A digitization pilot project runs a small, deliberately chosen batch of your real originals through every stage of the planned work, from handling and capture to text recognition, cataloging and a test import into your own system, so you can measure results before you commit a collection or a budget. It should include your hardest items as well as typical ones, and it should be judged against pass criteria you wrote down before the first folder moved. A good pilot ends in one of three decisions: proceed as planned, proceed with named changes, or stop.
A disclosure first. Dastok sells digitization and offers a pilot box as a first step, described on our engagements page. The method below does not depend on us. It works the same way with any vendor, with a vendor and your volunteers together, or with your own staff and a copy stand in the back room.
The pilot is one step in a longer sequence of decisions, set out in our overview of what digitizing a collection involves. This article covers the pilot itself: criteria, sample, measures and how to read the results.
What a digitization pilot project is for
A digitization pilot project replaces guesses with measurements taken on your own collection. It tells you whether the planned handling is safe for your originals, whether the outputs meet the written specification, how long each stage takes per item, and whether the records and files work in the system where they will live.
The idea is old and well argued. In the Northeast Document Conservation Center’s Handbook for Digital Projects (2000), Stephen Chapman of Harvard University Library wrote that “empirical evidence gathered from one’s own collections is more convincing than anecdotal reports from other projects.” For outsourced work he went further: the project budget “should be finalized only after a sample has been put through an entire workflow,” which he listed as scanning, processing, metadata creation including full text, and quality control, with the results “inspected and approved by the appropriate stakeholders in the project.”
A pilot is not the same as a sample scan
Several kinds of small test get called a pilot. They answer different questions, and confusing them is how institutions end up with a nice demonstration and no numbers.
| Test | Question it answers | Who usually runs it | What you get |
|---|---|---|---|
| Proof of concept | Can this method work on this material at all? | You or a vendor, a few items | A yes, no or maybe on one technique |
| Sample scans with a proposal | Can this vendor produce acceptable images? | Each bidder, on items you send | Images to compare between bidders |
| Pilot | What happens to our material, at every stage, at what cost? | The team that will do the work | Measured quality, error rates, times and a tested import |
| Batch acceptance in production | Does this batch meet the agreed specification? | You, on a random sample of each batch | Accept, or re-inspect and redo |
A proof of concept might show that handwriting recognition can read one writer’s hand. Only a pilot tells you what that costs once the names are corrected and the records are loaded.
Why the pilot must cover more than the scanner
Capture is the most visible stage, so it is tempting to pilot only capture. The Federal Agencies Digital Guidelines Initiative’s project planning outline (2009) offers a rough rule that shows why that is a mistake: on average, one third of the effort goes to “project planning, preservation preparation, management, and oversight,” one third to “archival description and indexing,” and one third to “the actual digitization.” The outline adds that the actual percentages “will vary depending on the project.” A pilot that times only the scanner measures about a third of the work.
Pilot, then production: a pattern from mass digitization
Large programs use the same pattern. The Smithsonian’s Digitization Program Office publishes a dashboard, Osprey, that it describes as “quality control and progress monitoring for Smithsonian digitization projects,” built to help “vendors, project managers, and museum units monitor progress and catch problems early.” Its list of Mass Digitization projects shows several pilots lasting from about a week to two months, each followed months later by a larger project with a matching name.
| Project (unit code) | Pilot dates | Pilot objects | Later project dates | Later project objects |
|---|---|---|---|---|
| Sears Collection (SG) | Aug 16 to 23, 2013 | 877 | Jan 12 to 28, 2015 | 3,394 |
| Numismatics (NMAH) | Feb 21 to 28, 2014 | 3,069 | Oct 20, 2014 to Mar 6, 2015 | 271,301 |
| Meserve (NPG) | Dec 4 to 8, 2017 | 1,900 (estimated) | Apr 9 to 13, 2018 | 3,853 |
| JPC Archive (NMAAHC) | Aug 16 to Oct 10, 2023 | 9,409 | Priority 1, Feb 26 to Oct 30, 2024; Production 1A, Nov 4, 2024 to Jun 3, 2026; Production 1B, from Mar 2, 2026 (ongoing) | 71,185; 379,147; 372,829 so far |
Most of the pilots took about a week, and the larger projects started months later. There is no fixed ratio: the numismatics and JPC Archive pilots were each about 1 percent of what followed, the Meserve pilot about half. The dashboard lists dates and counts only, not what each pilot tested or changed, so read it for the shape of the method.
Write the pass criteria before the first folder moves
Decide in writing, before the pilot starts, what you will measure, how you will measure it, what result counts as a pass, who judges it and what happens if it fails. Criteria written after the results arrive tend to bend toward whatever the pilot produced.
The standards agree. Chapter 10 of the FADGI Technical Guidelines (third edition, 2023) says quality plans “should address accuracy requirements for and acceptable error rates for all aspects evaluated.” In the NEDCC handbook, Janet Gertz of Columbia University Libraries advises institutions to “write up criteria and benchmarks for evaluation of image quality and metadata accuracy” and to “decide who will make the evaluation.” The LC Labs AI Planning Framework at the Library of Congress includes a worksheet to “articulate success criteria, measures, risks, and benefits” for an AI use case.
Three kinds of criteria
- Hard limits. No new damage. Every item returned in its original order. Every page and side that carries information captured. Restricted items handled under your rules. One failure stops the pilot until the cause is fixed.
- Tolerances. Results measured against a number: target measurements within the agreed FADGI star level, the share of images recaptured, character error rate per kind of page, accuracy of each required catalog field.
- Planning measurements. Images per hour, minutes per record, hours of your own staff time. They have no pass mark, but they set the budget. Write down the rate you would have assumed, so the gap is visible afterward.
Set thresholds from use, not from a brochure
A threshold should follow from what people will do with the output. Eileen Gifford Fenton, writing in the NEDCC handbook about OCR at JSTOR, put it this way: “If you are using the text files only to support searching, and they will not be displayed to the user, you may be able to tolerate lower accuracy.” Transkribus, a handwriting recognition platform, gives a similar rule of thumb in its guidance on character error rate: a rate of 10 percent or below “can be considered very efficient for automated transcription,” while 20 to 30 percent is “sufficient to work with powerful searching tools.” Those are a vendor’s own figures, not a standard, but the logic holds: text that only feeds search can pass at a looser threshold than a transcription people will read.
A criteria sheet you can adapt
The criteria in the fourth column are examples to adapt to your collection and purpose, not standards.
| Area | Measure | How it is measured | Example criterion (set your own) | Judged by |
|---|---|---|---|---|
| Handling | New damage; order on return | Condition notes and photos before and after; folder list check | No new damage; every item back in order | Collections manager |
| Capture | Target results per session | Target analysis report for each session | Every session within the agreed star level | Project lead |
| Capture | Images failing visual review | Every image checked at 100 percent | Recapture rate at or under an agreed share, no repeated cause | Project lead |
| Completeness | Pages and sides captured | Count against the inventory | 100 percent | Registrar or volunteer |
| Text | Character error rate by kind of page | Reference transcriptions | A rate per kind of page, before and after correction | Archivist |
| Description | Required fields; machine suggestions | Review of every pilot record | All required fields present; machine values labeled | Cataloger |
| Import | Records loaded and linked to files | Test import into your system | Every record loads and opens the right images | Database lead |
| Time | Minutes per item for each stage | Logs kept during the pilot | Recorded; no pass mark | Project lead |
Have the person who will approve the full program sign the sheet before work starts.
Choose a sample that includes your worst items
A pilot sample should cover every format, every major hand and every condition problem you expect to meet, with your hardest items included on purpose. Add a small random draw as well, so your typical rates are not distorted by hand-picked items.
Gertz’s chapter in the NEDCC handbook asks vendors answering a request for proposals to “scan a representative sample that represents a fair cross section of the materials, including both easy and difficult items.” She also describes what happens when the sample is too easy: “if the vendor bids on the basis of a sample of uniformly legible and nonbrittle materials, inclusion of a significant amount of low contrast or brittle items can seriously affect the anticipated workflow and productivity.” A pilot of clean typescript tells you nothing about the faded pencil letters in the next box.
Build the sample in two parts
The purposive part is chosen from your inventory to cover each kind of difficulty:
- Format: loose letters, bound volumes, prints, negatives, oversize sheets, newsprint, scrapbooks.
- Condition: brittle edges, tight or broken bindings, folds, faded pencil, tape.
- Content: each major hand, each language, forms filled in by hand, columns and tables.
- Sensitive items: one or two restricted items, to test how restrictions carry through to the records.
The random part is a few folders drawn by random number from the inventory, to estimate typical rates. Report the two parts separately. Error rates from the worst hand alone will overstate correction costs; rates from a random draw alone may miss the material that fails.
How big should the sample be?
Coverage matters more than volume. Each kind of difficulty needs enough items to show a pattern, and each major hand needs a few pages of reference text. The sample should still be small enough to finish and review in a few weeks: for many collections, one or two boxes or a few hundred images.
Chapman suggested that for forecasting cost, “in many cases, a half dozen items will be sufficient.” That is fair for a time estimate on uniform material. Error rates are different: as the section on reading results shows, a few dozen images cannot tell a 1 percent defect rate from a 10 percent one.
When the hardest item is too precious to risk
Gertz added a caution: where originals are valuable or fragile, a proposal sample “should consist of reasonably similar items that are less valuable or are expendable.” In a pilot under contract, with agreed handling and custody, you may include the difficult original itself. That is a judgment for your conservator or collections manager, recorded in the pilot plan.
Example (fictional). The Aster County Historical Society holds about 60 boxes of family papers plus a shelf of store ledgers. It chose Box 3, the Okerlund family papers and the Wren Gulch Mercantile day book, 1893 to 1896, as its pilot because the box held nearly every difficulty in the collection: 41 letters in two hands, seven pages in pencil, 18 cabinet cards with pencil notes on the backs and gilt mount edges, a 212-page day book with a split spine that opens safely to about 110 degrees, brittle newspaper clippings and a telegram. That is 69 items and 410 images. A gap the society noticed later: every item in the pilot was chosen by hand, so there was no random part. It treated the pilot’s rates for letters as cautious and added a random check to the first production batch. The whole box is followed stage by stage on our example collection page.
Handling and capture: what the originals and images must show
The pilot is where handling methods and the capture specification meet real originals. Check that every item came back unharmed and in order, that target measurements met the agreed level in every session, and that every image passes a full-size visual check.
Handling and custody
Chapman’s chapter on printed text and manuscripts is blunt: “Decisions regarding appropriate handling are complex, and any method must be tested and confirmed with a sample of materials before undertaking full production.” A pilot is that test. Record, for every item:
- Condition before and after, with photographs of existing damage, using a list such as our condition assessment checklist.
- Packing and transport, with a signed folder-level list at each handover, as in our packing and transport checklist.
- The handling method for each format: how far bound volumes were opened and on what support, whether folded letters were relaxed.
- Any incident, however small, and what was done about it.
The same record applies when the work happens on site. A pilot that ends with “nothing happened” and no notes is weaker evidence than a page of notes showing nothing went wrong.
Capture against the written specification
Write the capture specification for each kind of original before the pilot: the FADGI material category, the star level, the sampling rate and the file format. A FADGI star level is not a scanner setting. The guidelines say a digital object’s star rating “is determined by meeting the FADGI evaluation parameters through image analysis,” which means capturing a reference target and measuring it. Our guide to FADGI star ratings explains the levels and what a target report should contain. Ask for that report for every capture session in the pilot.
Federal rules offer a useful model, though they bind federal agencies, not your society or museum. The National Archives regulation for digitizing permanent federal records, 36 CFR Part 1236 Subpart E (2023), requires agencies to “evaluate internal or external vendor imaging systems against image quality performance parameters” and says an agency “cannot rely solely on equipment specifications, such as scanner ppi settings or camera sensor megapixels.” For a smaller institution, the pilot is that evaluation.
Inspect every image in the pilot
In production, FADGI recommends inspecting “at a minimum, 10 images or 10 % of each batch of digital images, whichever quantity is larger.” A pilot is small enough to look at every image at full size. Check:
- Completeness and order. FADGI asks you to verify “that 100% of the source materials have been captured and accounted for” and kept in the original order.
- Image defects. Focus, cropping, glare, color cast, skew, dust, soft gutters.
- Files. Format, bit depth, color profile and file names, checked with a validation tool such as JHOVE, which FADGI recommends.
- Targets. Each session’s target analysis within the agreed tolerances.
Record every defect with its file name, its type and its cause. Causes matter more than counts. A defect from a setting is fixed once; a defect from the material, such as a binding that will not open, recurs on every similar item. Our image QC checklist lists what to look at.
Example (fictional). Every one of the 410 images from Box 3 was reviewed. Nine were recaptured, 2.2 percent: four cabinet cards with glare on gilt mount edges, three day book pages soft in the gutter, and two envelopes with the flap cropped. All six session targets measured within the agreed three-star tolerances. The glare and the cropping had one-time causes and were fixed in the next session. The soft gutters came from a binding that opens only to about 110 degrees, so they would recur on every ledger. The production plan gave bound volumes their own setup and a slower rate, rather than more inspection.
Text recognition: measure error rates on your own pages
Measure the character error rate on reference transcriptions of your own pages, separately for each kind of page and each major hand, before and after correction. Published accuracy figures describe someone else’s documents and someone else’s model.
Character error rate (CER) is the number of substituted, missed and extra characters in the machine text, divided by the number of characters in a reference transcription made by a person. Word error rate (WER) does the same for whole words, and it is the number that matters most for search, because one wrong letter hides a whole word. Transkribus defines CER as “the percentage of characters that have been transcribed incorrectly,” measured against “your Ground Truth.”
How to measure it in a pilot
- Choose reference pages from each group: each major hand, each printed source, each ledger. Use typical pages, not the cleanest.
- Transcribe them exactly, by a person, with written rules for abbreviations, spelling and illegible words.
- Run recognition with the engine and settings planned for production, and compare.
- Classify the errors that matter to users: names, places, dates and numbers.
- Correct to the planned level on other pages, measure again on fresh pages, and time the correction.
The OCR accuracy calculator turns error counts into CER, WER and correction hours for a batch. Time the corrections in the pilot rather than assuming a rate.
Use published ranges as context only
Transkribus’s guidance says that “a good CER for printed text is between 0.5 and 2%, while for handwritten text, it ranges between 2 and 8%, depending if the model has been trained on one or multiple hands,” and that a model may “perform worse” on hands not seen in training. Fenton’s warning from the NEDCC handbook still applies: “Initially selected software may not perform on actual data as it did on test data. You may find processing limitations in the full production phase that were hidden during the pilot phase.” Check the first production batches against the pilot’s numbers.
Example (fictional). Box 3’s reference pages gave these results before correction.
| Material | Reference sample | CER | WER |
|---|---|---|---|
| Ida Okerlund’s letters | 6 pages | 3.96% | 11.57% |
| Walter Okerlund’s letters | 3 pages | 7.84% | 17.91% |
| Day book | 3 pages | 10.00% | 27.50% |
| Clippings (print) | 2 clippings | 0.61% | 2.59% |
In Ida’s uncorrected text the model read “Wren Gulch” as “Wren Gutch” eleven times, which would have hidden those letters from a search for the town. The society chose targeted correction of names, places, dates and low-confidence lines for Ida’s letters, full correction for Walter’s letters and the day book, and uncorrected, labeled text for the clippings. Fresh samples after correction measured 1.12 percent for Ida’s letters, with no errors left in names, places or dates, and 0.38 percent for the day book.
Catalog records and machine suggestions: measure field by field
Review every pilot record and count, for each field, how many machine suggestions were accepted, edited or rejected, and how long the review took per item. Check that every required field is present and that machine-made values stay labeled until a person has reviewed them.
Metadata quality is harder to score than image quality. FADGI’s chapter on quality management admits that “there are no clearly defined metrics for evaluating metadata quality” and says assessment “will likely require skilled human evaluation rather than machine evaluation.” So the pilot measure is a person’s review of every record, with counts kept as they go.
Accuracy differs by field
Machine suggestions are not equally good across fields, which is why you count field by field. The Library of Congress’s Exploring Computational Description experiment (2022 to 2024) found that identifiers, author and title “can be extracted at around 85% accuracy, while others like subject headings, genre, and dates were only generated accurately at or under 50% of the time,” and concluded that “no current AI tools are currently good enough to run fully automatically for this task.” Those figures come from the Library’s work on book cataloging, not from letters or photographs, but the pattern is worth testing on your own material.
What to record
- For each field: suggestions made, accepted unchanged, edited, rejected, and left for a decision.
- Errors that would mislead a user, such as a wrong person or an invented event, listed apart from small edits.
- Minutes of review per record, by kind of item.
- Whether required fields are present and follow your vocabularies and date formats.
- How labels for machine-made values appear in the records. Our AI policy describes how we label machine output; whoever does your work, settle the scheme in the pilot.
Example (fictional). For the 68 loose items in Box 3, a model suggested names, places, dates and a short summary, and every suggestion was reviewed. Names: 152 suggestions, 131 accepted (86 percent), 17 edited, 4 rejected. Places: 74 suggestions, 66 accepted (89 percent), 6 edited, 2 left for the society to decide. Dates: 68 suggestions, 57 accepted (84 percent), 11 edited, mostly cabinet cards dated from mount style. Summaries: 68 drafts, 48 accepted (71 percent), 18 edited, 2 rejected and rewritten. One draft said a dam had failed, which the letter does not say. The society had set 80 percent accepted unchanged as its bar for each field. Names, places and dates passed. Summaries did not, so for production they stay drafts that a cataloger always reviews, and the budget carries that review time. Subject suggestions were not reviewed in the pilot and stay labeled as machine suggestions.
The test import: prove the records land in your system
Load the pilot’s records, text and files into the system where they will live before production starts, and check that every record arrives whole and opens the right images. A mapping problem found on a few hundred records is a small fix; the same problem found after production means reworking every batch.
The capture side of a project is often easier than the joins between systems. OCLC Research’s Rapid Capture report (2011) quotes staff at The Henry Ford, who found fast capture “straightforward in comparison to the issues involved in integrating” it with other workflows and systems. One of their issues was “understanding how the collection management system will ingest large volumes of image metadata, and then handle that data within an archival hierarchy.” Their digital processing archivist, Brian Wilson, concluded that integration “will require addressing issues that extend beyond the basic operation of the capture workstation.”
What to check in a test import
- Counts. Records and files loaded equal those delivered.
- Identifiers. Each record opens the right images and text. FADGI notes that identifiers “must be accurate to bring together distributed metadata about a resource.”
- Field mapping. Each value lands in the intended field, as agreed in a metadata mapping worksheet before the pilot.
- Dates. Approximate dates and ranges load without rejection or false precision.
- Characters. Accented letters and symbols survive import and display.
- Hierarchy. Pages attach to their volume, items to their folder.
- Labels. Labels on machine-made values arrive in a visible field.
- Restrictions. Restricted items stay hidden.
- Search. A search for a known name finds the right page.
- Export. Records and files come back out in an open format.
Example (fictional). The society’s first test import of Box 3 loaded 263 of 281 records. The 18 cabinet card records failed because the database’s date field accepted only full dates, and those cards were dated “about 1895”. The fix took one change to the mapping: a display date field for the words and a separate year field for sorting. The second import loaded all 281 records, with the machine-suggestion labels in a custom field. Finding this on 18 records was cheap. Finding it after 9,600 items would not have been.
Throughput, staff time and cost per item
Time every stage for each kind of original, including the hours your own staff spend, and turn those times into rates such as images per hour and minutes per record. Multiply by the collection’s counts to get a schedule and a budget built on your measurements rather than on someone else’s rates.
Rates from other projects travel badly. The OCLC Rapid Capture report collected throughput from several institutions and warned that “comparing throughput rates across the vignettes must take other factors into account.” Resolution was one: “for some of these projects 300 ppi was adequate resolution; others required as much as 2500 ppi. It can take significantly longer to capture images at higher resolution.” It also reported that at the Smithsonian’s Archives of American Art, “processing takes about the same amount of time as capture.”
What to log
- Start and stop times for each stage, for each kind of original: preparation, capture, recapture, review, text correction, catalog review, return.
- Setup separately from steady work. The first session on a new format is slower; log it, but do not let it set the production rate.
- Your own staff time: questions, approvals, reviews, rights decisions, the test import.
- The unit a quote will use, per image, page, item or hour, so the numbers drop straight into a budget.
Build in an allowance
Fenton’s chapter puts one lesson plainly: “Costs will be higher by more than you expect.” She advised building in “an allowance for these unexpected developments” from the beginning. A pilot narrows the uncertainty; it does not remove it. Carry a contingency line, and say in the budget which numbers were measured and which were estimated.
Example (fictional). The society’s pilot logs gave these steady rates, including handling: letters 48 images per hour, cabinet cards 36, the day book 24 (half the rate of the letters), clippings 20. Review of machine suggestions averaged 7 minutes per loose item.
| Work | Pilot rate | Full collection (estimate) | Hours |
|---|---|---|---|
| Capture of loose papers | 48 images per hour | 30,000 images | 625 |
| Recapture at the pilot rate | 2.2 percent of images | 660 images | 14 |
| Capture of each ledger | 24 images per hour | about 216 images per ledger | 9 per ledger |
| Review of machine suggestions | 7 minutes per item | 9,600 items | 1,120 |
The image and item counts come from the society’s inventory sample, not from a full count, and the budget says so. Description, not capture, is the largest line, which surprised the board. The project lead entered these rates in the project timeline calculator and the digitization budget calculator, with a contingency line on top.
Judging the results: proceed, change or stop
Compare each result with the criterion you wrote, then decide for the pilot as a whole: proceed as planned, proceed with named changes, or stop and rethink. Read small samples with care. A pilot that finds no defects in a few dozen images shows only that the true defect rate is probably below a fairly high ceiling.
Small samples have wide margins
A defect rate measured on a small pilot is an estimate with a margin of error. The NIST/SEMATECH e-Handbook of Statistical Methods gives the Wilson score interval for a proportion such as a defect rate. Applied to a pilot in which every image is inspected, and assuming production behaves like the pilot, it gives the ceilings in the chart below.
If you inspect 20 images and find nothing wrong, the true rate could still be about 16 percent. After 410 clean images the ceiling falls below 1 percent. For the fictional Box 3 pilot, 9 recaptures in 410 images is 2.2 percent, with a 95 percent interval of about 1.2 to 4.1 percent. That is precise enough to budget recapture time, not to promise a 1 percent rate, and it says nothing about material the pilot left out.
The pilot estimates; production quality control decides
The NIST handbook’s section on acceptance sampling draws a useful line: the “main purpose of acceptance sampling is to decide whether or not the lot is likely to be acceptable, not to estimate the quality of the lot.” A pilot is mainly an estimate; production quality control is a decision, batch by batch. Use the pilot to set the production rule. FADGI’s rule is that if more than 1 percent of randomly sampled images and metadata in a batch are defective, “the entire batch should be re-inspected”; below that, only the defective items are redone. The QC sample size calculator sizes the sample for each batch.
Sort every failure by its cause
- Systematic: a setting, a habit or a mapping. Fix it once, then re-test on a few items.
- Random: dust, an occasional misfeed, a slip in review. Budget for it at the measured rate.
- Inherent: the material itself, such as a cramped ledger hand. Change the method, correction level or scope, and budget the difference.
What a pilot report should contain
The report is the bridge to the full program: the sample and why each item was chosen; criteria and results side by side; every defect with its cause and fix; changes to the specification, handling, correction levels, review rules and import mapping; measured rates and the budget built from them, with the uncertainty stated; and the decision, with the name of the person who made it.
Example (fictional). The society’s report recorded these verdicts. Handling: pass, no new damage, every item returned in order. Capture: pass, with a separate setup for bound volumes. Text: pass for both sets of letters and the clippings; the day book missed the target before correction, so full correction was budgeted for every ledger. Description: pass for names, places and dates; fail for summaries, which now always get a cataloger’s review. Import: failed the first time, passed after the date mapping fix. Decision: proceed with five named changes, starting with a first production batch that includes a random draw from the 60 boxes.
When should a pilot end in “stop”? When a hard limit fails and no change to the method fixes it, when no affordable correction level makes the text fit for its use, or when the measured budget is beyond what you can fund. Stopping is a result too. It usually leads to a narrower plan, such as folder-level description with linked images, search-only text, or a first phase limited to the material that passed.
A pilot checklist for any vendor or in-house team
A pilot can be planned on a few pages and run in a few weeks. These steps work whether capture happens at a vendor’s studio, on site, or in your own building.
- Name the decision the pilot must inform: funding, vendor choice, correction level.
- Write the criteria sheet: hard limits, tolerances and planning measurements, with who judges each.
- Choose the sample in two parts: purposive, covering every format, hand and condition problem, and random.
- Write the capture specification for each kind of original, with FADGI category and star level.
- Agree the handling and custody record, with condition notes before and after.
- Prepare reference transcriptions for each hand and kind of page.
- Agree the metadata mapping and labeling scheme for machine-made values.
- Run every stage, logging times, defects and causes, and inspect every image and record.
- Do the test import into your own system.
- Turn the measured rates into a budget, write the report and make the decision in writing.
If you are going out to tender, make the pilot phase one of the contract; our guide to writing a digitization RFP shows where it fits. If you would like to run a pilot box with us, the engagements page lists what it includes, and you can plan one with us. Either way, keep the criteria sheet: it is the most useful page you will write before the work begins.