We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Accuracy
We measure the checks below on real public documents against an answer key, and we publish the misses. Pick a check to see its numbers.
We ran the tender pack breakdown on 12 public Australian tender packs (521 pages) and checked every row against an answer key. Here is what it found, what it missed, and exactly how we measured it.
Last run 6 October 2026 · 12 packs from NSW, SA, TAS, VIC, WA · measured with GPT-6.1 Sol
The bar we set before publishing: find at least 90% of must-haves with at least 95% of quotes correct. This run meets it.
We also gave the same AI model each whole pack in a single request with a one-line instruction: “list everything a bidder must do or provide, with page numbers and quotes”. That is roughly what you get by pasting a tender into a chatbot. Both were scored the same way.
| Measure | FileAI | Single prompt |
|---|---|---|
| Must-haves found before you submit | 98% 731 of 746 | 91.8% 685 of 746 |
| Contract obligations found if you win | 98.2% 380 of 387 | 93.5% 362 of 387 |
| Quotes found on the cited page | 99.7% 3,386 of 3,395 | 99.8% 1,973 of 1,977 |
| Rows that are real requirements estimate, 95% range | 99% range 94.5–99.8% | 95.9% range 90.4–98.4% |
| Rows per pack | 282.9 | 164.8 |
| Pages per pack | 43.4 | 43.4 |
| Minutes per pack | 2.6 | 3.3 |
The single prompt found most of the forms and documents to return, but only half of the evaluation criteria and about four in five key dates, and it listed a little over half as many rows. FileAI reads each pack in sections of about eight pages with written rules for every type of requirement, then merges the rows and checks every quote against its page.
Must-haves found by type. The misses are listed honestly below, so you know what to double-check yourself.
| Type | FileAI | Single prompt |
|---|---|---|
| Condition for participation | 100% 25 of 25 | 100% 25 of 25 |
| Mandatory requirement | 100% 77 of 77 | 98.7% 76 of 77 |
| Returnable document | 97.7% 338 of 346 | 99.4% 344 of 346 |
| Insurance | 95.7% 22 of 23 | 100% 23 of 23 |
| Licence or certification | 94.7% 18 of 19 | 100% 19 of 19 |
| Key date | 100% 28 of 28 | 78.6% 22 of 28 |
| Evaluation criterion | 98.7% 76 of 77 | 49.4% 38 of 77 |
| Submission rule | 97.4% 147 of 151 | 91.4% 138 of 151 |
| Type | FileAI | Single prompt |
|---|---|---|
| Condition for participation | 100% 1 of 1 | 100% 1 of 1 |
| Mandatory requirement | 97.8% 270 of 276 | 96% 265 of 276 |
| Insurance | 98% 49 of 50 | 98% 49 of 50 |
| Licence or certification | 100% 20 of 20 | 100% 20 of 20 |
| Key date | 100% 40 of 40 | 67.5% 27 of 40 |
When a schedule lists dozens of rates, FileAI can summarise part of it and drop a line, such as a site establishment rate, one trade in a list of four, or one size of pavement numeral. Check the schedule itself before you price.
Where one sentence asks for two things, like signing the offer and answering every selection criterion, or giving certificate details and attaching a copy, the second part is occasionally lost.
One pack listed what it would consider in two places with different wording. FileAI captured the weighted table and only part of the second, unweighted list.
Cleaning checklists and similar schedules get one row per area, not per task, so a once-a-year task is only in the schedule itself. The other contract misses were a payment reduction scale, a no-claims clause, an intellectual property term and part of an auditor's reporting duty.
Scanned pages, drawings and tables saved as images have no text to read, so they are outside these numbers. FileAI names them as pages it couldn't read so you can check them yourself.
12 real tender and quotation packs (14 PDF files, 521 pages) published on official council and state agency websites in NSW, SA, TAS, VIC, WA. They cover cleaning, IT, building and civil works, planning and audit consultancies, marketing and a supply panel, from 9 to 111 pages each.
The packs are used for testing only. We don't republish them, so they are described by state, buyer and work type, not by name.
WA local government · Facilities / cleaning services
79 pages · 92 items in the answer key
FileAI found 97.8%, single prompt 94.6%
WA local government · Professional services
51 pages · 56 items in the answer key
FileAI found 96.4%, single prompt 100%
WA local government · IT managed services / cybersecurity
23 pages · 81 items in the answer key
FileAI found 97.5%, single prompt 91.4%
WA local government · Construction
44 pages · 66 items in the answer key
FileAI found 98.5%, single prompt 93.9%
WA local government · Professional services
43 pages · 85 items in the answer key
FileAI found 98.8%, single prompt 95.3%
TAS local government · Professional services
41 pages · 33 items in the answer key
FileAI found 100%, single prompt 69.7%
SA local government · Professional services
22 pages · 57 items in the answer key
FileAI found 98.3%, single prompt 87.7%
NSW local government · Construction
30 pages · 28 items in the answer key
FileAI found 96.4%, single prompt 82.1%
NSW local government · Council works and services panel
43 pages · 50 items in the answer key
FileAI found 94%, single prompt 88%
NSW local government · Civil services
9 pages · 35 items in the answer key
FileAI found 97.1%, single prompt 94.3%
VIC state agency · Construction
111 pages · 86 items in the answer key
FileAI found 98.8%, single prompt 89.5%
VIC state statutory body · Marketing / communications services
25 pages · 77 items in the answer key
FileAI found 100%, single prompt 97.4%
| Document set | Pages | In the key | FileAI found | Single prompt |
|---|---|---|---|---|
| WA local government · Facilities / cleaning services | 79 | 92 | 97.8% | 94.6% |
| WA local government · Professional services | 51 | 56 | 96.4% | 100% |
| WA local government · IT managed services / cybersecurity | 23 | 81 | 97.5% | 91.4% |
| WA local government · Construction | 44 | 66 | 98.5% | 93.9% |
| WA local government · Professional services | 43 | 85 | 98.8% | 95.3% |
| TAS local government · Professional services | 41 | 33 | 100% | 69.7% |
| SA local government · Professional services | 22 | 57 | 98.3% | 87.7% |
| NSW local government · Construction | 30 | 28 | 96.4% | 82.1% |
| NSW local government · Council works and services panel | 43 | 50 | 94% | 88% |
| NSW local government · Civil services | 9 | 35 | 97.1% | 94.3% |
| VIC state agency · Construction | 111 | 86 | 98.8% | 89.5% |
| VIC state statutory body · Marketing / communications services | 25 | 77 | 100% | 97.4% |
An AI review pass (a different AI model from the one FileAI uses) read every page and listed every obligation: 1,133 in all. Each entry's quote was then checked by a program against the text of its page.
No person has reviewed the whole key. It will have its own misses and judgement calls, so treat these numbers as a careful estimate, not a guarantee.
Must-haves before you submit (746) were listed exhaustively. Contract obligations if you win (387) were listed for the material items only, such as insurances, licences, plans and key dates.
A row has to cite the same file and a page within one page of the answer key's page, and say the same thing. A grading model decides the meaning with strict rules: a related but different item, a row that leaves out the deadline or amount, or a different number does not count.
We checked the grader by hand: on 40 decisions we reviewed ourselves, it agreed with us 37 times (92.5%).
An earlier version of the grading rules was too strict: on 40 other decisions it agreed with us only 28 times, always by marking a found item as missed. We fixed the rules, then ran the check above on new decisions.
The quote must appear on the page it cites, word for word, ignoring only spacing, capital letters and the style of quote marks and dashes. A quote may start on the cited page and run onto the next. Rows from tables count when every word of the quote sits together in that part of the page (21 of FileAI's 3,395 rows). This check is separate from the one FileAI runs on itself.
FileAI lists more rows than the key, because the key is selective: it holds the material items, not every line. To check those extra rows are real, we graded a random sample of 50 of them against their page.
Time is the AI processing time for a pack, with sections read in parallel the way the service runs them. It leaves out waiting in the queue and sending the email, and the person who checks every row, so a breakdown takes longer than this to reach you: after you upload, the free preview is usually ready within 15 minutes, longer for big packs, and the checked full breakdown usually within one business day.
The grader read 50 sampled FileAI rows that matched nothing in the answer key and found 49 of them to be real requirements. The one it rejected turns something the buyer may do into a task for you. We hand-checked the grader on an earlier run (see above), not on this sample.
While volumes are low, a person reviews every row of a paid breakdown against its page before it is delivered. Every row still links to its page so you can check it yourself.
We ran the RFP compliance matrix on 12 real public US solicitations (952 pages) from federal, state and local buyers, and checked every row against an answer key written from the documents themselves. Here is what it found, what it missed, and exactly how we measured it.
Last run 6 October 2026 · 12 solicitations: CA, Federal, MO, MS, NY, TN, TX · measured with GPT-6.1 Sol
The bar we set before publishing: find at least 90% of proposal requirements with at least 95% of quotes correct. This run meets it.
We also gave the same AI model each whole solicitation in a single request with a one-line instruction: “list everything an offeror must do or provide, with page numbers and quotes”. That is roughly what you get by pasting an RFP into a chatbot. Both were scored the same way.
| Measure | FileAI | Single prompt |
|---|---|---|
| Proposal requirements found for your proposal | 96.4% 1,603 of 1,663 | 68.5% 1,140 of 1,663 |
| Contract terms found if you win | 95.5% 621 of 650 | 64.5% 419 of 650 |
| Quotes found on the cited page | 99.1% 7,551 of 7,623 | 99.7% 3,297 of 3,308 |
| Rows that are real requirements estimate, 95% range | 91.3% range 84.5–95.5% | 98.2% range 93.9–99.5% |
| Rows per RFP | 635.3 | 275.7 |
| Pages per RFP | 79.3 | 79.3 |
| Minutes per RFP | 4.9 | 4.7 |
Proposal requirements found by type. The misses are listed honestly below, so you know what to double-check yourself.
| Type | FileAI | Single prompt |
|---|---|---|
| Submission instruction | 97.2% 411 of 423 | 79.9% 338 of 423 |
| Evaluation factor | 96.6% 228 of 236 | 11.9% 28 of 236 |
| Required form or certification | 94.5% 191 of 202 | 78.7% 159 of 202 |
| Eligibility and registration | 96% 24 of 25 | 72% 18 of 25 |
| Key date | 95.1% 117 of 123 | 52.8% 65 of 123 |
| Mandatory requirement | 97.5% 426 of 437 | 85.8% 375 of 437 |
| Past performance or personnel requirement | 97% 97 of 100 | 85% 85 of 100 |
| Pricing instruction | 93.2% 109 of 117 | 61.5% 72 of 117 |
| Type | FileAI | Single prompt |
|---|---|---|
| Key date | 75% 3 of 4 | 75% 3 of 4 |
| Mandatory requirement | 100% 6 of 6 | 100% 6 of 6 |
| Pricing instruction | 100% 1 of 1 | 100% 1 of 1 |
| Clause of note | 95.3% 407 of 427 | 60.9% 260 of 427 |
| Insurance or bonding requirement | 96.4% 106 of 110 | 80.9% 89 of 110 |
| Reporting or deliverable | 96.1% 49 of 51 | 72.5% 37 of 51 |
| Payment and invoicing | 96.1% 49 of 51 | 45.1% 23 of 51 |
When an amendment replaces a price workbook, adds a line item or repeats a revised calendar, FileAI lists the new deadline but can leave out what else the amendment changed. Read every amendment's cover page against your matrix.
Federal and grant-funded contracts (FTA, CDBG) list dozens of clauses. FileAI gives each clause that carries a duty its own row, but can still drop one, such as a change-order rule or a facility security requirement. Check the clause list yourself before you sign.
Section K representations that appear only as a clause number and title (for example covered telecommunications equipment, or delinquent tax and felony convictions) are sometimes missed. Complete Section K from the solicitation itself.
Instructions inside a pricing schedule, such as how option periods or an extension period are priced, or milestone payment percentages, are the proposal group FileAI misses most often after evaluation factors. Read the price instructions and the workbook together.
Where a solicitation scores in phases with pass/fail gates, a gate's exact criteria (minimum project values, what counts as similar work) can be shortened or missed. Treat every go/no-go statement as a must-have.
12 real public solicitations (39 PDF files, 952 pages) published on official federal, state and local government procurement sites. They cover services and construction, from 28 to 141 pages each.
The solicitations are used for testing only. We don't republish them, so they are described by buyer level and work type, not by name.
We improved FileAI's instructions while looking at 6 of the 12 sets (marked in the table). The other 6 were not used to make those changes and were first scored in the final run, so they show how it does on documents it was not tuned on.
| Sets | Proposal requirements found | Contract terms found |
|---|---|---|
| Used for tuning 6 sets, 310 pages | 97.9% 546 of 558 | 97.4% 294 of 302 |
| Held out 6 sets, 642 pages | 95.7% 1,057 of 1,105 | 94% 327 of 348 |
CA city · Construction (renovation) (used for tuning)
52 pages · 73 items in the answer key
FileAI found 97.3%, single prompt 83.6%
Federal agency · Medical services
125 pages · 172 items in the answer key
FileAI found 100%, single prompt 66.3%
Federal agency · Transport services (RFQ) (used for tuning)
28 pages · 63 items in the answer key
FileAI found 98.4%, single prompt 73%
Federal agency · Research services
87 pages · 218 items in the answer key
FileAI found 93.1%, single prompt 59.6%
Federal agency · Healthcare services
136 pages · 232 items in the answer key
FileAI found 95.3%, single prompt 73.7%
Federal agency · Construction (used for tuning)
45 pages · 77 items in the answer key
FileAI found 98.7%, single prompt 0%
Federal agency · Design-build construction (used for tuning)
83 pages · 185 items in the answer key
FileAI found 97.3%, single prompt 66%
MO state agency · Construction (used for tuning)
65 pages · 71 items in the answer key
FileAI found 97.2%, single prompt 63.4%
MS state agency · IT services
141 pages · 179 items in the answer key
FileAI found 96.1%, single prompt 84.9%
NY state agency · Training services
74 pages · 180 items in the answer key
FileAI found 95.6%, single prompt 70%
TN public authority · Professional services
79 pages · 124 items in the answer key
FileAI found 94.3%, single prompt 79.8%
TX city · Planning consulting (used for tuning)
37 pages · 89 items in the answer key
FileAI found 98.9%, single prompt 83.2%
| Document set | Pages | In the key | FileAI found | Single prompt |
|---|---|---|---|---|
| CA city · Construction (renovation) used for tuning | 52 | 73 | 97.3% | 83.6% |
| Federal agency · Medical services | 125 | 172 | 100% | 66.3% |
| Federal agency · Transport services (RFQ) used for tuning | 28 | 63 | 98.4% | 73% |
| Federal agency · Research services | 87 | 218 | 93.1% | 59.6% |
| Federal agency · Healthcare services | 136 | 232 | 95.3% | 73.7% |
| Federal agency · Construction used for tuning | 45 | 77 | 98.7% | 0% |
| Federal agency · Design-build construction used for tuning | 83 | 185 | 97.3% | 66% |
| MO state agency · Construction used for tuning | 65 | 71 | 97.2% | 63.4% |
| MS state agency · IT services | 141 | 179 | 96.1% | 84.9% |
| NY state agency · Training services | 74 | 180 | 95.6% | 70% |
| TN public authority · Professional services | 79 | 124 | 94.3% | 79.8% |
| TX city · Planning consulting used for tuning | 37 | 89 | 98.9% | 83.2% |
The answer key was written by reading every solicitation page by page and listing what a proposal manager's compliance matrix must contain: 2,313 rows in all. It was written before FileAI was run on the documents, by an AI assistant following written rules, not by FileAI. A program then checked that every quote is on the page it cites.
No procurement professional has reviewed the key. It will have its own misses and judgement calls, so treat these numbers as a careful estimate, not a guarantee.
Proposal requirements (1,663) were listed exhaustively: submission instructions, evaluation factors, required forms and certifications, key dates, mandatory requirements of the work, past performance and key personnel, pricing instructions. Contract terms if you win (650) were listed for the material items only, such as clauses of note, insurance and bonding.
A row has to cite the same file and a page within one page of the answer key's page, and say the same thing. A grading model decides the meaning with strict rules: a related but different item, a row that leaves out the deadline or amount, or a different number does not count.
We have not yet checked the grader by hand on RFP rows. It uses the same strict rules as the tender grader, which agreed with our own labels on 37 of 40 tender decisions, so treat the RFP recall as a careful estimate. Nine answer-key rows got no verdict from the grader and are counted as missed.
The quote must appear on the page it cites, word for word, ignoring only spacing, capital letters and the style of quote marks and dashes. A quote may start on the cited page and run onto the next. Rows from tables count when every word of the quote sits together in that part of the page (283 of FileAI's 7,623 rows). This check is separate from the one FileAI runs on itself.
FileAI lists more rows than the key, because the key is selective: it holds the material items, not every line. To check those extra rows are real, we graded a random sample of 50 of them against their page.
Time is the AI processing time for a RFP, with sections read in parallel the way the service runs them. It leaves out waiting in the queue and sending the email, and the person who checks every row, so a breakdown takes longer than this to reach you: after you upload, the free preview is usually ready within 15 minutes, longer for big RFPs, and the checked full breakdown usually within one business day.
We graded a random sample of 50 of the rows that matched nothing in the answer key: 42 were real requirements, 5 described something the buyer does rather than a task for you, and 3 quoted text the grader could not find on the cited page.
While volumes are low, a person reviews every row of a paid breakdown against its page before it is delivered. Every row still links to its page so you can check it yourself.
We ran the HOA document check on 6 real sets of public HOA and condominium association documents (632 pages) and checked every row against an answer key written from the documents themselves. Here is what it found, what it missed, and exactly how we measured it.
Last run 6 October 2026 · 6 document sets: AZ, CA, DE, FL, VA, WA · measured with GPT-6.1 Sol
The bar we set before publishing: find at least 90% of findings with at least 95% of quotes correct. This run meets it.
We also gave the same AI model each whole set in a single request with a one-line instruction: “list everything a home buyer should know before closing, with page numbers and quotes”. That is roughly what you get by pasting an HOA package into a chatbot. Both were scored the same way.
| Measure | FileAI | Single prompt |
|---|---|---|
| Findings found in the association's documents | 91% 193 of 212 | 80.2% 170 of 212 |
| Quotes found on the cited page | 100% 2,204 of 2,204 | 99.9% 940 of 941 |
| Rows that are real findings estimate, 95% range | 91% range 80.9–96.1% | 67% range 55.6–77.2% |
| Rows per set | 367.3 | 156.8 |
| Pages per set | 105.3 | 105.3 |
| Minutes per set | 3.4 | 3.2 |
The single prompt found about eight in ten of the findings and listed under half as many rows, and a third of the rows we sampled from it were wrong or not findings. It did well on insurance, rules and special assessments, and worst on reserve figures and repair schedules. FileAI reads each set in sections of about eight pages with written rules for every type of finding, then merges the rows and checks every quote against its page.
Findings found by type. The misses are listed honestly below, so you know what to double-check yourself.
| Type | FileAI | Single prompt |
|---|---|---|
| Dues and fees | 93.8% 15 of 16 | 93.8% 15 of 16 |
| Special assessment | 100% 6 of 6 | 100% 6 of 6 |
| Reserves and budget | 84.3% 43 of 51 | 66.7% 34 of 51 |
| Litigation and disputes | 100% 6 of 6 | 100% 6 of 6 |
| Insurance | 94.7% 18 of 19 | 89.5% 17 of 19 |
| Rule or restriction | 93.5% 43 of 46 | 93.5% 43 of 46 |
| Major repair or project | 88.5% 46 of 52 | 65.4% 34 of 52 |
| Delinquency and red flag | 100% 14 of 14 | 92.9% 13 of 14 |
| Key date | 100% 2 of 2 | 100% 2 of 2 |
A reserve study's own date or inspection date is sometimes not listed on its own, so a study that is six years old may not be flagged in the rows. Check the date on the first page of every study.
A one-line ban, such as on antennas, can be left out of a long set of rules, while the larger rules (rentals, pets, parking, fines) are found. Read the rules yourself for anything you plan to do.
A figure like a 40-year replacement total or a payment spread over several years can lose its period in the summary row. Open the page the finding links to.
Some balance sheet lines (such as deferred assessments) are summarised under another line. The headline reserve numbers, percent funded and special assessments are found reliably.
6 sets of real public documents (23 PDF files, 632 pages) that homeowners and condominium associations publish on their own websites, in AZ, CA, DE, FL, VA, WA. They are budgets, reserve studies, governing documents, board minutes, insurance certificates and policies, from 63 to 166 pages each.
The documents are used for testing only. We don't republish them, so they are described by state and type of community, not by name.
AZ villa community · Minutes, insurance and reserve study
63 pages · 27 items in the answer key
FileAI found 88.9%, single prompt 88.9%
CA planned community · Budget package, fines and insurance
89 pages · 31 items in the answer key
FileAI found 90.3%, single prompt 93.5%
DE townhome community · Minutes, audit and reserve study
78 pages · 38 items in the answer key
FileAI found 86.8%, single prompt 68.4%
FL condominium · Structural integrity reserve study and declaration
166 pages · 24 items in the answer key
FileAI found 83.3%, single prompt 79.2%
VA master association · Budget, covenants and reserve study
95 pages · 38 items in the answer key
FileAI found 94.7%, single prompt 76.3%
WA high-rise condominium · Minutes, rules and reserve study
141 pages · 54 items in the answer key
FileAI found 96.3%, single prompt 79.6%
| Document set | Pages | In the key | FileAI found | Single prompt |
|---|---|---|---|---|
| AZ villa community · Minutes, insurance and reserve study | 63 | 27 | 88.9% | 88.9% |
| CA planned community · Budget package, fines and insurance | 89 | 31 | 90.3% | 93.5% |
| DE townhome community · Minutes, audit and reserve study | 78 | 38 | 86.8% | 68.4% |
| FL condominium · Structural integrity reserve study and declaration | 166 | 24 | 83.3% | 79.2% |
| VA master association · Budget, covenants and reserve study | 95 | 38 | 94.7% | 76.3% |
| WA high-rise condominium · Minutes, rules and reserve study | 141 | 54 | 96.3% | 79.6% |
The answer key was written by reading the documents page by page and listing what a buyer should know: 212 rows in all. It was written before FileAI was run on the documents, by an AI assistant following written rules, not by FileAI. A program then checked that every quote is on the page it cites.
No real estate professional has reviewed the key. It covers the material findings, not every line of every document, and it will have its own misses and judgement calls, so treat these numbers as a careful estimate, not a guarantee.
Most of the findings are about the association as a whole (212). Findings about one home (0) are rare because these public documents are not about a particular home; a resale certificate is.
A row has to cite the same file and a page within one page of the answer key's page, and say the same thing. A grading model decides the meaning with strict rules: a related but different item, a row that leaves out the deadline or amount, or a different number does not count.
The grader is an AI model with a strict rubric, and we did not hand-label its decisions for this kind. It is sometimes too strict: among the nineteen misses below, FileAI's rows do state at least one of them (a Florida condominium's 0.34% percent-funded figure), so the real recall is a little higher than shown.
The quote must appear on the page it cites, word for word, ignoring only spacing, capital letters and the style of quote marks and dashes. A quote may start on the cited page and run onto the next. Rows from tables count when every word of the quote sits together in that part of the page. This check is separate from the one FileAI runs on itself.
FileAI lists more rows than the key, because the key is selective: it holds the material items, not every line. To check those extra rows are real, we graded a random sample of 50 of them against their page.
Time is the AI processing time for a set, with sections read in parallel the way the service runs them. It leaves out waiting in the queue and sending the email, and the person who checks every row, so a breakdown takes longer than this to reach you: after you upload, the free preview is usually ready in 5 to 15 minutes, and the checked full breakdown usually within one business day.
Five of the fifty sampled rows that matched nothing in the answer key were not findings a buyer needs (for example a general procedure) or were slightly off. FileAI lists everything it can support with a quote, so expect more rows than a person would write; the preview and the ratings lead with the ones that cost money.
While volumes are low, a person reviews every row of a paid breakdown against its page before it is delivered. Every row still links to its page so you can check it yourself.