Use caseKnowledge and searchDocument AI
Finding public tenders worth bidding on, and answering them from your own bid library
Tender monitoring by CPV code, bid/no-bid scoring, a compliance matrix with page references, and draft answers only from your approved bid library.
A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, .
The short version
An AI integration for firms that bid on public tenders in the EU and UK. It monitors TED and national portals by CPV code and keyword, scores each notice against your capabilities, references and capacity, and extracts every requirement from the tender pack into a compliance matrix with page references. Answers are drafted only from approved past answers, project sheets and CVs in your SharePoint bid library, with citations. Where no approved source exists, a person writes it. Clarification and submission deadlines are tracked throughout.
- Best for
- Engineering consultancies, construction subcontractors and agencies that bid on five or more public tenders a month.
- Connects to
- TED and national tender portals, SharePoint bid library, Word, Excel compliance matrix, Microsoft Teams
- The AI does
- Reads 100 to 300 page tender packs into a requirement list, finds matching approved content, and drafts answers that cite their sources.
- People do
- The bid manager decides whether to bid, experts write what the library cannot answer, and a director signs off every submission.
- Built as
- AI Integration, usually $15K - $35K
Five tenders a month, and every one starts from a blank matrix
Take an 80-person engineering consultancy in the Netherlands or Germany: civil and structural engineering, building services, some project management. Public clients bring in half its revenue. It bids on five to ten tenders a month, from municipal framework agreements to design work for a regional hospital, and one bid manager coordinates them all with engineers who write their sections in the evenings.
Every tender pack runs to 100 to 300 pages: the notice, instructions to tenderers, specification, draft contract, award criteria, and an ESPD for the exclusion and selection questions. Before anyone writes a word, someone reads all of it and builds a compliance matrix by hand, requirement by requirement, with the page each one came from. Many of the answers are ones the firm gave last month, rewritten from old files because nobody can find the approved version.
The expensive failures are not the lost bids. They are the bids that should never have been started, the clarification window that closed while the team was busy, and the submission excluded on a formality: a page limit exceeded, a certificate that expired in March, a reference project from outside the permitted window. Bid capacity is the real constraint, and too much of it goes on reading and searching.
One tender, the way it runs today
From notice to upload, before any automation.
- The bid manager spots the notice in a daily portal email, or a colleague forwards it a week after publication.
- Someone downloads the pack and skims it to decide whether to bid, mostly on instinct and contract size.
- The team reads everything and builds the compliance matrix in Excel, one requirement per row.
- Engineers dig through SharePoint and old submissions for answers, CVs and project sheets, then rewrite them to fit.
- The bid manager gathers certificates, insurance proofs and the ESPD, and checks each against the requirements.
- Clarification questions are drafted if someone notices a gap in time, often days before the question deadline.
- Formatting, page limits and the final check happen the night before, and the portal upload is done with an hour to spare.
None of this is done badly. It is done by hand every time, and the reading and searching consume the hours that should go into the answers that win.
How the assistant works a tender, notice to submission
Retrieval and drafting are the model's job. Scoring, deadlines and the rule that every sentence needs an approved source are code, and every decision to commit belongs to people.
01 Trigger · TED API, national portals
A matching notice is published
A daily job queries TED and the national portals you bid on by CPV code, keyword and region. New notices, and changes to ones you follow, corrigenda included, arrive with their documents.
02 AI model · Structured output
Read the notice and the pack
A model extracts the contracting authority, value, lots, deadlines, exclusion and selection criteria, award criteria and weights, and every requirement, each with its document and page reference.
03 Plain code
Score bid or no-bid
A weighted score in code: scope fit, value range, region, certifications held, approved reference projects inside the required window, team capacity in the bid calendar, and time left before the deadline.
04 Decision
Is this worth a bid?
The bid manager decides, with the score and its reasons in front of them.
- Strong fit and free capacity then a go request in Teams; accepting it opens the bid workspace
- Borderline, or a likely incumbent then the bid manager reads the reasons and decides
- A hard fail: missing certification, turnover threshold, no eligible references then logged as no-bid with the reason, so patterns show up over time
05 Plain code · Excel
Build the compliance matrix
Each requirement becomes a row with page reference, type (mandatory, scored, informational), owner and status. References are checked against the PDF text, so every row can be traced back to its sentence.
06 AI model · SharePoint via Microsoft Graph
Draft from the library
For each scored requirement, the system retrieves approved answers, project sheets, CVs and certificates, reranks them, and drafts an answer that cites each source. With no approved source, it drafts nothing and assigns the row to a person.
07 Person
Experts write, the bid manager reviews
Engineers write what the library cannot answer and edit the drafts. The bid manager checks compliance, and a director signs off the price and the final submission.
08 Result · Teams
Checks before upload, and a library that grows
Clarification and submission deadlines carry reminders. Before upload, code checks page limits, required attachments and certificate expiry. After the bid, newly approved answers return to the library tagged with their source tender.
A compliance matrix, as the assistant builds it
An excerpt from a design services tender. Every row points at its page, and the answer-basis column shows whether drafting has something approved to stand on.
| Ref | Requirement | Source | Type | Answer basis | Status |
|---|---|---|---|---|---|
| 2.1 | Minimum annual turnover of EUR 2 million in each of the last three years | Instructions, p. 14 | Selection, pass or fail | Annual accounts in the library | Ready |
| 2.4 | Two references for comparable design services completed in the last three years | Instructions, p. 15 | Selection, pass or fail | Project sheets: hospital wing, school campus | Check dates |
| 2.6 | ISO 9001 certification | Instructions, p. 16 | Selection, pass or fail | Certificate valid until March 2027 | Ready |
| 4.2 | Approach to BIM coordination, two pages at most | Award criteria, p. 22 | Scored | Two approved answers from 2025 | Drafted |
| 4.3 | Approach to reducing embodied carbon | Award criteria, p. 22 | Scored | No approved source | Assigned to a person |
| 6.1 | Professional indemnity insurance of EUR 5 million | Draft contract, p. 41 | Contract condition | Current policy covers EUR 3 million | Flagged to a director |
Drafting requirement 4.2, and declining 4.3
Two scored requirements from the same tender. One has approved material behind it. The other does not, and the assistant says so instead of improvising.
Bid workspace, Wednesday 11:30
- retrieve(requirement: "4.2 BIM coordination", approved_only: true)9 candidates, reranked / top 3: BIM answer July 2025, BIM answer March 2025, project sheet: hospital wing
- check_sources(ids)all approved / hospital project completed 2024, inside the window / named BIM coordinator marked as no longer employed
- draft(requirement: "4.2", page_limit: 2, sources: 3)draft of 780 words / every paragraph cites its source / 2 items flagged for an expert
Assistant · Word
Draft for 4.2 inserted into the bid template, with sources as comments. Two items need an expert: the July 2025 answer names a BIM coordinator who has since left, and the software versions it lists differ from what this tender asks for.
- retrieve(requirement: "4.3 embodied carbon", approved_only: true)2 candidates below the relevance threshold / 1 unapproved draft in a personal folder, excluded
Assistant · Teams
No approved source for 4.3 (embodied carbon, scored award criterion). Assigned to Pieter as sustainability lead, due Friday. Nothing has been drafted.
- When Pieter's answer is reviewed and approved, it is filed in the library with this tender as its source, so the next tender that asks has something to draw on.
Where bids fail on detail, and the rule that prevents each
Exclusion grounds and the ESPD
Exclusion grounds are pass or fail, and ESPD answers are formal declarations. The assistant pre-fills the ESPD from your company record and highlights anything that changed since the last one, but a director confirms every exclusion answer. The model never answers them alone.
Page limits and formatting rules
'Four A4 pages at most, Arial 11' is checked in code on the rendered document rather than estimated from word counts, and an answer over the limit is flagged at review, not discovered at upload.
Certificates that quietly expired
ISO certificates, insurance proofs and registrations carry expiry dates in the library. Anything that lapses before the submission date or the contract start is flagged when the matrix is built, which leaves weeks to renew rather than hours.
References that must really exist
Reference projects come only from approved project sheets with completion dates, values and client contacts, and the window rules (often three years for services and five for works) are checked in code. The model cannot introduce a project that has no sheet.
Clarifications and corrigenda
Authorities publish answers to bidders' questions and corrections to the documents while the tender runs. Every change to a followed notice is pulled and compared with the previous version, and the affected matrix rows reopen automatically.
Confidential packs and personal CVs
Tender documents can be confidential and CVs are personal data. Processing runs in an EU region under a data processing agreement, retrieval respects SharePoint permissions, and CVs of people who have left are excluded automatically.
Who is accountable for which part of a bid
The AI model
Read tender packs into requirements
Hundreds of pages of mixed legal and technical text, where reading is the bottleneck.
Retrieve and rerank approved content
Finding the right past answer among thousands is a search by meaning, not by keyword.
Draft answers with citations
A tailored first draft from approved material, with every source visible to the reviewer.
Plain code
Score bid or no-bid
Weights agreed with the directors, applied the same way to every notice.
Track deadlines and formal rules
Dates, page limits and expiry checks are arithmetic, not judgment.
Enforce 'no approved source, no draft'
The one rule that keeps unapproved claims out of a submission.
A person
Decide to bid, and sign the submission
A bid commits the firm's capacity, its prices and its legal declarations.
Write what the library cannot answer
New methods and new commitments need an expert's words first, then approval.
Altura, Tendium, Loopio, or your own library with a thin layer on top?
For discovery and bid management there are good products. Stotles covers the UK market well, Tendium the Nordics, Altura bid management in the Netherlands and increasingly Germany, and GovDash US government contracting. Loopio and Responsive are strong answer libraries for commercial RFPs and security questionnaires. If one of them covers your market and your team will genuinely maintain its content, buy it.
A custom integration wins when the value sits in what you already have: a SharePoint bid library, years of submissions in Word, project sheets and CVs your team keeps current, and a review process that works. Rather than migrating all of it into another platform, the assistant reads it where it lives, respects its permissions, and adds the matrix, drafting and deadline layer. It also wins when you bid across several countries and languages. The draft contract in each pack deserves its own review, which contract review against a playbook covers.
Timing matters too. The Commission proposed a single EU Public Procurement Act on 9 September 2026 to replace the three current directives, but it is years from applying, so today's rules and portals are what a bid team works with. The retrieval design, and why reranking matters for a library like this, is covered in the RAG architecture tutorial and the post on rerankers.
How you would know it is working
A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.
- Hours per bid, by stage
- Reading, matrix, drafting and review, logged per tender, so you can see where the time actually moved.
- Draft retention
- How much of each drafted answer survives expert editing, measured by edit distance per requirement type.
- Bid decision quality
- Win rate of strong-fit tenders against borderline ones. If borderline bids win just as often, the weights are wrong.
- Formal failures
- Submissions with an exceeded page limit, an expired certificate or a missed clarification. The target is zero.
- Library coverage
- Share of scored requirements with at least one approved source. It should rise every month as new answers are approved.
What a build like this costs
This is built as AI Integration, which runs $3.5K - $45K overall. A build like this one usually lands in the retrieval over your documents, with evaluation tier: $15K - $35K, 3-6 weeks. The first working version runs on your real data well before the end of that window.
What it costs to run
Reading a 200-page tender pack is the largest model cost, from well under a euro to a few euros per tender depending on the model, and drafting adds less. Embedding the library is a one-off cost plus small updates. At five to ten tenders a month, running costs are small next to one engineer's evening spent searching old files.
What moves the price
- How many portals and countries you monitor, and whether each offers an interface or only email alerts
- The state of the bid library: tagged, approved and current, or ten years of submissions in folders that need sorting first
- Languages: tenders and answers in one language, or Dutch, German, French and English side by side
- Whether drafts must fit strict Word templates, with styles, tables and page limits checked automatically
Who this is for
- Engineering and architecture consultancies bidding on five to ten public tenders a month with a single bid manager
- Construction subcontractors and specialist contractors answering framework and works tenders in several countries
- Agencies bidding for public communication, design or IT contracts with a SharePoint full of past submissions
- Firms that have lost a bid on a formality and never want that to happen again
Questions people ask about this
Can AI write tender responses?
It can write first drafts from your approved material, and that is the only way it should. Here every drafted paragraph cites an approved past answer, project sheet or CV, and where nothing approved exists, the assistant drafts nothing and assigns the requirement to a person. Experts edit, the bid manager checks compliance, and a director signs the submission.
How do I automate a tender compliance matrix?
Have a model extract every requirement from the tender documents with its page reference, then check each reference against the PDF text in code before it becomes a row. Add type, owner and status, and link each row to its answer. The matrix exists on the first day of the bid, and every row can be traced to its sentence.
Can it monitor TED for relevant tenders?
Yes. TED publishes EU notices in a structured format through an API, so a daily query by CPV code, keyword and region is straightforward, and national portals are added one at a time. Each match is scored against your capabilities, references and capacity, so the bid manager gets a score with reasons rather than a raw list of notices.
Does it work with our SharePoint bid library?
Yes, through Microsoft Graph, reading documents where they already are and respecting existing permissions, so nobody sees content they could not open before. Approval status comes from a metadata column or a folder convention, and only approved content is used for drafting. Nothing has to move to a new platform.
How much work is it to set up?
Most of the effort goes into the library rather than the software. Marking which answers, project sheets and CVs are approved and current usually takes the bid team a few weeks alongside the build, and it is worth doing anyway. Monitoring and the matrix can run on live tenders early, before drafting is switched on. It is an AI integration into tools you already use.
Is it allowed to use AI for public tender responses?
Generally yes. You remain responsible for every statement you submit, so read the instructions to tenderers for any conditions the authority sets. The rule that every claim traces back to an approved source is what makes AI assistance defensible: nothing goes into a bid that your firm cannot stand behind.
Sources