Start here

User guide & FAQs

The workbench turns a folder of past bids into a defensible starting number for the next one. It is deliberately conservative: it shows you what it knows, what it does not, and how far it is willing to go — and it never invents a figure to fill a gap.

The idea in one paragraph

Every past RFP, pricing workbook and staffing plan in the two Drive folders has been parsed into 156 comparable records. When a new solicitation arrives, you load it here, the tool extracts its scope, and then five independent methods each try to answer the same question — how much effort is this? Where they agree you get a tight range; where they disagree the range widens and the confidence drops. You leave with a ROM, an indicative price, a shareable brief and a working Excel model.

The six-step workflow

1
Look at what the archive already knows

Start on Corpus overview for the shape of the archive, then Historical opportunities to filter all 156 records by domain, jurisdiction, year, value band or basis. Open any row to see its Drive folder, source documents, staffing, scope and the caveats attached to it.

2
Load the new RFP

Go to New RFP → estimate and either paste the text or drop the file — PDF, Word, Excel, CSV, text, or a whole ZIP of the procurement package. For workbooks and ZIPs you pick which sheets and members to read, so cover tabs and boilerplate stay out.

3
Extract the requirements

Press Extract requirements. The page pulls candidate requirement statements out of the text and tags each one to a capability area. Edit, re-tag or delete anything it got wrong — the estimate uses this list, so five minutes here is the highest-value five minutes in the workflow.

4
Add what you know

Fill in the opportunity facts you have — jurisdiction, users, counties, duration, budget ceiling. Optionally upload a product maturity inventory (a CSV template is on the page) to unlock the fit-gap method, and enter bottom-up WBS effort if you already have a build-up.

5
Run the estimate

Press Run estimate. Five methods run independently; each one either answers with a range or reports what it is missing. The reconciled ROM spans the methods that answered, with dispersion and a confidence band alongside it, plus an indicative price from the archive rate card and role mix.

6
Share it and build scenarios

Download the PDF brief to send to Slack or email, and the Excel model to hand to whoever will build the commercial scenarios — its Inputs and Rate card tabs are live, so changing margin, mix or effort recalculates the price. Export the JSON if you want the run to be reproducible later.

How to read the numbers

Aggressive / Expected / Conservative is a range, not three forecasts. Quote the expected figure with the range attached, and say which methods answered.
⚠ Three things to say out loud when you share an estimate
  • The archive is bid data — proposed effort, not delivered cost.
  • Blocked methods were excluded, not guessed; more input narrows the range.
  • Anything flagged as template carryover means an original workbook repeated a total it never recomputed.

Frequently asked questions

What is this tool?

A workbench for pricing new government-IT bids using what the firm has already bid. It holds a parsed archive of 156 past procurement folders — RFPs, pricing workbooks, staffing plans — and uses them to produce a rough order of magnitude (ROM) effort and indicative price for a new solicitation, with every number traceable back to a source.

Where does the data come from?

Two Google Drive folders of past pricing responses, RFIs and RFPs. They were parsed once into a fixed snapshot: 156 records, 147 of which are opportunities. The archive explorer links back to the original Drive file for every record, so you can always open the source.

Is the archive actual delivered cost?

No. Almost everything is BID data — what was proposed, not what was delivered. Exactly one record in the archive carries delivered effort. Every screen labels the basis (bid, mixed, actuals) so you never quote proposed effort as if it were history.

Why does an estimate sometimes say a method is blocked?

Because the inputs that method needs are missing. Five methods run independently: comparable historical, parametric, bottom-up WBS, product fit-gap and external benchmark. A method with no inputs reports blocked and is excluded from the range. Nothing is filled in with a plausible-looking number.

What is dispersion, and why does the range widen?

Dispersion is the ratio between the highest and lowest expected answer across the methods that answered. When the methods disagree beyond the threshold, the reconciled range is widened to span them rather than averaged, because averaging disagreement hides it.

What does the confidence band mean?

It scores how much real evidence the run stands on: how many methods answered, how close the comparables actually were, whether the underlying records are bid or actuals, and whether any matched record is flagged as template carryover. Low confidence means use the number to start a conversation, not to sign a bid.

What is template carryover?

Some archive records share byte-identical totals with unrelated opportunities, because a pricing workbook was copied and the numbers were never changed. Those records are flagged wherever they are used, so a suspiciously repeated total is visible rather than silently averaged in.

What file types can I upload as an RFP?

PDF, Word (.docx), Excel (.xlsx/.xlsm/.xls), CSV/TSV, plain text and Markdown — or a whole ZIP of a procurement archive, which is unpacked and listed so you can choose which members and which workbook sheets to read. You can also just paste the text.

Does my RFP get uploaded anywhere?

The file is parsed in your browser; the file itself is never sent to a server. If you run the research brief, the extracted text is sent to the AI gateway for that one call.

How much does the research brief cost?

It depends on how much text is read. Fast reads about 25k characters on the cheapest model, Balanced about 60k, Deep up to 120k on the strongest model — roughly 1×, 4× and 20× relative cost. Start with Fast for triage and only re-run Deep on the bids you shortlist.

Does the research brief search the web?

No. It reads the solicitation you loaded and the archive comparables. Its Market context section is unverified model knowledge about similar programmes elsewhere — a lead to check, never evidence. Live web research can be added later if you want verified citations.

What do I get to send to a client or to Slack?

Two downloads on the estimate page. The PDF brief is a shareable read-only summary: ROM range, method-by-method results with blocked reasons, indicative price, comparables, caveats and the snapshot stamps. The Excel model is a working file: blue input cells on Inputs and Rate card, live formulas on Pricing and Summary, so anyone can build scenarios from the baseline.

Can I save and reload a run?

Yes. Export estimate JSON writes the inputs, requirements, module inventory and every method result to a file; Import estimate JSON restores the page exactly, including the snapshot stamps a reviewer needs to reproduce it.

Does the archive update when new files land in Drive?

Not yet. The corpus is a fixed snapshot from the initial parse. Incremental Drive sync — detecting new folders and changed files and queuing them for review — is the next thing being built.

Is there a login?

No, not at the moment. Anyone with the link can use the workbench, so treat it as internal-only until access control is added.

Where to go next