By Sagar Shankaran, Founder of CallSphere
Fourteen box scores, a faxed stat sheet, a penciled scorebook photo. How 2026 document readers cut keying at the sports desk and where they still get names wrong.
Key takeaways
You tried this in 2024. A coach texted a photo of the scorebook page, you fed it to a scanning app, and it gave you back a column of numbers with no idea which ones were points and which ones were jersey numbers, plus a player named "M. Sxhultz." Your sports editor typed the whole thing in by hand anyway, at 10:50 p.m., with the 11:15 page deadline sitting on him. You concluded that this was not ready. You were right then. It is a different situation now, and the reason is narrow enough to explain in a paragraph.
What changed in 2026 is that models got genuinely good at reading the ugly documents real businesses receive — a phone photo taken at an angle under gym lights, a fax that went through a machine twice, a handwritten page with initials in the margin — and, more importantly, at understanding what kind of document they are looking at. A model reading a basketball scorebook now knows it is a scorebook: that the left column is players, that the running tally goes by quarter, that FG‑FGA is a made-attempted pair and not a subtraction, that a circled number means a made free throw in most scorebooks and something else in a few. Multimodal document reading means the model reads the artifact you actually received — the photo, the fax, the handwriting — instead of requiring a person to retype it into a form first.
Here is the Friday your sports desk actually has, whether you are a weekly with a Monday deadline or a small daily going to press at 11:15. Fourteen games across football, volleyball and the start of basketball. Three coaches text photos of the penciled scorebook page. Two athletic directors fax a stat sheet, because that is what the school office has always done. A stringer calls in from a parking lot reading off a legal pad. Two coaches send nothing and get chased Saturday morning. The junior high results arrive on a sheet somebody's parent filled out in pen. All of it has to become agate — the 6-point type on the scoreboard page — before the page closes.
The person doing it is usually one full-time sports editor and, if you are lucky, a part-time stringer being paid $35 a game. The work is pure transcription, it happens in the last ninety minutes of the night, and it is where your paper generates the errors readers notice most. Nothing draws mail like a wrong point total on somebody's kid.
The sequence now: the coach's text photo comes in. The model reads it as a scorebook page, pulls the roster names and jersey numbers, the quarter-by-quarter scoring, the field goals made and attempted, the three-pointers, the free throws and the rebounds. It returns the box score already formatted to your house style — the same column order your paginator has used since the Vision Data days — with the team totals added and checked against the quarter line score. If the totals do not reconcile, it says so instead of publishing a box score that does not add up.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
What makes this usable on deadline is the flagging. A good setup does not guess at the smudged cell; it marks it, and your sports editor sees four highlighted cells instead of a wall of 180 numbers. He looks at four, fixes two, and the page closes at 11:04 instead of 11:19.
flowchart TD
A["Coach texts photo of penciled scorebook"] --> B["Read roster names and jersey numbers"]
A --> C["Read quarter-by-quarter line score"]
A --> D["Read FG-FGA, threes, free throws, rebounds"]
B --> E{"Totals reconcile and every cell legible?"}
C --> E
D --> E
E -->|No| F["Flagged cells go to sports editor's screen"]
E -->|Yes| G["Box score set in agate for the 11:15 page"]
F --> G
Sports is the loudest example, but a community paper keys in far more. The faxed sheriff's sale from the courthouse. Village board minutes that arrive as a scan of a printout with the clerk's handwritten correction in the margin. Church bulletins and club submissions dropped at the front counter. Real estate transfers and court dispositions from the county. Handwritten letters to the editor. "Looking Back" photos brought in by a reader with the date and names in pen on the back of the print — and that pen-on-the-back is exactly what the model can now read and turn into a cutline.
On the business side: signed ad proofs faxed back with "OK w/ changes" and initials in the margin, insertion orders on the advertiser's own form, obituaries emailed from funeral homes as a scanned PDF of a typed page, and classified line-ad tickets from the front counter. Every one of those gets retyped today by somebody whose job title is not "typist."
Assume a paper covering four high schools plus junior high, with 14 events on a typical Friday during the winter overlap. Assume six minutes to key and proof each box score by hand, and that your sports editor is loaded at $28 an hour. Assume the reader cuts keying time to two minutes per box score including the check on flagged cells.
| Line | Assumption | Result |
|---|---|---|
| Box scores keyed per week | 14 Friday + 8 midweek | 22 |
| Minutes each, by hand | key + proof | 6 |
| Weekly keying time today | 22 × 6 min | 2.2 hours |
| Minutes each, reading the photo | review flagged cells only | 2 |
| Weekly keying time after | 22 × 2 min | 0.73 hours |
| Hours saved over an 18-week season | 1.47 × 18 | 26.5 hours |
| Value at $28 loaded | $742 | |
| Typos avoided | 1 error per 40 keyed figures, 180 figures/wk | ~81 per season |
Seven hundred dollars a season is not the argument. Two things in that table are. First, the 26.5 hours come out of the worst hour of the week — the 90 minutes before the page closes, which is when your sports editor decides whether he still likes this job. Second, the typo line: at one error per forty keyed figures, an ordinary winter produces roughly eighty wrong numbers in agate. Most go unnoticed. The ones that do not generate a phone call from a parent, a correction, and a small permanent dent in the thing you actually sell, which is being right about your own town.
Handwriting that is genuinely bad stays genuinely bad. A model reading a scorebook filled in by an assistant coach in the dark during a comeback will flag half the fourth quarter, and it should. The test of a good setup is not that it reads everything — it is that it never quietly invents a number it could not see. Check that in your first month by handing it your three worst photos.
Names are the other permanent hazard. It will read "Schultz" where the scorebook says "Shultz," because it has seen more Schultzes. Your roster list is the fix: give it the actual rosters from each school's athletic department at the start of the season and require it to match every name to that list or flag it. Nothing else in this piece matters as much as that one step.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
And keep a person on anything with a legal consequence. A faxed foreclosure notice or sheriff's sale is read by the model and then compared, character by character, to what your legals clerk sets — it is never allowed to be the sole source of statutory wording. Same for the minutes: the model can lift the text, but whether a village board actually voted 4‑1 on the sewer bid is a fact your reporter confirms.
Do not start live. Take the twelve worst artifacts of last season out of the folder — the crooked photos, the double-faxed stat sheets, the parent's handwritten junior high results — and run them through Claude Cowork or ChatGPT Work with your house box-score format pasted in and the rosters attached. Compare what comes back with what you published. That is a two-hour exercise for your sports editor on a Tuesday, and it tells you whether this survives a real Friday. If it does, use it in parallel for one full week before it touches the page.
If the fax lands as a PDF or an image in an inbox, yes — that is exactly the kind of degraded input the 2026 readers were built for. If it lands as paper on a tray, the fix is a fax-to-email number, which is a cheap change your phone provider can make and which pays for itself on the courthouse notices alone.
Different formats, same job, and each one needs its own house template and its own week of parallel running. Wrestling brackets are the hardest because the information is in the lines on the page, not just the text. Start with the sport that produces the most box scores; get that clean before you add the next.
Frontier AI is down roughly ten times from 2025 prices, and reading a few hundred documents a week sits well inside a small paper's software budget. Run one month, look at the bill, and set a spending cap — the enterprise plans now include spend limits and alerts at 75 and 90 percent so it cannot quietly run away from you.
It can produce a competent three-paragraph roundup from a box score, and for junior high and freshman results that nobody was going to write up anyway, that is a real gain. For your Friday night varsity lead, no — that story exists because a person was in the gym, and readers can tell within two sentences whether anyone was.
The other half of Friday night is the phone: coaches calling in results at 9:40, a parent asking whether the volleyball score made the paper, and the Saturday morning calls from the two teams nobody could reach. CallSphere builds AI voice and chat agents that answer a business line and website chat and capture the details 24/7, so a called-in score at 9:40 p.m. arrives as a written record on the sports desk instead of a voicemail nobody plays until Monday. The reading of the scorebook photo is a separate tool — the point is that neither the call nor the paper should be waiting on somebody with a keyboard at 10:50.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Carbon-copy scale tickets, margin notes and container photos still get keyed by hand. What 2026 document readers catch, and the unbilled extras they surface.
Broadcast reconciliation eats 26 hours a month at a mid-market agency. How 2026 document readers handle faxed affidavits, margin notes and makegoods.
Handwritten shorts on back-door delivery tickets leak thousands a year across a 14-store group. What 2026 document reading changes, with the arithmetic.
Newsprint is ordered nine weeks ahead and the draw is a guess. How 2026 forecasting models cut returns, catch sellouts and set renewal prices by cohort.
Heat numbers get retyped four times between the mill and your AS9102 Form 2. What 2026 document reading changes at an aerospace supplier's receiving dock.
Colo cross-connect orders still arrive as faxed scans with handwritten port pairs. Here is what 2026 document reading does to the provisioning desk and the MMR.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI