Pubs Advisory Service. A production clause-extraction system for commercial pub leases
A pub tenant signs fifty pages of legalese to get the keys. The system read the lease and pulled out the thirteen clauses that decide whether the deal works.
Two-stage extraction Batches of 10,000 Refuses what the lease does not say
Clause library · 13 entries
Every answer comes back with the lease’s own wording underneath it.
Where Pubs Advisory Service were.
Pubs Advisory Service advise tenants and operators on the commercial leases they are about to sign, which means someone had to read every one of them properly before a single piece of advice could be given. The documents are long, frequently scanned badly, and written to be precise rather than readable.
OpenKit built and ran a bespoke document-analysis system for them: a clause library specific to pub leases, a two-stage extraction and answering pipeline behind it, and a batch path for when the work arrived in volume. It ran as an internal production tool rather than as a one-off analysis, with the maintenance and capacity reporting that implies.
- 01
Thirteen lease clauses configured with their own extraction notes and search questions, not a generic field list.
- 02
A prompt architecture that refuses questions the documents do not answer rather than inventing something.
- 03
The same engine ran on a different clause set by swapping the configuration.
What the service had to get right.
Pub leases carry the terms that decide whether a tenant can make a living: how the rent gets reviewed, what the property can be used for, who pays for the repairs, and whether there is a way out at the end. Those terms sit anywhere in fifty pages, often on a page that scanned crooked. Reading fast is how a clause gets missed, and reading slowly is what made detailed pre-entry advice expensive enough to price certain conversations out.
- 01
The clauses that matter to a pub tenant, specifically, rather than whatever a general contract tool thinks is important.
- 02
An excerpt attached to every answer, so an advisor confirms it against the lease instead of trusting it.
- 03
A plain refusal when the lease does not say, because a confident wrong answer is the expensive failure here.
A clause library built by the advisors.
The library encodes the advisors’ own judgement about what matters in a pub lease. Each clause carries its own extraction notes, its own natural-language search question and a description of what a correct answer looks like, which is a domain expert’s knowledge written down in a form the system can execute.
- Names of the parties
- Commencement date
- Length of the agreement
- Physical address
- Permitted use
- Amount of basic rent
- Deposit amount
- Cost of other contractual services
- Rent review terms, annual
- Rent review terms, open market
- Repair or maintenance fund amount
- Redecoration terms
- Option to renew
Inside the extraction pipeline.
- 01
Split the lease before reading it
The document is cut into overlapping extracts so a clause that straddles a page break is still whole in at least one of them. Textract handles the OCR on the scanned originals, which is where most of the damage happens in this kind of work.
- 02
Pull the clauses as question, excerpt, answer
Each of the thirteen clauses has its own extraction notes and its own natural-language search question. The output is structured: the question asked, the excerpt it was answered from, and the answer itself, so nothing arrives without the text it rests on.
- 03
Answer only from what was pulled
The second stage answers free-text questions over the extracted set, constrained to the supplied documents and instructed to say plainly when the lease does not answer the question.
- 04
Batches of up to ten thousand documents
Work comes in through an email-triggered path into a batch scheduler, and a post-processor turns the model’s output into clean list markup without touching the words.
How an answer comes back.
The answer and the wording it came from
Each clause comes back as the question that was asked, the answer in plain terms, and the lease’s own wording underneath it, so an advisor reads the answer and checks the excerpt in the same glance. A rent figure arrives with the review mechanism quoted beneath it.
The system in numbers.
clauses in the library, each with its own extraction notes and its own search question, written for UK pub commercial leases.
documents the scheduler takes in a single file, with individual items handled rather than the batch treated as one job.
of that capacity in use, measured across every component of the running system.
stages in the prompt architecture: pull the clauses out first, then answer strictly from what was pulled.
How the service was run.
OpenKit operated the system as well as building it, measuring capacity on every component and holding a documented review cadence.
- 01
Headroom recorded and countersigned on every component, with the finding stated plainly that no upgrades were needed.
- 02
A six-monthly review cadence and a documented escalation path.
- 03
Running costs measured from the infrastructure, on capacity sized to the volume of work.
The build
- Custom OCR + LLM pipeline
- Citation-anchored retrieval
- Email integration for advisors
- Google Sheets export
OpenKit certifications
- ISO 27001
- ISO 9001, UKAS-accredited
- Cyber Essentials
Controls on this project
- UK GDPR
- UK data residency
Their deep understanding of our business needs, coupled with their expertise in GPT and Cloud (AWS) software-services, enabled them to navigate complexities and deliver a bespoke AI solution tailored to our operations. The team demonstrated a high level of professionalism and adaptability, ensuring a smooth project delivery.
More of the work.
Find your first workflow.
We start with a conversation, audit where AI actually pays back, and build the first automation into how your team already works. We reply within one working day.
Start the conversation AI Audit