Who confirms the tender requirements AI extracts?

One missed minimum requirement can end a bid before anyone scores your answers, and several of the rules that can end it sit outside the questions. In SEQUESTO, specialist agents in Agent Force pull every question, requirement, evaluation criterion, mandatory attachment, word or character limit and dated requirement into one list, and the bid manager confirms that list before anyone drafts.

That list is what tender requirements extraction produces. It's the first step of a tender response, and every later compliance check relies on it. The stakes are plain in the administrative specifications for the European Commission's call for tenders EC-CNECT/LUX/2026/OP/0056, an open procedure run by DG CONNECT. A bid that doesn't meet the minimum requirements is rejected. A ground for rejection at any stage of the evaluation, from access and administrative compliance through exclusion and selection to the award criteria, stops the bid from being fully evaluated. For where extraction sits among the other stages from notice to award, see the supplier-side tender process.

How a tender pack becomes a confirmed list

In SEQUESTO aOS, a tender pack becomes a confirmed list in five steps. The files go into one project, the agents pull out the requirements, and James builds the timeline around your sign-off gates. The bid manager then confirms the structure, and the team tracks each item to submission. The Commission call serves as the illustration throughout. It's a public tender anyone can download, not a SEQUESTO customer project, and it's useful because several of its rejection grounds, such as an upload list, a rule on covering every requirement and scoring thresholds, sit outside the questions.

Every file goes into one project

Each tender is its own project, with its own files, deadlines and contributors. In the Commission call, the administrative specifications are Part 1, and the technical specifications sit in a separate Part 2. Bidders are told to take into account the full set of documents, including the draft contract's provisions. All of it goes into the project. Word, PDF, Excel, images and slide decks are converted and indexed. Scanned pages are read with OCR, and the dropdown options behind each field in an Excel questionnaire are read too.

Completeness matters more than order here. The agents work from what's in the project, so a draft contract left in someone's inbox adds nothing to the list.

What the agents pull out

Agent Force's specialist agents extract the questions, requirements, evaluation criteria, evidence requests, mandatory attachments, and word and character limits into one working response structure. Several of the call's rejection grounds are not questions at all. Annex 1 lists every document to submit for each bidder type, including its file name and where to upload it. Failing to upload the technical or the financial response in the Commission's eSubmission system leads to rejection. A technical response that doesn't cover all the requirements may be rejected as non-compliant and not evaluated further.

The award criteria carry their own pass marks. Price weighs 30% and quality 70%. Quality splits into three criteria worth up to 40, 40 and 20 points, with minimums of 24, 24 and 12. On top of those, a response that scores below 60 out of 100 overall on the technical evaluation is rejected. A criterion on the list therefore needs its weight and its threshold beside it, or the team drafts towards a score without knowing where the floor is.

The list also has to tell pass/fail items apart from scored ones. The call's minimum requirements can't be met with any assumption, limitation, condition or reservation. That makes them a different kind of item from a criterion where the team chooses how much to invest. James reads the pack like an experienced bid manager and flags which requirements depend on one another. For example, a quality criterion might score how the bid meets a requirement set out in the technical specification.

The timeline and its sign-off gates

Specialist agents extract every dated requirement into a task list from day one. James then reads your organisation's governance process alongside the tender. He inserts its sign-off stages, financial verification points, and resource review gates as milestones, each with a time and a named reviewer. The result is a plan worked back from the submission date that holds your own approvals, not a generic template.

Some dates fall after submission. The Commission call says it will request certain documentary evidence later on short deadlines, and it invites bidders to prepare that evidence in advance. Put those items on the list from the start, each with an owner, so that preparing them doesn't wait for the buyer's request.

Confirming the structure before drafting

Before anyone drafts, the bid manager confirms the extracted structure, accepting, adjusting or rejecting what the agents returned. This is where a person checks the items that sit outside the questions. On the Commission call, the bid manager would tick off:

• the page limit on the technical response: 40 pages excluding the cover page and table of contents, because the evaluation committee doesn't consider pages beyond it
• the typography rules: Times New Roman, at least 11 points, A4 pages and margins of at least 15 mm
• every Annex 1 upload for your type of bidder
• each quality criterion with its threshold, and the overall technical threshold
• each minimum requirement marked pass/fail, apart from the scored items
• evidence due after submission, with its owner
• requirements from Part 2 and the draft contract, not only from Part 1

Each confirmed item is a decision the bid manager can stand behind at submission, which is when the team makes its own declaration.

Tracking each item to submission

Once the list is confirmed, the team works from it. Each question or section carries an owner, a status and a deadline. Each question can have its own reviewer and approver, so compliance, legal and technical reviews each take their own route. The team follows progress in Kanban, timeline or table views, and items stay tracked until the bid manager signs them off.

That sign-off feeds a declaration the tool never makes. In the Commission call, bidders declare in eSubmission, at the moment they submit, whether their response complies with the minimum requirements. The tracked list is what the bid manager reads before answering that question.

What extraction leaves with your team

Extraction gives your team a list to confirm. The compliance decision, the declaration and the submission stay with people. SEQUESTO extracts and tracks, but a person on the team confirms the structure, signs off each item and decides what goes out. Final wording and sign-off stay human.

Submission stays with your team too. SEQUESTO doesn't submit on the team's behalf. The export keeps the document's original structure, sections and order, and a person on the team uploads it in the buyer's portal. Whoever does that acts on decisions the bid manager has already confirmed and signed off on, not on the agents' output.

Extraction only covers the documents in the project. It works from what the team has received and added, so a clarification response or an amended document changes nothing until someone adds it. Within a project, James follows its rules and asks before taking any action outside the agreed scope. That boundary is what InfoSec and legal need to know: the tool acts only within what the project allows, and anything beyond it goes to a person.

Nobody should promise that an extracted list is complete, and SEQUESTO doesn't. The list is a draft until the bid manager confirms it, which is why confirmation comes before drafting rather than after. Every item a person accepts, adjusts or rejects is a judgement the team owns. The declaration at submission is made in the team's name, not the tool's. If an item turns out to be missing at that point, it's the team that answers for it. That's why we keep the confirmation step human, no matter how good the extraction gets.

Where extraction sits in SEQUESTO aOS

Requirement extraction belongs to the Bid Capabilities layer of SEQUESTO aOS and runs on the Agentic Harness and Context & Knowledge beneath it. In our view, it carries two of SEQUESTO's pillars: governance and human-in-control.

The layers underneath

The Bid Capabilities layer is where the confirmed list drives the work. It holds the configurable workflows with their review chains and approval logic, the question-level draft answers and the export that keeps the buyer's structure. Beneath it, Context & Knowledge converts and indexes the tender files the agents read. The Agentic Harness is the runtime under Agent Force. Agents act through defined tools and typed function calls, so you can inspect the inputs and outputs of each extraction step, and trace retrievals and tool calls. Every confirmation then lands in the action log kept by Security & Governance.

The roles stay distinct across the layers. James orchestrates: he reads the pack, flags dependencies and builds the timeline. Agent Force's specialist agents carry out the extraction. The bid manager owns confirmation and sign-off.

The pillars it carries

We see extraction as governance first. The list runs through the workflow your organisation configured, with your sign-off stages and named reviewers in the timeline, rather than through steps someone reassembles for each bid. It also keeps the bid manager in control. The bid manager owns the structure the team drafts against and the final declaration. The agents do the reading, and your team keeps the judgement.

Test it on a tender you have already submitted

The fairest test of any extraction method is a tender your team has already submitted, with the compliance matrix you built for it as the pass mark. Your current method goes through the same test as every other option, including any general-purpose AI tools your team uses, which write a decent first draft fast, and any purpose-built response platform. Run it in this order:

1. Choose a closed tender with several files and at least one rule outside the questions, and clear it with InfoSec for each option.
2. Give each option the documents only, never the matrix, so that no run can be tuned to your answers. For your current method, have someone who didn't work on that bid build the list fresh.
3. Judge each pass/fail item on your matrix per item, allowing no misses, and check every extra item an option lists against the documents.
4. Record hours on a confirmed list for every role that touched it: the bid manager, subject matter experts, legal, and InfoSec.

Set-up, InfoSec clearance and training repeat for each option, so name the people and hold their hours per option before the first run. An option that misses a pass/fail item on your matrix fails, no matter how quick it is. Among the options that pass, compare total hours across all roles with your current method, and count each gain once: an item your matrix missed is a finding, not also time saved. If nothing beats your current method on hours without missing an item, keep your current method.

Start by pulling the matrix from your last submitted tender and counting how many of its pass/fail items sat outside the questions.