A 12-person management consultancy we audited in April was burning 22 director-hours on every UK public sector bid response. They submitted four tenders per quarter and won one — a 25% win rate, above the sector average of 17%. The problem was not what they wrote. It was how many bids they could afford to write. At 22 hours per response, submission capacity was the constraint. Had they doubled their submission rate at the same quality level, the wins would have followed.
We mapped where those 22 hours went: eight reading and annotating the Invitation to Tender, six writing method statements from scratch, four in internal senior review, four formatting and checking compliance for portal submission. The goal was not to eliminate all 22 hours — four of them, the senior review, are non-negotiable. The goal was to recover the rest through automation. After a three-week build sprint, the same consultancy produced a 40-page first draft in 14 hours. Here is what that pipeline looks like, where it holds, and where it breaks.
The UK public procurement cycle and exactly where 22 director-hours disappear per bid
Public sector procurement in the UK follows a structured cycle regardless of whether the authority uses the new Procurement Act 2023 framework or a legacy procedure still running under transitional provisions. From an SME supplier's perspective, the stages that consume director-hours are: opportunity identification, pre-qualification via Selection Questionnaire, ITT receipt and annotation, response drafting, and submission through a portal such as Jaggaer, Delta eSourcing, or the government's own Find a Tender service.
A standard ITT for a professional services contract in the £200k–£2m range runs to 30–50 pages of specification, evaluation criteria, and compliance requirements. Attached is a response template: typically 12–20 questions with individual word limits, quality weightings, and mandatory pass/fail criteria. Evaluators score method statements on evidence, specificity, and relevance — not creative prose. That structure is both what makes bid writing amenable to automation and what makes low-effort automation dangerous.
| Stage | Hours | Automation potential |
|---|---|---|
| Reading ITT, marking up criteria | 8 | Low — interpretation requires human context |
| Writing method statements | 6 | High — library retrieval plus model adaptation |
| Internal senior review | 4 | Low — human accountability required |
| Formatting, portal compliance, submission | 4 | High — rule-based pipeline |
The 10 automatable hours are recoverable without sacrificing accuracy. The 12 that remain need human attention and should not be compressed.
Building a compliance library: past answers, accreditations, and capability statements as structured data
The compliance library is the single piece of infrastructure that separates a firm getting faster at bidding from one that plateaus. It is a structured repository of everything the firm can truthfully say about itself: past question-and-answer pairs from submitted bids, accreditation certificates with expiry dates, insurance levels and policy numbers, staff CVs in a standard format, methodology statements, equality and diversity policies, financial account summaries, and references from prior public sector clients.
The mistake most firms make is storing this as a flat folder of Word documents. Retrieval is then manual and adaptation is slow. The right structure is a set of JSON records with explicit metadata tags:
{
"entry_id": "lib-047",
"type": "method_statement",
"topic": "project_management",
"contract_types": ["professional_services", "consultancy"],
"word_count": 380,
"last_updated": "2025-11-12",
"certifications_referenced": ["ISO 9001:2015"],
"text": "Our project management approach follows a structured four-stage ...",
"used_in_bids": ["bid-2025-08", "bid-2026-02"],
"win_record": [
{"bid": "bid-2025-08", "result": "won"},
{"bid": "bid-2026-02", "result": "lost"}
]
}
Embedding each entry and storing the vectors alongside the JSON — we use pgvector on Postgres for this — enables fast semantic retrieval when a new ITT question arrives. Building the first library from ten historical bids takes three to five days. After that, every bid adds new entries and retrieval accuracy improves. The retrieval architecture that underpins this is the same pattern we describe in our Document RAG case study.
Question mapping: matching ITT questions to the right library section automatically
With the library embedded, question mapping becomes a nearest-neighbour lookup. When an ITT arrives, the pipeline extracts each question box — question text, word limit, scoring criteria, pass/fail flags — into a structured list. Each question is embedded and compared against the library using cosine similarity.
# question_mapper.py
def map_questions(itts_questions: list[dict], library: VectorStore) -> list[dict]:
mapped = []
for q in itts_questions:
embedding = embed(q["text"])
results = library.search(embedding, top_k=3, threshold=0.78)
mapped.append({
"question_id": q["id"],
"word_limit": q["word_limit"],
"weight": q["weight"],
"library_matches": results,
"confidence": results[0]["score"] if results else 0.0,
"draft_source": "library" if results else "generate"
})
return mapped
Questions with a score above 0.78 are tagged draft_source: library — the model adapts an existing answer. Questions below that threshold are tagged draft_source: generate — the model must draft from scratch with explicit source constraints. In the April audit client, 63% of questions matched at 0.78 or above after the initial library build. For a deeper look at the retrieval trade-offs between vector and keyword approaches, see document RAG: vector vs keyword retrieval.
LLM-assisted response drafting: what the model writes reliably and where it confidently invents
The split between library-matched and generated content determines where you spend review effort. For matched questions, the model's job is adaptation: take a 380-word method statement written for a previous contract, reduce it to 250 words to fit this ITT's limit, and replace specific references to the previous contract with the current one's terminology. That task the model does consistently well. Output quality is high because the factual base is fixed and injected.
For unmatched questions, the risk profile changes entirely. The model drafts from the general context: your firm's name, the ITT specification, any background injected. It has no library entry to ground it. This is where hallucinated statistics, invented certifications, and fabricated case study details appear. Ji et al.'s survey of hallucination in natural language generation identifies "intrinsic hallucination" as output contradicting source material, but in bid writing the problem is more often extrinsic: the model generates plausible-sounding facts with no source at all.
The model writes reliably: Methodology descriptions adapted from a library entry. Team structure descriptions from a provided CV list. Social value commitments derived from an existing policy statement. Evaluation-weighted response structure.
The model invents confidently: Percentage improvement figures ("we delivered a 34% efficiency gain"). Named accreditations not mentioned in the prompt. Client names and contract values from previous work not provided in context. Named personnel for roles where no CV was injected.
Hallucination controls: the pre-submission review checklist that catches invented certifications
Every draft response goes through a structured pre-submission check before it reaches the director. This is not a read-through — it is a mechanical scan against specific rule categories.
- Percentage and numeric claims — every figure in the response must trace to a source document in the library. Mark the source inline; if unmarked, delete the figure before escalation.
- Certification and accreditation names — compare every named standard or scheme against the certification register. ISO 9001, Cyber Essentials, G-Cloud framework references — all must be current and held. Check expiry dates.
- Named personnel — every person named in a method statement must be a current employee or confirmed subcontractor with a signed CV in the library.
- Contract and client references — every named past contract must appear in the firm's reference list with a verifiable point of contact.
- Turnover and financial figures — verify against the most recent filed accounts. LLMs will round figures or advance a financial year without flagging it.
This verification pattern appears in similar form in our post on LLM contract review and NDA clause extraction, where clause-level accuracy controls follow the same principle: no number or named party passes without a traced source.
Word count compliance, mandatory formatting, and PDF submission pipeline for UK portals
UK portal submission requirements are inconsistently documented and inconsistently enforced — until a bid is disqualified for a formatting breach. Common requirements: word counts per question box (not per section), character limits distinct from word limits in some portals, and mandatory cover pages with specified field values.
The pipeline step between draft and submission handles three things automatically. First, word count validation: compare each generated answer against the ITT limit before export, and flag over-limit answers for human editing — never truncate automatically, because truncation cuts meaning. Second, format application: inject finalised answers into a Word or PDF template matching the authority's specified layout. Portals including Jaggaer require PDF/A format specifically; test this early, not the day before the deadline. Third, portal compatibility check: Delta eSourcing and some NHS procurement portals reject uploads over a certain file size or with embedded fonts. Strip non-essential metadata and flatten vector graphics before upload.
Do not automate the final submission click. The submission confirmation is a legal act. A director must review the formatted output and submit manually.
Win rate tracking by tender type: the feedback loop that improves the library after each result
A bid library that does not receive feedback is a static document store. After each result — win, loss, or no-award — log the outcome against the bid record and tag whether each question response was library-retrieved, model-adapted, or fully generated. When the authority provides evaluation scores (increasingly available under the Procurement Act 2023 transparency provisions), map the score per question against the source type.
Over eight to ten bids, patterns emerge. In our client work, library-retrieved method statements scored higher than fully generated ones in roughly 75% of audited bids. Questions answered with named personnel and verifiable references consistently outperformed those without. Questions where the model generated freely — no library match, no retrieved context — scored lowest and attracted the most evaluator comments about lack of evidence.
Use that data to identify library gaps and commission new entries before the next relevant ITT drops, not during it.
What changed in 2025–2026: Procurement Act 2023 changes and new Find a Tender portal requirements
The Procurement Act 2023 came into force on 24 February 2025, replacing the Public Contracts Regulations 2015. The transition has three direct effects on bid automation pipelines.
First, Pipeline Notices now require contracting authorities to publish anticipated contracts over £2 million in advance. For SMEs, this means ITT content is more predictable — you can start building library entries for a specific contract type before the specification is published, based on the notice text.
Second, the Central Digital Platform (CDP) creates a single supplier profile that authorities can access. Standard company information questions — turnover, insurance, accreditations, policies — are increasingly pre-populated from the CDP record. Update your CDP entry whenever an accreditation renews or a financial year changes; discrepancies between CDP data and ITT answers trigger compliance queries. The Crown Commercial Service supplier guidance carries the most current information on which procedures remain under transitional rules.
Third, the Find a Tender service now aggregates contracts across devolved administrations more consistently. Filter by CPV code, contract value, and authority type to build a reliable forward pipeline of opportunities your library already covers.
Good / Bad / Ugly: three bid automation approaches and their accuracy at submission stage
| Approach | What reaches the evaluator | Hallucination risk | Typical drafting time |
|---|---|---|---|
| Good: Embedding retrieval + library + human checklist | Library-grounded drafts with model adaptation | Low — constrained to verified content | 4–6 hours for a 20-question ITT |
| Bad: Zero-shot LLM with no library | Fluent, confident, ungrounded responses | High — numbers and certifications invented | 1–2 hours to draft, unpredictable review |
| Ugly: Manual copy-paste from Word documents | Previous bid sections without structured adaptation | None from LLM; factual staleness accumulates | 16–20 hours — no improvement over baseline |
Good: Embedding-based retrieval from a structured compliance library, with the model adapting library content to the specific ITT and a human conducting the pre-submission checklist. The 14-hour first-draft time in the opening example used this approach. Time drops further on subsequent bids as the library grows and retrieval coverage improves.
Bad: Zero-shot LLM given the ITT and a brief company description. Output reads confidently and often sounds more polished than the Good approach. It will also reference ISO certifications the firm does not hold, cite a client satisfaction score that was never measured, and name a project manager who left eighteen months ago. Evaluators with domain knowledge catch this at scoring stage.
Ugly: A flat folder of previous bids, sections manually copied into new responses. No hallucination problem — but no time saving either, and the approach accumulates outdated content faster than anyone cleans it up. By bid five, the folder contains three contradictory versions of the social value statement, none current.
The firm we audited in April moved from the Ugly approach to the Good approach over a three-week sprint. For the proposal generation pipeline that preceded this work, see the post on automated proposal generation for professional services firms.