?

The capability argument runs toward the public frontier models — they perform better on most tasks a discretionary manager wants to run. The decision against them rests on three issues specific to regulated wealth management. First, FCA operational resilience requirements under PS21/3 create documentation and control obligations for any technology that becomes part of an important business service; a third-party API relationship complicates that compliance picture in ways that a self-hosted system does not. Second, UK GDPR and ICO guidance create data residency and audit obligations that are easier to satisfy on dedicated UK sovereign infrastructure than across a cross-border API. Third, the proprietary investment process — the rebalancing logic and decision history that constitutes the firm's competitive differentiation — sits in the prompts sent to a third-party system. For firms whose process is their product, that exposure is the deciding factor.

How does the FCA operational resilience framework treat a private LLM built and operated by the regulated firm itself?

The FCA has not published specific rules for private LLMs, and its AI Discussion Paper DP23/3 did not resolve the question definitively. The informal regulatory dialogue that has taken place — between the FCA's Supervisory Technology and Authorisations division and at least one discretionary manager known to this reporter — has proceeded on the basis that a self-hosted system on UK-sovereign infrastructure, with final portfolio decisions retained by a named human portfolio manager and with full audit logging of system outputs, is more likely to be classified as an internal analytical tool than as a material outsourcing arrangement under the FCA's SYSC 8 outsourcing rules. That classification matters because material outsourcing triggers additional notification and due diligence obligations. The internal-tool classification does not eliminate all FCA obligations — the system's outputs still carry MiFID II suitability documentation requirements — but it significantly reduces the regulatory complexity of the deployment.

What is the economic case for a discretionary manager at £18 billion AUM?

At Ashworth Calder's scale, the case rests on capacity creation rather than cost reduction. The pension reform tailwind has driven a 22 per cent increase in client contact hours for the wealth management team in twelve months; without intervention, that trend compresses the firm's capacity to onboard new mandates. The LLM project is designed to absorb the analytical workload that currently sits between investment committee model decisions and individual wealth manager client implementation — a task that consumes roughly a third of each wealth manager's working day. If the system handles that analytical layer, the firm estimates a 30 per cent increase in mandate capacity without headcount growth. Against an average annual mandate fee of £8,400 and 30 per cent additional capacity, the revenue case justifies the £6.8 million build cost within four months at full utilisation. Below £5 billion AUM, the per-basis-point cost of this approach becomes harder to defend without a strong strategic or regulatory argument alongside the capacity case.

Why does the quality of the training corpus matter more than the choice of base model?

The base model — in Ashworth Calder's case, Mistral 7B Instruct — determines the system's general linguistic and reasoning capability. The corpus determines whether that capability is applied to the firm's specific investment context. A frontier-model capability on a thin or generic corpus produces confident outputs that are not grounded in the firm's decision history. A smaller base model fine-tuned on thirty years of structured, firm-specific investment committee reasoning produces outputs that reflect how that firm actually makes decisions. For a discretionary manager whose competitive differentiation is the consistency and quality of its investment process, the corpus is the value. Ashworth Calder's thirty-one-year archive of working documents — not just final decisions but the intermediate reasoning — is an asset that a competitor could not replicate in less than a decade of equivalent documentation discipline, regardless of which base model it chose.

Is this pattern visible at other UK discretionary managers in the same AUM bracket?

Not yet at scale. Among discretionary managers in the £5 billion to £30 billion AUM band — the tier that sits below Rathbones, Quilter, and Evelyn Partners by total size but above the regional boutiques that cannot sustain the infrastructure cost — advisors and technology consultants working in the space estimate that three to five firms are in active evaluation as of early 2025. None have publicly disclosed a build decision. The signals that indicate a firm has moved from evaluation to build are indirect: specialist ML engineering hires with financial document parsing experience, dedicated GPU capacity reservations at UK Tier IV data centres, and a quieting in the frequency with which the firm's technology leadership engages with frontier model enterprise sales teams. Those signals are beginning to appear, unevenly, across the sector.

The desk view

Ashworth Calder's system is six months into a production environment that Aldridge describes as limited but deliberate. The wealth management team uses it on portfolio rebalancing analysis for 22 per cent of mandates reviewed in the most recent quarter — below the 35 per cent target Aldridge set for the end of 2024, but above the 12 per cent he had projected for the same period when the system entered production in September. The gap between target and actuality runs in the firm's favour on adoption pace, and against it on the analytical task types where wealth managers are not yet confident enough in the system's reasoning to rely on it. Rebalancing analysis for straightforward multi-asset mandates is well within the system's demonstrated capability. Suitability cross-referencing for the more complex pension transfer mandates — the clients arriving through the reform pipeline with layered tax positions and occupational scheme histories — remains a task where wealth managers are using the system as a first-draft tool rather than a final-step tool. Aldridge is running a structured evaluation of the suitability task type that he expects to conclude in Q1 2025. If it passes, the 35 per cent adoption target becomes achievable before mid-year.

The broader pattern that Ashworth Calder's build represents is visible to anyone who has been paying attention to the UK wealth management sector's infrastructure investment over the past eighteen months — not as a trend with momentum, but as a leading indicator from a firm that has made a bet early enough to be running real production data while its competitors are still in evaluation. The pension reform dividend is arriving. The question of whether the UK discretionary sector's operating infrastructure can collect it without breaking is not rhetorical. Ashworth Calder's answer is a Salford data centre running Mistral fine-tuned on thirty-one years of investment committee argument. Whether that answer proves correct will be visible, in the aggregate, in the sector's mandate growth figures by the end of 2025. The individual firm's numbers will not be disclosed. They will be inferred from what the team's capacity allows.