A growing number of accounting firms are facing the same choice. The partners are convinced that AI document analysis is now a real productivity tool. The question is whether to build it in-house - assembling the components of a retrieval-augmented generation system from cloud AI APIs, vector databases, and a custom interface - or to use the docs-only mode built into a practice management or specialist AI platform. The honest answer for most firms is the second, but the reasoning matters and the cases where building in-house makes sense are worth being clear about.
This guide compares the two paths in 2026 terms - the components involved, the realistic cost, the ongoing operational burden, the regulatory positioning, and the speed to value. It is written for partners and operations leads who are being told by one of their team that they could build it themselves over a weekend.
What a RAG system actually consists of
A working retrieval-augmented generation system has at least seven components, each of which needs to be selected, integrated, kept current, and operated. The components are:
- An ingestion pipeline that takes documents (PDFs, DOCX, XLSX, emails) and prepares them for the index, including OCR for scanned material
- A chunking strategy that breaks documents into passages the model can work with, balancing context size against retrieval precision
- An embedding model that converts text passages into vectors representing their meaning
- A vector database that stores the embeddings and supports fast similarity search
- A retrieval layer that takes a user query, embeds it, finds the most relevant passages, and assembles them into context
- A language model with an API contract that supports the firm's data handling requirements
- A user interface, including conversation history, citation display, and refusal handling
The hidden eighth and ninth components
In practice there are two more components that the weekend-project estimate routinely omits. The eighth is the security and compliance layer - authentication, authorisation, role-based access control, audit logging, data encryption at rest and in transit, secrets management, and the contractual paperwork with each underlying supplier. The ninth is the operations layer - monitoring, alerting, version pinning of the language model, evaluation and regression testing of changes, incident response, and the resourcing for someone to be on the hook when it breaks.
These two components are where in-house RAG projects most often founder. The first six or seven components can be assembled in a fortnight by a capable engineer. The compliance and operations layers consume the rest of the year and never finish.
Realistic cost of building in-house
A defensible internal estimate for a small to mid-sized firm building a production-grade in-house RAG system covering the practice's documents looks like this. Initial build: three to six months of engineer time, with the lower end achievable only by someone who has built this before. Ongoing operations: the equivalent of half a competent person, indefinitely, plus the underlying cloud and API costs which typically run into the low thousands of pounds per month at firm scale.
Annualised, a serious in-house build costs in the order of £80,000 to £150,000 in the first year and £40,000 to £80,000 in subsequent years. These are not crazy numbers for a firm above £5 million in fees; they are very large numbers for a firm below £2 million in fees.
Realistic cost of using a built-in docs mode
The comparable cost of using a built-in docs-only mode in a practice management or specialist AI platform is, for most platforms, a per-firm or per-seat subscription in the low thousands of pounds per year - and often less, particularly for smaller firms where per-firm pricing applies. There is no engineering cost, no operations resource to fund, and no compliance paperwork to maintain beyond the supplier's own commitments.
The cost differential is therefore in the order of 20 to 100 times. That alone does not settle the question - there are reasons to build in-house that are not about cost - but it does mean the in-house case has to clear a high bar on the non-cost dimensions.
Where in-house genuinely makes sense
There are scenarios where building in-house is the right answer, and they should not be dismissed. The clearest are:
- A very large firm (top 20 by fee in the UK) where the bespoke nature of the document corpus and the volume of use makes the per-seat economics of a third-party platform genuinely unfavourable
- A firm with a specialist practice (forensic, insolvency, niche tax) whose documents and workflows are not well served by general-purpose platforms
- A firm with an existing in-house engineering team that has built and operated production systems before and has spare capacity
- A firm with a data localisation requirement that no third-party supplier can meet within the firm's required time-to-market
- A firm with a strategic view that proprietary AI tooling is part of its competitive position
Where built-in docs mode is the right answer
For most firms - by number, almost all firms below the top 50 by fee in the UK and most firms in the UAE - the built-in docs mode is the right answer. The reasons are:
- The supplier has already done the engineering, security, and compliance work the firm would otherwise be doing
- The supplier amortises the cost of keeping current with the latest models across all customers
- The integration with the practice management workflow, the client portal, the working paper file, and the time tracking is already done
- The contractual basis for processing client data has already been negotiated and is reviewable on day one
- The total cost is small enough that the productivity gain pays for it many times over even on conservative assumptions
- The firm's scarce engineering and operations attention can be spent on the engagements, not on the infrastructure
The regulatory positioning
A built-in platform from a serious supplier comes with a pre-existing data processing agreement, an established security posture, a documented commitment on whether data is used for training, and a track record of meeting professional services requirements. The firm signs the agreement on day one and has a defensible regulatory position from that point.
An in-house build requires the firm itself to assemble all of that - the contracts with the underlying cloud AI vendors, the documentation of the security controls, the policy on data handling, the evidence of staff training, the incident response plan. This is doable but it is real work and it is not optional. Insurers and regulators will ask for the same things from the in-house firm as from the supplier, and the firm has to produce them itself.
The hybrid case
A small number of firms run a hybrid - a built-in platform for the day-to-day client work, and a narrowly scoped in-house tool for a specific high-value workflow where the bespoke nature of the work justifies the investment. This works when the in-house tool is genuinely narrow (one workflow, not a general firm AI platform), when the firm has the engineering capacity to maintain it, and when the partner board has a clear view of why the bespoke investment pays back.
It does not work when the in-house tool is "general firm AI" duplicating the built-in capability of the platform. That is the scenario where the firm ends up paying for both, using neither well, and stuck with operational debt.
The build-versus-buy decision in practice
The honest version of the decision in 2026 for most firms is: buy the built-in docs mode for general use, evaluate carefully whether any specific high-value workflow justifies a bespoke tool, and revisit the question annually as the supplier landscape and the firm's scale change. A partner board that takes this position can deploy AI in the next quarter rather than the next year, and can redirect the engineering and operations capacity it would otherwise have spent into the engagements that pay the bills.
How Accupe helps
Accupe's built-in docs-only AI mode gives firms the RAG capability without the build cost. The platform handles ingestion, embedding, retrieval, citation, refusal, and the working-paper integration; the three modes (Fast, Planning, Ultra-Detailed) cover the range of work from quick extraction to deep document interrogation; and the per-firm pricing from £20/month removes the cost-per-seat friction that drives some firms to consider building in-house in the first place. The compliance positioning - docs-only, source-cited, no training on client data - is documented and reviewable.