Proposal and bid drafting

• The tender pack is read within an hour of download, and a compliance matrix is built: every question, word limit, scoring criterion, mandatory requirement, deadline and any rule the buyer sets on AI use.
• A first draft of each answer is written from the firm's approved library only, not from general knowledge, and tailored to the buyer's scoring criteria.
• Every factual claim in a draft links to the evidence behind it: a case study, a certificate, a CV, a policy or a figure from finance. Claims without evidence are highlighted, never smoothed over.
• Out-of-date certificates, people who have left, and case studies the client hasn't agreed to be named are caught before a person reads the draft.
• Subject matter experts spend their time improving answers rather than starting from a blank page.
• The bid manager signs off every answer and the bid director approves the submission. Pricing, commitments and declarations are always written by people.

• Cabinet OfficeProcurement Act 2023 (in force 24 February 2025): exclusion grounds
  →

  A buyer may exclude a supplier that acted improperly in a procurement, which includes providing incomplete, inaccurate or misleading information, or for professional misconduct that brings its integrity into question. Every claim is traced to evidence, and the bid manager signs off each answer, so there is a record of who checked what.

• Cabinet OfficePPN 017: Improving transparency of AI use in procurement (replaces PPN 02/24 for Act procurements)
  →

  Suppliers' use of AI isn't prohibited, but buyers may ask whether AI was used, for what, and how its output was checked, and may restrict the use of their confidential information to train AI. The platform records which answers were AI-assisted and who checked them, so any such question can be answered accurately. Tender documents are never used for training.

• Cabinet OfficeNational Procurement Policy Statement and PPN 002: the Social Value Model
  →

  In-scope central government buyers must use the Social Value Model for Act procurements from 1 October 2025, usually with at least 10% of the score. Social value commitments become contract obligations, so the AI drafts from the firm's real programmes and data and never proposes a new commitment. People decide what the firm promises.

• Cabinet OfficeThe Central Digital Platform (Find a Tender)
  →

  Core supplier information, connected persons and exclusion declarations are held on the platform, and the supplier confirms they are accurate. They come from the firm's official records, owned by the company secretary or legal team, and are never drafted by the AI.

• Home OfficeFraud Act 2006 and the failure to prevent fraud offence (Economic Crime and Corporate Transparency Act 2023, in force 1 September 2025)
  →

  A knowingly false claim made to win work can be fraud by false representation, and large organisations can now be liable if someone working for them commits fraud intended to benefit them, unless they had reasonable procedures to prevent it. The evidence check and the sign-off log form part of those procedures. In private bids, a false statement that induces a contract can also be a misrepresentation.

• ICOUK GDPR and the Data Protection Act 2018
  →

  CVs and staff details are personal data. Staff are told how their CVs are used in bids, only bid-relevant details are indexed, CVs are refreshed from the HR system rather than copied, and leavers drop out of the index. Clearance levels and anything sensitive stay out of the library unless a bid needs them.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• A named owner and an expiry date for every item in the evidence library.
• Written client permission recorded against every case study that names the client.
• A bid/no-bid decision by people before drafting starts, so the AI never makes bidding cheaper than deciding not to bid.
• A record of which answers were AI-assisted and who checked them, ready for any buyer question on AI use.
• A monthly audit of a sample of submitted answers against their evidence.
• Bid staff and experts trained before go-live, including how to challenge a confident but unsupported draft.

1. 3–4 weeks

   Discovery

   Map today's bid process, measure the baseline on recent bids, audit the content library, and agree the evidence rules with legal and the bid director.

2. Gate · must be true to passBaseline agreed, library owners named, DPIA started, client-naming permissions checked.

3. 8–10 weeks

   Pilot

   Live bids in one sector, in shadow mode at first: the AI drafts alongside the team, and drafts are compared with what was submitted.

4. Gate · must be true to passEvery claim in pilot submissions linked to evidence, zero unsupported claims in the audit sample, bid director sign-off.

5. 4–6 weeks

   Production

   AI drafting becomes the starting point for every bid in that sector, with monitoring, support and a rollback plan.

6. Gate · must be true to passMonitoring, support, library upkeep and bid team training in place.

7. Ongoing

   Scale

   Add sectors and bid types one at a time, then reuse the library for proposals, framework call-offs and capability statements.

8. Gate · must be true to passEach new sector passes the same evidence audit.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   The AI writes fluent answers with figures nobody can source.

   How the design avoids it

   Drafts come only from the approved library, every claim links to its evidence, and unsourced claims are flagged, not smoothed over.

2. 2

   Goes wrong

   The library is full of old bids, so drafts quote leavers, lapsed certificates and old numbers.

   How the design avoids it

   An owner and an expiry date on every item, and HR and policy changes fed in as events.

3. 3

   Goes wrong

   A case study names a client who never agreed to it.

   How the design avoids it

   Client permission recorded against each case study, and a guardrail that blocks names the library doesn't allow.

4. 4

   Goes wrong

   Promises in the bid become contract terms nobody planned to deliver.

   How the design avoids it

   The AI never proposes commitments, KPIs or social value pledges. Delivery leads approve every one.

5. 5

   Goes wrong

   The firm can't say how AI was used when the buyer asks.

   How the design avoids it

   A record of AI-assisted answers and who checked them, kept with the bid.

1. From 1 October 2026
   PPN 022 on steel: central government buyers must state in the contract details notice whether UK-produced steel will be used, for procurements of £10 million or more, or over 500 tonnes of steel.

   Engineering bids will be asked for steel origin data from the supply chain. Add it to the library as evidence with an owner, not a figure written fresh for each bid.

2. 1 May 2027
   Below-threshold contract notices move from Contracts Finder to Find a Tender, deferred from October 2026 by the Procurement (Amendment) (No. 2) Regulations 2026.

   Point opportunity monitoring and CRM feeds at the Central Digital Platform, so smaller opportunities aren't missed during the switch.

3. Late 2026–2028
   The Cyber Security and Resilience Bill brings medium and large managed service providers into regulation, with duties set by secondary legislation.

   Firms bidding for managed IT or digital work should expect new security questions in tenders. Build the evidence now, with owners and dates.

4. 2026–2027
   The ICO's statutory code of practice on AI and automated decision-making, required since May 2026.

   Keep the DPIA and the record of how staff data is used in bids current, so the library can be shown to meet it.