Bid Knowledge Base: The Bid Office’s Second Brain
TL;DR
- Construction teams spend 11.5 hours a week searching for data (Deloitte, 2023, via Skanska)
- Prices, exclusions and clarifications sit in archives that drafting tools and content libraries rarely reach
- Search with sources and per-document permissions finds them; price tables also need structured extraction
- Start with one team’s closed tenders and one question type, logging what nobody could find
Why do large bid teams need a bid knowledge base?
A bid office already has a second brain: its archive. Much of what decides a large bid has been written down somewhere in the company; finding it before the deadline is the hard part. FMI estimates that 96% of construction data goes unused, and 90% of it is unstructured, as Skanska’s director of data and analytics noted in 2024. From the software side, the documents are there; what is missing is a search that reaches all of them.
11.5 hours
a week spent searching for data (Deloitte, 2023, cited by Skanska)
41%
of the industry’s most experienced workers expected to retire by 2031 (US figure)
Sources: Deloitte (2023) and Dusty Robotics, both cited by Skanska (2024); US industry figures.
Search that shows its source gives people something to check, which generated text on its own cannot.
None of this will surprise a bid manager. Earlier prices, clarifications, exclusion history and lessons each sit in a different place, from an estimating file to a partner’s records or someone’s head. Some of it arrives in the final week.
To an engineer, each of those places is a separate system with its own search, or none at all. Skanska saw the same in its own data: “We realized we had a lot of siloed data. It did not talk to each other.” Often the only link between a document and its reader is a colleague’s memory, and memory retires. “A lot of knowledge is walking out the door,” said Brian Karas of Skanska USA Building in 2026.
According to Microsoft’s customer story, Mott MacDonald rolled out its search agent EMMA (Every Mott MacDonald Answer) to 20,000 employees in nine months. EMMA has since handled over 220,000 unique employee queries. At Skanska USA Building, a safety assistant answers from internal documents and safety regulations; it “provides a hand when someone is making a decision”, Karas explains. Neither is a bid tool. Both do the same thing: search a company’s own documents while people decide.
Our view
For a large bid team, the first AI project should be search over its own archive that always shows the source document, not a tool that writes new text faster. We build software, not bids. The price, the exclusion reason or the clarification exists. The systems around it just do not bring it to the person deciding. We see three ways in: search, integration with the places files already live, and data work on dates and access rights.
Most bid automation tools help teams draft, with content libraries, templates and AI suggestions, as Melissa Fletcher, CF APMP, described the category in 2025. That helps you write faster. It does not reach documents outside the tool’s own library. In our view, the commercial and legal knowledge of a bid rarely lives in one.
Where we would start
If the knowledge already exists, the first thing to build is a way to find it with its source:
- one archive: the closed tenders of a single bid team;
- one kind of question, such as earlier prices or earlier exclusions with their grounds;
- every answer tied to the document it came from and its date;
- a person deciding, and a log of what people asked and could not find.
That’s the short version, a three-minute read. Below are the details and data behind it, for anyone who wants to go further.
The details
Where does bid knowledge sit, and why is it hard to find?
Bid knowledge sits in several systems at once (folders, estimating files, inboxes, tender portals) and in people’s memory, and no single search reaches all of them. A 2012 cross-industry estimate by the McKinsey Global Institute (MGI) put the time interaction workers spend looking for internal information or colleagues at nearly 20% of the workweek; it is old, but the pattern has not gone away.
For a bid, in our assessment, the picture looks like this:
| Bid knowledge | Where it usually sits | Who asks for it |
|---|---|---|
| Past offers and unit prices | Folders, estimating files | Estimators |
| Buyer clarifications | Tender portal, inbox | Bid managers |
| Exclusions, penalties | Legal and partners’ records | Bid managers, lawyers |
| Lessons learned | Close-out notes, memory | The bid team |
Each row has its own search, or none. Timing makes it harder: under Directive 2014/24/EU, Art. 53(2), clarifications can arrive as late as six days before the deadline. Senior estimators know where things are; an index keeps that findable after they retire.
How does RAG for bid management find a past price?
Retrieval-augmented generation (RAG) for bid management searches the archive by exact terms and by meaning, then grounds each answer in the passages it found and can show its source. Each part of RAG has a job:
- Exact terms: keyword search matches item codes, clause numbers, dates and names (hybrid search documentation).
- Meaning: vector search finds “delay penalty” in a clause headed “liquidated damages”.
- One ranking: Reciprocal Rank Fusion (RRF) merges both result lists.
- Sources and permissions: answers can carry their source, and the index can inherit SharePoint or Microsoft Entra ID permissions.
The index reads documents where they live, as in RAG on historical project data.
Example: an estimator looks up a past unit price
Imagine an estimator who asks what the team priced for a similar item before. Keyword search finds the item code in closed offers; vector search finds the same work, worded differently, in another, each with its source and date. Hits from 2021 and 2022 stand out: tender prices vary with time, place and competition, Eurostat notes. Whether any price still holds is the estimator’s decision.
Why do price tables need more than vector search?
Most past prices live in spreadsheets and priced bills of quantities, not in prose. Vector search handles text well and tables poorly: a unit price only means something together with its item, unit, currency, VAT basis and date. So price tables are first extracted into structured records, brought to the same unit and basis, and searched next to the documents they came from.
Scanned offers go through OCR before any of this. The quality of answers is measured on a fixed set of real questions from the bid team, by how often the answer points to the right source. The examples here use Azure AI Search documentation; the same approach works on other search stacks.
A content library cannot do that for documents nobody put into it. Many bid teams already pay for software with a library, which raises a fair objection.
Doesn't your bid software already have a knowledge base?
Most bid automation tools keep a library of curated answers; a bid knowledge base searches the archive itself, including prices, exclusions and clarifications nobody curated. Melissa Fletcher, CF APMP (2025), lists “content libraries” first among their features.
Some will say
A well-kept library already is our knowledge base. For reusable answers, it is. Prices, exclusions and clarifications live in the archive the library never sees.
Four ways to search a bid archive, in our assessment:
| Approach | Finds | Misses | Source and permissions | Effort to set up |
|---|---|---|---|---|
| Folder and file names | Matching file names | Content inside files | The file; share rights | None |
| Document system keywords | Exact terms and codes | Same idea, other words | The document; system rights | Low, if a document system exists |
| Content library | Curated answers | Anything not curated | The entry; library rights | Ongoing curation |
| Hybrid search with sources | Terms and meaning | Documents outside the index | Every answer; per-document rights | Highest: indexing, permissions, upkeep |
| Verdict | Hybrid search finds the most, at the highest set-up cost |
Where does a bid knowledge base stop working?
A bid knowledge base stops working where the data is uncurated, the permissions are loose or the figures are old, and it never makes the decision itself. Each limit has a software side and a person who decides:
| Limit | What software can do | Who decides |
|---|---|---|
| Uncurated data | One curated archive first; log unanswered questions | The bid office |
| Loose permissions | Access per document, from indexing to query | Management, lawyers |
| Old figures | Date and source on every figure | The estimator |
Price workings are sensitive: a trade secret needs “reasonable steps” under Directive (EU) 2016/943, and whether access rights count as such is a legal question. Figures also age: Eurostat’s construction producer prices for new residential buildings rose 5.8% across the EU in 2021, with double-digit rises in most countries in 2022.
Example: a bid manager asks about a partner’s history
Imagine a bid manager forming a consortium who asks whether a partner had a contract terminated in past tenders. The answer lists the documents the bid manager may see, with sources and dates.
A bid knowledge base is only as good as its data, its permissions and the date on each figure. That is why a first slice should start small, where each limit can be watched.
What does a first slice look like in practice?
A first slice starts with one team’s closed tenders and one question type, and the query log is read before anything is added. Mott MacDonald began with a proof of concept, according to Microsoft; Brian Karas of Skanska USA Building says “stick with the knowledge base that you can maintain” (EHS Today, 2026).
The log shows, in the words of Mott MacDonald’s Paul Skidmore, “the information they’re trying to find”. Built this way, the same index can later feed a pre-submission check, which we describe in our article on tender automation for large bids.
How we do it
At DO OK, every change an AI helps make is approved by an engineer; nothing reaches production automatically. That is the rule behind AI search built with a human in the loop.
What a bid knowledge base changes, and what it does not
A bid knowledge base changes how fast the documents behind a decision reach the person making it; it does not change who makes the decision. The records exist. What fails is their route to someone on a deadline.
That route is what software can change: search, integration, dated and permissioned data, and a team to build it.
Frequently asked questions
Is a bid knowledge base the same as a bid AI assistant?
Not quite. An assistant that answers from the archive is one interface to a bid knowledge base; a list of search results is another. Either way, the software must ground answers in your own documents, show each source and enforce access per document.
Does AI for tenders train on our bid documents?
It depends on the provider’s terms. Microsoft, for example, states that prompts, outputs and embeddings in Foundry Models sold by Azure are not used to train generative AI foundation models without your permission. That covers one service; other providers’ terms are worth reading the same way.
Can a bid knowledge base flag outdated prices?
Partly. Software can show the date and source with every figure and set older ones apart. That matters because construction prices moved fast in 2021 and 2022, according to Eurostat, so a figure’s age changes what it is worth as a reference.
Can software decide whether a past exclusion must be declared?
No, and it should not try. Software can put every related document in front of the person deciding, with sources and dates. In Poland, KIO 251/24 (2024) described the bidder’s task as presenting such material to the buyer. What must be declared stays the legal team’s decision.
Talk to us about a bid knowledge base for your team
If your bid archive holds more than your tools can find, send us a short note through the form. We will come back to talk through your archive and the one question type worth indexing first, and whether we build it with you or add engineers to your team.
Sources
- Skanska, 2024
- EHS Today, May 27, 2026
- Microsoft Customer Stories, Mott MacDonald
- McKinsey Global Institute, July 1, 2012
- Melissa Fletcher, CF APMP, LinkedIn, August 12, 2025
- Eurostat, Statistics Explained, October 2025
- KIO 251/24, February 16, 2024
- Microsoft Learn, Hybrid search in Azure AI Search, 2026
- Microsoft Learn, RAG and generative AI in Azure AI Search, 2026
- Microsoft Learn, Document-level access control in Azure AI Search, 2026
- Microsoft Learn, Data, privacy, and security for Foundry Models, 2026
- Directive 2014/24/EU, Article 53
- Directive (EU) 2016/943, Article 2
